Savannah, Georgia Correspondent

In port logistics, we talk about throughput because ships have to move, cargo has to be located, and somebody eventually has to get a container onto a truck. A port that leaves every vessel at anchor while managers admire a new scheduling system has not improved its operations. Yet the opposite mistake is just as common: an impressive dashboard reports that containers are leaving the terminal faster while customers discover that the wrong boxes have arrived, customs holds are multiplying, and drivers are wasting hours waiting for the last piece of documentation. That is how I approach artificial intelligence in medicine. The technology may be reading images faster, classifying a suspicious finding more consistently, or finding a promising compound in a database. I want to know whether the patient reaches the right diagnosis and receives useful treatment with less delay, less harm, and a smaller bill. Those are not the same measurements as the ones displayed at a software demonstration.

Medicine is obviously not shipping, and I do not mean to reduce a sick person to a unit moving through a system. Quite the opposite. A misplaced container can usually be recovered. A delayed diagnosis can alter a life in ways no spreadsheet can repair. What I recognize is the architecture of complicated organizations: separate departments, narrow responsibilities, specialist vocabulary, fragile handoffs, and software that assumes the next person will do exactly what the previous person expects. When an algorithm notices an abnormal scan, that information still must reach a clinician who can interpret it, a patient who understands it, a scheduler who can arrange the next test, and a treatment team with available capacity. A wonderful first step does not compensate for a broken journey.

There is now credible evidence that the first step really is improving in selected applications. Germany’s PRAIM implementation study examined more than 463,000 mammography screenings, with AI-supported reading used for roughly 261,000 women. Investigators reported a breast-cancer detection rate of 6.7 per thousand with AI assistance against 5.7 per thousand without it, an adjusted relative improvement of 17.6 percent. The difference matters, though the study was observational and participating radiologists selected whether to use the system. It does not establish that an AI deployment anywhere in America will reproduce the result. The Swedish MASAI randomized trial provides stronger evidence in an organized screening setting. Its 2026 analysis, involving approximately 106,000 women, found higher sensitivity with AI-supported screening, unchanged specificity, and an interval-cancer rate that met the trial’s noninferiority criterion. Those are serious findings, not promotional slogans.

But even here we must ask what exactly improved. Detecting additional cancers at screening can be valuable, and finding disease sooner may enable less invasive treatment. It is not identical to demonstrating fewer deaths from breast cancer, better long-term quality of life, or a lower overall burden of unnecessary intervention. A screen can detect a lesion that would never become clinically dangerous. A recall can frighten a person who has to take a day off work, find childcare, and drive to another county. The true benefit has to include the resulting clinical course. I would be delighted to learn that AI helped identify a dangerous cancer earlier; I would be less impressed by a vendor counting every extra biopsy as success regardless of what followed. Numbers need a destination.

Something similar is happening with diabetic-retinopathy screening. A clinic can use an authorized autonomous system to evaluate retinal images without asking an ophthalmologist to read each picture. In a small 2026 real-world integration study in an endocrinology clinic, the system produced definitive results for most participating patients. That is promising because diabetic eye damage is one of those problems that can advance without much warning until treatment becomes difficult. Yet the study also documented that fewer than half of AI-positive patients completed confirmatory examinations at the institution; outside follow-up was not captured. This does not mean the software failed. It means success depends on a second operation, the one that gets a patient from a warning to an actual specialist visit. In logistics terms, the alert is a manifest, not a delivery receipt.

People discussing these systems often assume the healthcare organization can simply assign a nurse or coordinator to chase each flagged patient. I have worked in seasonal operations long enough to know that labor does not appear because a planning spreadsheet includes it. Medical offices are already managing phones, records, insurer approvals, missed appointments, and staff turnover. A new alerting system may reveal unmet need faster than the practice can address it. That is a genuine capacity problem, not evidence of technophobia. If the technology discovers a hundred patients who need care and the local service can see only twenty, the bottleneck has moved rather than disappeared. We should measure the time between screening and effective treatment and budget honestly for the people who make that happen.

The larger language models now being proposed as diagnostic partners raise another issue familiar to anyone who has suffered through a poorly designed operational system: confident output. When software prints a highly polished explanation, people can confuse the quality of the presentation with the reliability of the underlying information. A 2024 randomized trial published in JAMA Network Open found that providing physicians access to a large language model did not significantly improve their diagnostic reasoning compared with conventional resources, even though the model alone performed strongly on the study’s cases. That should make us curious rather than triumphant or dismissive. Perhaps the interface was badly suited to clinical practice. Perhaps clinicians relied on the wrong portions of the answer. Perhaps the cases did not reflect actual wards. What it does not justify is installing a chatbot and assuming teamwork automatically produces the sum of human and machine intelligence.

A separate JAMA experiment involving clinical vignettes illustrates a more worrying outcome. Clinicians shown accurate AI assistance improved in the study, but systematically biased suggestions substantially lowered diagnostic accuracy. Explanations attached to the erroneous outputs did not undo the harm. That is the medical equivalent of a navigation system presenting a beautiful map that directs every truck across a bridge closed for repairs. The problem is not merely the incorrect instruction; it is the way a familiar, orderly interface persuades users to trust the instruction more than competing evidence. In a hospital, such over-reliance may be especially dangerous late at night or when a team is short-staffed.

Consider how many kinds of AI are being grouped under one fashionable label. Image enhancement, suspicious-lesion detection, medication-dose adjustment, laboratory-data interpretation, genetic-variant ranking, and free-text conversational advice are not one product with one risk profile. A tool that highlights a region of a scan so a radiologist can inspect it is different from a program that writes a differential diagnosis, and both differ from an autonomous screening device with a defined indication. The U.S. Food and Drug Administration reports more than 1,600 authorized AI-enabled medical devices as of September 2026. That scale reflects real engineering and regulatory activity. It does not mean 1,600 technologies have independently demonstrated improved survival or can safely replace a professional. Authorization and a demonstrated patient-outcome benefit are related but distinct achievements.

The drug-discovery story deserves similar care. Software can sort through molecular possibilities far faster than a conventional laboratory team and may help researchers identify targets that deserve testing. In 2025, a randomized phase 2a study of the AI-discovered compound rentosertib for idiopathic pulmonary fibrosis reported encouraging signals in a small group, including a difference in forced vital capacity at the highest studied dose. Seventy-one patients entered that trial, which followed treatment for twelve weeks. Those details are important. The study did not establish a broadly available cure or prove that the development process can consistently deliver safe drugs more cheaply. It showed that a candidate arising through AI-enabled discovery could advance into meaningful clinical testing and produce findings worth investigating further.

That is still a considerable accomplishment. Many people have a legitimate interest in more efficient drug research because existing therapies can be extraordinarily expensive and slow to develop. I would like researchers to find treatments for rare disorders, cancers, and chronic conditions that have frustrated families for years. But the expensive part of medicine is often not finding something that looks effective in a computer. It is establishing through animal work, manufacturing controls, human trials, longer observation, and adverse-event surveillance that the intervention genuinely helps. A faster start is valuable. Calling the race won because the first mile was quick insults the patients who must live with the consequences of an unfinished product.

Genetic diagnosis may be particularly well suited to machines that can assemble information from many places. A family with a child who has unexplained symptoms can spend years moving among specialists. Sequencing may identify enormous numbers of genetic variations without explaining which, if any, causes the illness. Recent systems such as DeepRare use clinical descriptions, genomic findings, and referenced medical evidence to rank possibilities, and a 2026 study involving previously unsolved childhood cases showed that AI-supported literature and genomic review can surface new diagnostic leads. But a lead is not a clinical diagnosis. It still needs review by qualified specialists, laboratory confirmation where appropriate, counseling about uncertainty, and an honest explanation of what knowledge remains incomplete. Families deserve hope without a sales pitch.

The difficulty of transferring technology between sites is another reason to resist universal claims. A radiology model trained on one combination of scanners, protocols, ages, and patient characteristics may behave differently after a hospital replaces equipment or begins serving another population. A system interpreting clinical records must navigate abbreviations, copied notes, missing entries, and ordinary human mistakes. I have seen terminal information systems struggle because one carrier uses a different code or timestamps events differently. In healthcare, the translation problems multiply. Anyone purchasing an AI tool should demand evidence from the kind of patients and settings in which it will be used, not simply from the institution whose logo looks best on a sales presentation.

Patient privacy matters even when the practical gains are real. Imaging files, genomic data, and medical histories contain information people cannot easily change after a breach. A port operator can issue a replacement access card; a patient cannot replace their genetic sequence. Hospitals therefore need clear rules on which information goes to outside vendors, how long it is retained, what purpose it serves, who can audit its use, and what happens when a vendor changes ownership. The safeguards should be intelligible to patients rather than buried in legal prose. There is no moral justification for forcing someone to surrender control of intimate information as the hidden price of obtaining an ordinary medical consultation.

Hospitals also need a way to turn systems off without creating chaos. If an AI-assisted imaging tool goes down, does the reading queue freeze? If a vendor updates the algorithm, who verifies that performance has not changed? If the software flags too many false positives, who notices the accumulated burden on clinicians and patients? I would insist on practical contingency procedures, local performance reports, audit trails, and named responsibility for each decision. In maritime operations we plan for weather, equipment failure, and staffing interruptions because these things happen. An algorithm is not an exemption from contingency planning. Software failure should be a manageable incident, not a surprise that exposes the absence of human capability.

The most useful evaluation would follow an entire group of patients through the system. How many eligible people were screened? How many received a useful result? How many abnormal findings were confirmed? How long did it take to reach treatment? Did clinicians miss fewer important diagnoses? Were avoidable admissions reduced? Did outcomes vary by language, insurance status, location, or income? How many false alarms required follow-up? Were staff freed for patient care, or did alerts create another stack of tasks? These questions sound pedestrian compared with claims about machine intelligence, but they are what turn innovation into accountable service. A hospital should not be graded on the brilliance of its software vendor’s press release.

Some critics would use the risk of error to stop most deployment. That would also be a mistake. Human medicine makes errors, including errors from fatigue, poor communication, and uneven access to specialists. A carefully assessed system that helps an overburdened radiologist detect a cancer, or helps a rural clinician obtain a useful second look, can improve care. The question is not whether humans are perfect, because they are not. It is whether a particular arrangement of people and tools performs better than the realistic alternative, for the actual patients who must use it. Comparative evidence should be the standard, and that evidence should be collected after installation as well as before it.

There is also a lesson about promises made from a national stage. Washington can encourage technology, fund trials, and issue regulatory guidance. But it cannot by itself create a specialist appointment at a rural clinic next Thursday or replace a health-system employee who resigned last week. Those operating details require local ownership and honest funding. I am tired of programs that announce success on the day the software license is signed. The hard work starts after the ribbon-cutting, when ordinary people are expected to reconcile the new system with everything that already has to happen.

My conclusion is neither that AI will save medicine nor that it will ruin it. In imaging and selected screening workflows, the evidence is increasingly serious and encouraging. In rare-disease interpretation, it offers clues that deserve disciplined verification. In drug discovery, it may accelerate portions of a long and uncertain process. In general diagnostic advice, the dangers of bias and unwarranted confidence remain substantial. The dividing line is not between innovators and skeptics. It is between organizations willing to measure real consequences and organizations satisfied to count outputs. You can build a modern terminal, install the finest scheduling software in the world, and still leave the customer waiting on the dock. A healthcare system that buys AI without completing the patient’s journey is making the same mistake, except the stakes are incomparably higher.

Phoenix, Arizona Correspondent

From where I sit in hospital operations, the conversation about artificial intelligence in medicine is taking place on two different floors. Upstairs, executives attend demonstrations of algorithms that can examine scans, organize clinical notes, forecast deterioration, and search the literature at astonishing speed. Downstairs, someone is looking for an available bed, a nurse is trying to reach a physician, and a patient who has already answered the same questions three times wonders when anyone will explain what happens next. Neither picture is false. But if you want to know whether AI improves care, start downstairs. A system that predicts deterioration is useful only if somebody can recognize the alert, determine whether it matters, and respond before the patient becomes sicker. A program that accelerates imaging interpretation helps only if the result reaches the right person and a meaningful intervention follows.

I am not a physician, and that distinction matters. I work with scheduling, contracts, paperwork, and the organizational machinery that clinicians depend on. I have enormous respect for the people who diagnose and treat patients, partly because I see the amount of work required simply to allow them to do their jobs. The public often imagines a hospital as a continuous flow of medical decisions. In reality it also contains insurance authorizations, transport delays, software incompatibilities, staffing shortages, handoff failures, pharmacy questions, and people who cannot afford to miss another day of work. Technology can reduce those burdens, but it can also create them. Anyone who tells you every new alert saves time has probably never watched what happens when the alerts arrive by the hundreds.

Let’s begin with a bright spot. Two substantial European mammography studies provide stronger evidence than the usual collection of screenshots and laboratory benchmarks. Germany’s large PRAIM study found a higher breast-cancer detection rate when radiologists used an AI-supported reading system. The Swedish MASAI randomized trial, with about 106,000 women enrolled, reported higher sensitivity for AI-supported screening than standard double reading, with the same specificity and a noninferior rate of cancers appearing between scheduled screens. It also reported reduced reading workload. I do not dismiss those outcomes. Screening programs are labor-intensive, and a tool that maintains or improves detection while reducing work could release scarce professional capacity. For patients waiting to schedule examinations, that is not an abstract advantage.

We do have to be precise about what a study says. Detection rates and sensitivity are important, but they are not the same as demonstrating a reduction in mortality. Earlier detection does not automatically produce a better outcome in every case; follow-up treatment, cancer biology, access, and the harms of unnecessary procedures still matter. The Swedish trial’s organization, training, image acquisition, and available expertise are part of the intervention. A Phoenix hospital cannot assume it will replicate every result by purchasing a similar product. It needs to determine whether local populations and workflows resemble those studied, especially when patients have different clinical backgrounds or histories of interrupted care.

Now consider sepsis. Any hospital understands how quickly a severe infection can deteriorate. If a system detects a dangerous pattern hours before human recognition, clinicians may be able to intervene sooner. That is an exciting possibility. Yet a 2026 multicenter prospective validation of an updated Epic sepsis prediction model found that performance differed across four major health systems. At a threshold chosen to yield 60 percent sensitivity, positive predictive values were about 13 to 26 percent depending on the site. Put plainly, many alerts did not correspond to a case meeting the study’s sepsis definition. That does not prove the model is useless; prevalence and thresholds matter, and a false alarm may be an acceptable price for not missing a serious event. But somebody must account for every alarm’s effect on bedside care.

Imagine a medical-surgical nurse who is caring for patients with complex conditions and receives another notification indicating elevated sepsis risk. What should happen? Does the nurse call the physician immediately? Repeat vital signs? Request additional laboratory work? Begin a protocol? What if the patient has already been evaluated and the system keeps firing? What if the score is misleading because of a chronic condition or a transient laboratory result? These are not objections invented by people who fear computers. They are exactly the questions that determine whether clinical decision support produces earlier treatment or simply adds noise to an overloaded shift. Alert fatigue is not a character flaw. It is a predictable response to too many interruptions without reliable prioritization.

Hospitals should measure not only how accurately an algorithm predicts an event but also how teams use its output. A retrospective score may look excellent when researchers know which patients later developed sepsis. That does not tell us whether staff can act in real time, how often the alert arrives before a human has recognized the problem, or whether a treatment protocol changes outcomes. Some warnings will arrive when action is already underway. Others may occur long before a treatable condition is clear. A clinically useful system must achieve the difficult middle ground: sufficiently early to matter, sufficiently reliable to warrant attention, and embedded in a response that is both safe and practical.

There is a temptation to compare the machine’s diagnostic accuracy with an individual physician’s score on test questions. But the hospital does not deliver care through an isolated physician answering a multiple-choice examination. It delivers care through teams. A 2024 randomized study found that physicians offered access to a large language model did not achieve a significant diagnostic improvement relative to physicians using conventional resources, despite the model’s strong performance when answering separately. I find that less mysterious than some technologists do. A clinical team has habits, obligations, distractions, and existing information systems. Adding a clever consultant who speaks in perfect paragraphs is not automatically helpful. It may even be one more place a busy clinician must look.

The more frightening problem is the perfectly worded wrong answer. Research using clinical vignettes showed that biased AI recommendations could reduce clinicians’ diagnostic accuracy and that explanatory text did not rescue the situation. A persuasive explanation can make an incorrect prediction feel safer. The danger is not that clinicians are foolish. People must rely on information systems every day, and they cannot independently reconstruct every computation. A good hospital therefore needs defined roles: what the algorithm may suggest, what a clinician must independently verify, which cases require escalation, and when human judgment overrides the screen. Accountability must remain attached to identifiable people and institutions.

Hospital leaders also have to resist a financial incentive that rarely gets mentioned at conferences. Software can be purchased with a capital or subscription budget. The follow-up work often appears somewhere else: nursing hours, care coordination, extra imaging, additional referrals, and calls to patients who do not understand what a message means. A vendor may claim it will reduce cost by helping identify risk. The hospital should ask for the full balance sheet, including training, quality monitoring, incident investigation, hardware, downtime plans, cybersecurity, and the staff needed to respond to the algorithm’s findings. A tool can be accurate and still make operations more expensive if it increases work that cannot safely be deferred.

The diabetic-retinopathy screening example is particularly revealing. An autonomous retinal-imaging system can extend basic screening to patients who may not have easy access to an ophthalmologist. One 2026 implementation study obtained definitive results in nearly 96 percent of participants without dilating their pupils. That is the kind of practical efficiency a medical clinic can use. Yet less than half of the patients with positive screens completed confirmation at the institution, with outside follow-up unobserved. The patient does not receive the benefit of early detection merely because a computer has generated a referral. Someone has to arrange it, explain it, and make sure the person can attend. In a state with long drives, intense summer heat, and uneven access to specialists, that last part can be decisive.

Rare genetic diseases present another area of enormous promise and uncertainty. Families sometimes arrive with long binders of reports and years of unanswered questions. A system able to compare symptoms, sequencing data, and fresh medical literature may identify a candidate explanation that a small team has not considered. Recent research on systems such as DeepRare reports strong performance in curated cases, and other 2026 work has shown that AI-assisted review can surface diagnostic leads from previously unresolved pediatric files. Those developments deserve serious attention. Yet a ranked hypothesis is not proof that a variant causes illness, and even a correct molecular diagnosis may not immediately yield a treatment. Counseling, confirmatory testing, and careful communication remain essential.

That last point matters because hope is a medical resource of its own. Patients with unexplained illness may have been dismissed, misdiagnosed, or exhausted by years of appointments. An AI tool that presents a possible answer in confident language can be deeply appealing. Clinicians must resist the temptation to transform possibility into certainty simply because the patient needs an answer. They also must avoid withholding legitimate new possibilities because the tool that found them is unfamiliar. Respect means explaining degrees of certainty, what evidence could confirm a finding, and what finding it would not prove. It means allowing patients to understand the limitations rather than presenting technology as a miraculous authority.

Our population in Arizona is hardly uniform. Patients arrive from different cultures, speak different languages, have distinct patterns of chronic disease, and bring varying access to regular primary care. A model trained elsewhere may underperform in subgroups that were poorly represented during development. The World Health Organization has explicitly warned that large multimodal models can reproduce bias from data and can create false, incomplete, or misleading medical information. For hospital operations, that warning has a concrete meaning. We should evaluate outputs by relevant patient groups, track disparities in referrals and completed care, and allow frontline staff to report anomalies without being told they are obstructing progress.

Data privacy is another operational problem hiding inside a technical one. A clinician may enter a lengthy patient history into an approved system, but what happens after the submission? Is the information used only to generate an answer, or retained for development? Can a vendor’s subcontractors access it? What if the hospital changes platforms? The organization must be able to explain its arrangements clearly, enforce access restrictions, and respond to incidents. Genetic details deserve special care because they concern relatives as well as the individual patient. Security should be designed before the first deployment, not added after a breach becomes a television story.

We are also hearing a great deal about AI-discovered pharmaceuticals. The 2025 phase 2a trial of rentosertib for idiopathic pulmonary fibrosis provided an important demonstration that a treatment emerging from AI-enabled discovery can reach rigorous human testing. But it was a short study of seventy-one people with encouraging signals requiring larger trials. I would be delighted if AI reduced years of wasted laboratory effort. That does not release developers from proving durable efficacy or monitoring harms. Hospitals will still have to deal with prescribing information, interactions, reimbursement, supply continuity, and the patients who respond differently from the trial average. We cannot order a pharmacy shelf full of hope and call that a treatment program.

The training issue is bigger than teaching clinicians which button to press. A radiologist or physician needs to know the tool’s intended use, known weaknesses, local validation, and the circumstances under which a suggestion should be ignored. Nurses need to know whether an alert creates a new duty or simply informs an existing one. Administrators need to know who owns a quality problem and how incident reports will lead to adjustments. Executives need to hear about false positives and near misses with the same attention they give early success stories. And patients need a way to ask whether automated advice influenced their care without being made to feel unreasonable.

I would require a staged rollout rather than a ceremonial launch. Begin with a defined clinical question, baseline measurements, and a process for reviewing cases. Test the model quietly before connecting it to active alerts, when appropriate. Then introduce it with clear response protocols and an escalation route for doubtful results. Monitor missed cases, unnecessary testing, delays, outcomes, and staff workload. Review performance after updates and after changes in the patient population. If the system does not help, stop it. A hospital should not keep a weak tool running because the vendor has an impressive booth at a national conference.

One of the best possible uses of this technology may be almost invisible to the public. If it helps reduce duplicated work, sorts urgent studies appropriately, organizes information for clinicians, or makes it easier to find a suspicious change in an image, the patient may never encounter a chatbot at all. There is nothing wrong with that. The objective is not to persuade patients that they live in a futuristic hospital. The objective is to provide competent, timely, humane care. The medical workforce will remain essential, and efficiency should be used to give clinicians more useful time with patients instead of merely making appointment slots tighter.

I am optimistic about specific tools and skeptical of grand announcements. Both attitudes can live together. The evidence in mammography is substantial enough to merit serious adoption discussions, not blanket rejection. Other applications require careful local testing because benchmark success is not clinical success. Diagnostic errors, automation bias, and imperfect follow-up are not minor side issues; they are central to whether an intervention helps anyone. We should welcome technology that makes a genuine difference while remembering what every crowded waiting room teaches: patients do not experience medicine as a collection of innovations. They experience it as whether somebody listens, makes the right judgment, and gets them safely to the next step. If AI cannot improve that experience and the outcomes behind it, its intelligence is not the part that matters.

Orange County, California Correspondent

Southern California has no shortage of people willing to announce that a new technology will transform an entire industry. We have heard the language in real estate, finance, transportation, communications, and nearly every service business where a venture-capital presentation can be confused with a completed project. Sometimes the claims prove right. More often, the technology solves a genuine problem but leaves the larger economics largely intact. That is why I am interested in artificial intelligence in medicine, but cautious about promises that it will suddenly make healthcare cheaper, more equitable, and more effective. An algorithm may read an image in moments; a patient can still spend weeks seeking insurance authorization, arranging an appointment, and waiting for treatment. Medical progress is not only a question of what machines can recognize. It is a question of how organizations distribute the gains, bear the risks, and change their behavior.

I do not work in healthcare. My professional experience is in commercial real estate analysis and local economic development, where investments are judged by occupancy, costs, financing, risk, and actual performance rather than by how attractive the original projection looked. Those instincts are applicable here. Every health system buying AI should have to say exactly what improves, by how much, and compared with what alternative. Does the software make clinicians more accurate? Does it reduce staffing pressures? Does it prevent expensive complications? Does it increase billable procedures without helping patients? Is it cheaper to purchase than to maintain? Who keeps the savings? Until those questions are answered, predictions about an inexpensive new age of medicine are incomplete.

The clearest advances appear in narrowly defined clinical tasks. The German PRAIM study, involving more than 463,000 screened women, reported higher rates of breast-cancer detection using AI-supported mammography than conventional double reading. The Swedish MASAI randomized study reported higher sensitivity, unchanged specificity, and a noninferior interval-cancer rate. Those results are meaningful because they emerge from large populations and clinical settings, not merely a contest in which software is asked to classify a carefully chosen set of images. I would expect hospitals and imaging centers to examine the findings closely. But I would also ask how much the entire screening process costs per clinically significant cancer detected, how many additional workups follow, and whether earlier discovery changes subsequent outcomes.

That distinction is not a request for perfection before adoption. Established medical procedures often have benefits that vary across patients, and evidence accumulates over time. My concern is that a new technology can create a powerful business incentive to measure what is easy rather than what matters. Software vendors can publish processing speed, sensitivity, or agreement with expert readers. A health system can report expanded screening capacity. Payers may calculate claims associated with additional examinations. The patient cares about whether a serious disease is identified and treated, whether an unnecessary alarm disrupts daily life, and whether the cost threatens household finances. An economic evaluation that omits those consequences is like a development pro forma that counts new leases but ignores the cost of keeping the building open.

There is another financial problem: costs and benefits frequently fall on different institutions. Suppose a hospital pays to install a system that detects retinal damage earlier among patients with diabetes. If it succeeds, the eventual savings from avoided blindness may accrue years later to insurers, public programs, employers, families, and patients rather than to the hospital that made the investment. That makes the project socially valuable but possibly unattractive to the organization being asked to pay for it. Conversely, a system that generates more imaging and procedures may improve one provider’s revenue even when the broader benefit is uncertain. We should not pretend that every institution is automatically rewarded for pursuing the best patient outcome.

The real-world retinopathy evidence illustrates the difficulty. In a 2026 study integrating autonomous screening into an endocrinology clinic, the device obtained definitive results for most participants, but many of those who screened positive had no documented confirmatory visit at that institution. External follow-up was not captured. The technology can identify a need while the larger system fails to meet it. Payment arrangements, transportation, the availability of ophthalmologists, and patient understanding all influence whether detection turns into care. That is why contracts for clinical AI should include meaningful measures of completed follow-up and outcomes, not only the number of scans analyzed. The public cannot afford to mistake more warnings for more treatment.

The growing inventory of AI-enabled medical devices is impressive. The FDA reported more than 1,600 authorized products by September 2026. I would not dismiss the underlying engineering effort. I would, however, resist interpreting that total as 1,600 independent proofs that patient outcomes are better. Authorization is tied to a device’s intended use and regulatory pathway; it is not a universal certificate of reduced mortality, shorter hospitalizations, or lower spending. The distinction is especially important when the label ‘AI’ encompasses image enhancement, measurements, screening, prediction systems, and clinical decision support with very different risk profiles. Consumers are being sold a category as though it were one invention.

From the purchaser’s perspective, I would divide products into several groups. One group has a narrow, objective task, a reasonable way to verify results, and a mature clinical workflow. Another offers useful predictions that clinicians must interpret amid uncertainty. A third generates broad text or diagnostic suggestions that can sound plausible even when wrong. The first group may be comparatively straightforward to evaluate; the third demands more skeptical attention. I would never apply the same diligence checklist to a building’s fire-alarm sensor and to a consultant recommending whether to demolish the building. Yet medical institutions sometimes speak of ‘AI adoption’ as one project instead of a portfolio of distinct decisions.

This matters when executives promise a return on investment. Some efficiencies are tangible: fewer redundant reads, faster triage, better scheduling, reduced administrative work, and less time spent locating information. Others are speculative. A system might help find cancer sooner, which could ultimately reduce expensive late-stage treatment. That benefit is important, but it needs longitudinal evidence and a fair estimate of follow-up costs. Still other programs might create more work because every false alarm leads to extra calls, tests, and visits. Hospital CFOs should examine all three possibilities rather than rely on a vendor’s preferred performance indicator. If a project is described as both life-saving and cost-saving, both claims deserve separate proof.

Clinical reasoning is where optimistic assumptions become particularly fragile. A 2024 randomized trial found that physicians given access to a large language model did not show a significant diagnostic improvement relative to physicians using conventional resources, even though the model alone performed well on the test cases. Some advocates treated that as an interface challenge; skeptics treated it as a refutation of medical AI. I see it as a warning about implementation risk. A capable technology does not become a capable work process simply by being purchased. Training, incentives, time pressure, and the design of the interaction can determine whether its theoretical advantage survives. That is familiar in every organization that has introduced expensive software and then discovered employees must perform duplicate work to keep operations functioning.

More concerning are experiments showing that systematically biased AI advice can impair clinicians’ decisions even when an explanation is provided. The explanations can create the appearance of transparency without giving a busy practitioner a reliable method of detecting error. A vendor’s marketing department may emphasize explainability, but explanation is not the same as accountability or accuracy. The system should have clear limits, documented performance on relevant populations, and a mechanism for recording and investigating incorrect suggestions. Managers must understand that the confident, articulate answer is often the one most likely to escape casual challenge, especially from overworked teams.

The same scrutiny should apply to software updates. Real estate owners do not ordinarily accept a major modification to the building’s fire system without verifying that it functions. Hospitals should not accept a continually changing clinical algorithm without a defined plan for validation and monitoring. The FDA’s 2025 guidance on predetermined change-control plans provides a framework for certain AI-enabled device modifications, and its 2026 clinical-decision-support guidance addresses important software distinctions. Regulation is not a substitute for competent purchasing or internal quality control. It is a floor. A hospital should know what version it uses, when it changed, what effects were anticipated, and whether actual results match expectations.

Another issue is concentration of market power. As medical organizations deepen their reliance on AI vendors, they may discover that switching providers is difficult. Data formats, workflow integration, proprietary models, specialized hardware, and training can produce expensive dependencies. In my field, landlords worry about a tenant who represents too much rental income; tenants worry about a single landlord who controls the only viable space. Hospitals should worry about mission-critical software controlled by a supplier whose prices, terms, or technology may change. Competition and interoperability matter for patient care because a trapped customer has fewer options when quality deteriorates.

It is tempting to argue that public policy should solve these problems through broad rules requiring disclosure, testing, and equal access. Some rules are necessary, particularly concerning data privacy, safety, and claims about clinical performance. But governments can also protect incumbents by making compliance so expensive that smaller innovators cannot compete. California has firsthand experience with well-intended requirements that accumulate into a costly maze. The objective should be clear duties and effective oversight, not the number of pages generated by a regulatory process. Officials should distinguish high-risk diagnostic interventions from routine tools that merely help organize information.

Privacy is especially complicated for genetics. An AI system that links symptoms to molecular findings may reveal something meaningful about a patient and the patient’s relatives. The information does not become less sensitive because computers can analyze it quickly. Consent must explain whether data can be used for model development, whether it leaves the health system, what protections apply, and what choices patients retain. The legal language should not become so broad that the patient effectively gives up meaningful control. A promising research project should be able to earn trust through clear rules rather than demand it on the strength of scientific prestige.

On rare diseases, the opportunity is substantial. Families can spend years obtaining inconclusive evaluations while medical literature grows faster than any specialist can read. Tools like DeepRare can rank candidate diagnoses by combining symptoms, sequence information, and cited sources; 2026 investigations have identified promising leads in previously unresolved childhood cases. But an algorithm’s ranked possibility may be wrong, clinically irrelevant, or insufficient to justify treatment. Confirmation and counseling remain indispensable. I would insist that any institution measuring success disclose how many leads were independently confirmed and whether a correct diagnosis changed management. A promising hypothesis is valuable research output, not automatically a finished clinical service.

Drug discovery is even more sensitive to promotional enthusiasm. Rentosertib, an AI-discovered candidate for idiopathic pulmonary fibrosis, produced encouraging preliminary clinical findings in a 2025 phase 2a study involving seventy-one people. That is an important milestone, but the trial’s size and twelve-week follow-up are not sufficient to establish durable benefit or large-scale commercial success. In pharmaceutical development, many seemingly promising candidates fail for reasons that become visible only with more patients, longer observation, or real-world use. AI may improve early-stage target identification and molecular design while leaving those later risks largely intact. Investors should price the distinction rather than pretend the use of algorithms makes biological uncertainty disappear.

The uncomfortable truth is that increasing scientific productivity could initially increase healthcare spending. More sensitive screens identify more cases. More capable genetics tools create demand for confirmatory tests and specialist consultations. Faster drug discovery might expand the pipeline of costly clinical trials. These developments can be desirable if they deliver better outcomes, but we should not confuse clinical expansion with immediate budget savings. An honest health system will decide which benefits justify expense and how to avoid displacing essential services. The standard should be value, not simple cost reduction or uncritical technological expansion.

Finally, the distribution of gains matters. Wealthy health systems may afford sophisticated tools, trained staff, and follow-up specialists, while clinics serving poorer communities are left with cheaper products, inadequate support, or none at all. The resulting inequity could be magnified if algorithms perform less reliably on underrepresented groups. Serious public policy should encourage validation across populations, reasonable interoperability, and payment for the labor required to convert diagnosis into treatment. It should not subsidize glamorous software purchases while allowing ordinary preventive care to disappear. The latter would turn an exciting innovation into another demonstration of who can afford to benefit first.

I expect AI to become an enduring part of medicine because selected applications already show worthwhile performance and the demand for skilled care exceeds available capacity. But the durable winners will not necessarily be the firms with the largest models or boldest announcements. They will be the organizations that integrate technology into real clinical pathways, document outcomes, manage liability, and deliver value to patients and payers. The public should demand evidence of that value. A building is not successful because a developer presents a beautiful rendering; it succeeds because tenants can use it, the structure is sound, and the economics hold. Medical AI deserves exactly the same distinction between a persuasive prospectus and a functioning enterprise.

Dublin, Ireland Correspondent

American arguments about artificial intelligence in medicine often proceed as though there are two possible positions: release the technology immediately because innovation must not be restrained, or impose so many restrictions that innovation cannot occur. From Dublin, where my work involves software compliance and international customers, that strikes me as an oddly provincial choice. Most countries want medical technologies that genuinely improve diagnosis and treatment. They also want a defensible basis for deciding which tools can be trusted, who is accountable when things go wrong, and whether sensitive patient information is being handled properly. These are not competing objectives in the simplistic way political slogans imply. When the law and the evidence are aligned, they can make adoption more predictable. When they are not, both innovators and patients suffer.

I am neither a doctor nor a European Commission official. I am a compliance analyst accustomed to discovering that a clever feature in a software demonstration can be difficult to sell into a regulated organization. Sometimes the rules are unnecessarily complicated. Sometimes the developer simply has not established what the product does. The distinction matters enormously in clinical AI because a claim about assisting a doctor is not interchangeable with a claim about making a diagnosis autonomously. Nor is a system that enhances an image necessarily the same kind of intervention as a conversational model offering treatment suggestions. One of the first things I would ask every vendor is to stop saying ‘our AI’ and specify the exact clinical purpose, user, patient group, and conditions under which the output should be trusted.

The strongest available studies show why such precision pays. The Swedish MASAI randomized trial assessed an AI-supported mammography workflow in a real organized screening program. Its 2026 results found that interval-cancer rates were noninferior to standard double reading, while sensitivity improved and specificity remained unchanged. The trial reported sensitivity of 80.5 percent with AI support against 73.8 percent for conventional reading. Those numbers describe a tested arrangement, not a universal property of artificial intelligence. A different model, screening population, staffing pattern, or mammography service might not produce exactly the same outcome. But the study supplies something more useful than a grand technological claim: evidence that a particular clinical workflow can perform at least as well on important measures while reducing reading burden.

Germany’s PRAIM project offers a complementary form of evidence. It followed more than 463,000 women in real-world screening and reported higher cancer detection among AI-supported readings. That is encouraging, particularly because healthcare systems face shortages of specialist time. Yet the radiologists chose whether to use the system, making its observational design important when interpreting the findings. The German and Swedish studies should be discussed together precisely because their methods differ. Random assignment provides protection against some selection effects; real-world observational deployment tells us something about operational feasibility. Neither, in isolation, establishes that every detected abnormality improves survival. Careful readers should know what question a study answers before borrowing its figures for another argument.

Regulators have a legitimate role in making these distinctions visible. The European framework involves medical-device requirements, data-protection law, and the EU AI Act’s rules for systems classified according to risk, with applicable obligations depending on the system and implementation timetable. It is unwise to describe this as one neat approval stamp that settles every question. A provider still needs to determine the device’s intended use, whether conformity requirements apply, whether personal data are processed lawfully, and how the organization will oversee the tool in actual practice. The United States relies on a different mix of FDA device authority, clinical practice, and privacy requirements. What matters to me is that each jurisdiction should state obligations plainly and avoid creating contradictory compliance demands.

The United States has already authorized a large number of AI-enabled medical devices. The FDA’s own account exceeded 1,600 by September 2026. That is a meaningful sign of commercialization and regulatory activity, but authorization is specific to intended use; it does not promise that all devices have proven superior patient outcomes. American debate sometimes swings between treating FDA authorization as the final word and dismissing regulation as an obstacle to useful innovation. Both instincts miss the point. A regulator’s decision can establish that particular evidence meets a particular legal standard. A hospital still has to decide whether the device improves care in its environment and whether people will use it correctly.

The distinction between deterministic software and adaptable models is a major concern. A device approved in one configuration may behave differently after a substantial change. Developers understandably want to update models as new information becomes available, but a clinical product cannot improve itself according to a vendor’s private judgment without an accountable change process. The FDA finalized guidance on predetermined change-control plans for AI-enabled device software in 2025. European conformity and quality obligations similarly require thought about software modifications and risk management, even though legal details differ. A responsible purchaser should demand a version history, validation evidence, an explanation of important changes, and the ability to investigate an adverse event against the precise software configuration used.

The clinical evidence also shows that technical competence and human benefit must be distinguished. A randomized diagnostic-reasoning study found that granting physicians access to a large language model did not significantly improve their performance compared with conventional resources, despite the model’s impressive stand-alone results. This is not an argument that physicians cannot use technology. It indicates that the interaction itself must be evaluated. How does the tool present uncertainty? Does it encourage the user to verify evidence? Is its answer likely to anchor a diagnosis prematurely? Can practitioners contest its suggestion without fighting the interface? Many software systems fail because designers assume that good functionality automatically produces good human behavior.

An earlier randomized vignette study demonstrated the opposite of the desired effect when clinicians encountered systematically biased AI advice. Their accuracy declined, and explanations did not reliably prevent the problem. That finding deserves particular attention from organizations that advertise explainability as a comprehensive safeguard. A model can offer an understandable explanation and still be wrong. Transparency about the steps or factors behind a conclusion is valuable, but it does not substitute for representative data, sound validation, and the professional capacity to recognize discordant clinical evidence. A patient does not become safer because an erroneous recommendation includes a paragraph describing itself.

Data protection is more than a standard privacy notice. Medical information can reveal pregnancy, inherited disease, mental health history, and conditions people reasonably regard as intimate. Genomic findings may affect relatives who never consented to the original analysis. Developers need lawful grounds for processing data, appropriate security, careful retention policies, and clarity about whether personal information is reused for training. Cross-border processing creates further questions about access, contractual safeguards, and the consequences of service-provider failure. In my experience, the most dangerous phrase in a contract is often a vague promise that data may be used ‘to improve services.’ In healthcare, patients deserve to know exactly what that includes.

Europe should also be honest about its own weaknesses. Complex compliance requirements can consume money that might otherwise be spent on trials, engineering, or service. A large incumbent can maintain legal and regulatory teams that a small clinical start-up cannot afford. If rules become inconsistent or impossible to interpret, governments may entrench existing suppliers while claiming to protect consumers. The objective should be a risk-proportionate system in which a narrow low-risk aid does not face the same practical hurdles as a tool that directs life-altering treatment. A sensible regulator can be strict about evidence and clear about process simultaneously. The two qualities reinforce each other.

Rare-disease genomics demonstrates both the importance of cross-border research and the hazards of rushing interpretation. A child with unexplained symptoms may undergo sequencing that identifies numerous variants of uncertain significance. Tools such as DeepRare can compare clinical features and genetic findings with medical literature and rank possible diagnoses. A 2026 project reviewing previously unresolved childhood cases also produced candidate leads for further investigation. These systems may help clinicians revisit cases when scientific knowledge changes. But a computational ranking is not equivalent to independent clinical confirmation, and the relevant biological evidence may remain limited. Cross-border collaboration can make research stronger if records, terminology, consent, and laboratory methods are sufficiently compatible.

The distinction is especially important for families seeking an answer after years of uncertainty. It is possible to help someone enormously by showing where qualified clinicians should look next. It is also possible to do harm by conveying speculative genetic associations as established causes. The right system should provide traceable references, convey uncertainty, and support specialist review. It should record when an interpretation was reached, based on which version of a database or model, and how later findings may change it. These details may sound bureaucratic, but they are how a family can obtain a reliable second opinion rather than being trapped inside a black box.

Drug discovery invites a separate set of promises. In 2025, a randomized phase 2a trial evaluated rentosertib, an AI-discovered compound being investigated for idiopathic pulmonary fibrosis. It reported encouraging but preliminary findings among seventy-one participants over twelve weeks. The result demonstrates that AI-supported research can produce candidates capable of meaningful human testing. It does not mean the treatment has established long-term efficacy or received approval for routine use. European and American regulators should continue to evaluate clinical benefit, safety, and manufacturing quality without granting special credit because software helped identify the molecule. The body does not care how attractively a candidate was generated; it responds to the actual chemical.

There is an understandable desire to accelerate pharmaceutical development because patients with serious diseases may not have years to wait. Faster candidate generation can be important. So can more efficient trial recruitment, literature analysis, and detection of harmful interactions. But accelerated research must be distinguished from shortened proof. Clinical trial endpoints, duration, sample selection, adverse-event monitoring, and independent review are not decorative obstacles. They are protections against the considerable number of promising scientific ideas that ultimately fail in patients. We can improve the trial process without pretending to abolish uncertainty.

One of the questions I would like European and American health systems to ask together is whether outcomes can be compared in a credible way. A vendor may report a strong result in Sweden, a health authority may commission a study in Ireland, and a hospital in the United States may publish a proprietary dashboard. If their measures and definitions differ, purchasers will struggle to know which result transfers. Shared reporting standards, structured documentation of the intended use, and clear subgroup analyses would be more helpful than another round of competing declarations about who leads the AI race. Competitive research can coexist with common expectations about evidence.

There is a broader democratic issue as well. When patients are told that AI has influenced their diagnosis, they should know whom to question and how to seek review. A doctor cannot reasonably be asked to validate every line of code, but neither should a hospital be permitted to blame an inaccessible vendor after an error. Responsibility must be allocated among developers, deployers, and clinical professionals according to what each controls. Incident reporting should permit investigation without encouraging concealment. The law should allow complaints to be heard and lessons to be learned. Above all, systems ought to preserve meaningful clinical discretion rather than using contractual or technical design to make automated recommendations effectively compulsory.

Low-resource settings also matter to a European conversation. A highly effective tool can widen the gap between patients if it requires expensive connectivity, standardized scanners, specialized training, and data infrastructure that many clinics lack. Models may perform poorly where diseases, languages, or imaging conditions differ from the development population. The World Health Organization has warned about bias, automation dependence, privacy, and the unequal distribution of advanced AI tools. Those are not hypothetical objections devised by cautious Europeans; they are practical requirements for making a technology generalizable beyond elite institutions. A product that only works under privileged conditions must be described honestly as such.

My position is that Europe should neither celebrate caution as a virtue in itself nor abandon governance in pursuit of American-style speed. The United States should not assume that every regulatory difference is evidence of fear, and Europe should not assume that writing elaborate rules will produce safe software. We should compete on evidence, clinical skill, and delivery, while cooperating on safety, research integrity, and interoperability. The promising findings in mammography warrant serious adoption; the weaker evidence for unrestricted diagnostic chatbots warrants restraint; the potential in genomics and drug discovery deserves continued research without exaggerated claims. Progress in medicine has never been guaranteed by either a laboratory breakthrough or a statute. It comes when useful discoveries survive careful testing and become trustworthy care for actual patients.

Johannesburg, South Africa Correspondent

When a health-technology executive says artificial intelligence will democratize expert medical care, I am inclined to ask which clinic, on which network, with which electricity supply, and at what price. I live in Johannesburg and work with operational risk and energy markets. I have watched institutions announce a technological leap while basic systems falter in public. I do not mean that sophisticated technology has no place in countries struggling with infrastructure. Quite the opposite: if a capable diagnostic tool helps a doctor in an underserved community, its value may be greater than in a wealthy hospital surrounded by specialists. But the benefits do not materialize merely because a corporation has produced a powerful model. The more distant patients are from reliable laboratories, transport, skilled personnel, and follow-up care, the more important the rest of the system becomes.

AI in medicine is often discussed as though clinical intelligence were a scarce commodity that can simply be shipped across the world in software. Expertise is scarce, but it is not the only thing that is scarce. Reliable medical records may be absent. Blood tests may take too long. A patient may not afford repeated travel. An imaging machine may be available only at certain facilities. Electricity, communications, and equipment maintenance can be uneven. In some places, a specialist opinion is available while the necessary treatment is not. If AI identifies a dangerous condition and there is no affordable pathway to intervention, the technology has discovered the location of a problem without solving it. We must not confuse knowledge with capacity.

There is legitimate reason for optimism. AI-supported breast screening has moved beyond laboratory claims. The German PRAIM project studied more than 463,000 women and observed higher rates of detected cancers with AI assistance. The Swedish randomized MASAI trial enrolled roughly 106,000 women and reported improved sensitivity, preserved specificity, and a noninferior rate of interval cancers under AI-supported reading. The findings are important because they concern actual screening operations. If imaging expertise can be used more efficiently without reducing clinical quality, countries with shortages of radiologists may eventually benefit. But the word eventually matters. These were structured programs, not improvised deployments in clinics without consistent imaging, pathology, or recall systems.

Transferability should be investigated rather than assumed. An algorithm trained on high-quality mammograms obtained with particular equipment may encounter different scanners, breast-density distributions, patient ages, disease patterns, or image-acquisition practices elsewhere. A health authority should not be asked to buy a promise that was established under unrelated conditions. Local validation can identify whether performance is materially different and whether adjustments are needed. I would be interested in how the system behaves when equipment is old, images are imperfect, records are incomplete, or a patient misses a scheduled screen. Those are not marginal situations in much of the world. They are daily clinical reality.

The experience with autonomous diabetic-retinopathy screening shows both the possibility and the limits. A small 2026 implementation study in an endocrinology setting found that an authorized system could provide definitive results for most patients using images collected by a non-ophthalmic operator. This is potentially valuable where ophthalmologists are scarce. But many patients with positive findings lacked documented institutional confirmation, and the investigators could not count outside follow-up. That problem can be even more severe when people live far from referral centres or cannot afford the journey. Governments should measure the number of people receiving appropriate treatment, not merely the number who were screened and told to seek another appointment.

I know something about systems that look reliable in a dashboard but are not reliable for the people who depend on them. An electrical grid can report capacity while a particular distribution feeder repeatedly fails. A bank can publish risk controls while customers encounter one unavailable service after another. The same distinction applies to AI-assisted medical care. A national authority may buy licenses for thousands of clinics, but if the clinics cannot maintain devices, protect records, or follow up on suspicious results, the initiative becomes a procurement achievement rather than a healthcare achievement. The measure should be service delivered under actual conditions, including on the least convenient day and at the least well-equipped facility.

General language models introduce a different risk. They can produce fluent, organized summaries of clinical literature and propose plausible diagnoses. Fluency may be helpful for education and for clinicians who need an entry point into an unfamiliar subject. It can also manufacture confidence. A randomized study of physicians found that access to a large language model did not significantly improve their diagnostic reasoning compared with access to usual resources, even though the model performed strongly when evaluated on its own. That suggests human-machine collaboration remains a design and training problem, not an automatic consequence of combining two intelligent actors. The conclusion is not that people or machines are incapable. It is that a workflow can erase theoretical advantages.

More troubling still, researchers found that systematically biased AI advice could reduce clinician accuracy and that explanatory text did not reliably correct the harm. This is a direct challenge to the idea that a model is safe if it can explain itself. The question is whether its explanation rests on reliable evidence, whether the model performs consistently across populations, and whether the clinician has enough independent information to reject it. In resource-constrained settings, a clinician may have fewer opportunities to verify an answer with immediate laboratory or specialist support. The danger of a confident but false suggestion may therefore be amplified precisely where the promise of AI assistance is greatest.

Healthcare data also reflect a world’s uneven patterns of research and wealth. Some diseases, ancestries, languages, and clinical environments are heavily represented in development data; others are barely visible. A genetic model evaluated primarily on families who resemble those in the development cohort may be less useful for populations whose variation is underrepresented. Even ordinary laboratory reference ranges and documentation practices can differ. If the resulting software is sold globally without scrutiny, existing inequalities become embedded in supposedly objective recommendations. The World Health Organization has warned of bias, missing information, automation dependence, privacy risks, and inequitable deployment. These are operational warnings, not merely philosophical objections.

Consider the excitement around AI-assisted genetic diagnosis. Tools such as DeepRare have demonstrated the ability to assemble symptoms, genetic data, and medical references into ranked diagnostic hypotheses across difficult rare-disease cases. Other studies have found new leads by revisiting previously unsolved records with advanced reasoning systems. This could be invaluable to families whose children have lived with unexplained conditions for years. But it is not a substitute for a reliable sequencing service, proper sample identification, a genetics professional, confirmatory testing, and sensitive counseling. A model can suggest a variant associated with disease while the causal interpretation remains uncertain. Offering a family false certainty can do real harm, especially where accessing a second opinion is difficult.

I would also ask who owns the data and who captures the commercial value. African populations have often been insufficiently represented in biomedical research, yet their clinical and genomic information is potentially valuable in building more comprehensive systems. Collaboration can benefit patients and science, but terms should be clear. Patients should understand what information is collected and how it may be used. Local institutions should be able to examine performance and share in legitimate research benefits. The worst arrangement would be for communities to supply data, receive products they cannot afford, and be told that any questions about governance amount to hostility toward innovation.

Drug discovery is another area where promises can travel faster than evidence. A 2025 phase 2a randomized trial tested rentosertib, a compound identified through AI-enabled target discovery and molecular design, for idiopathic pulmonary fibrosis. The trial enrolled seventy-one people and produced encouraging short-term findings that require larger, longer studies. I welcome that. A quicker path from biological hypothesis to human testing could save time and eventually money. But the biology of disease is not abolished by a faster computational search. An experimental candidate can fail in a larger population, produce delayed adverse effects, or prove economically inaccessible even if it works. The history of pharmaceutical research is full of reasons to distinguish a promising candidate from a proven treatment.

A rich country may reasonably emphasize whether AI reduces hospital staffing burdens or increases the productivity of an expensive specialist. Elsewhere, a more important question may be whether a trained non-specialist can safely provide a service that was previously unavailable. Autonomous retinal screening and properly supported imaging interpretation deserve special consideration on that basis. But such arrangements require clear referral protocols and escalation routes. A system that classifies uncertain cases must be able to say it does not know, and the clinic must have somewhere to send them. A product that refuses to express uncertainty is not more advanced. It is less suitable for safety-critical work.

The financing arrangement also needs attention. Technology companies may offer favorable pilot prices that become difficult to maintain after dependency has developed. Software may require continual cloud connectivity, international payments, licenses priced in foreign currency, or support supplied from another continent. Governments should examine the long-term cost of equipment, bandwidth, updates, cybersecurity, training, and replacement. They must know what happens if the vendor ceases service or raises prices sharply. A pilot program underwritten by a foundation can be admirable, but a ministry must decide whether it can sustain the program at national scale without diverting money from vaccinations, primary care, or other essential services.

The relationship between algorithmic performance and political accountability deserves scrutiny. Governments sometimes embrace a digital-health project because its launch is easier to photograph than improvements in staffing or public administration. The technology then becomes a shield: officials refer to a sophisticated new platform whenever patients ask about long queues and unavailable medicines. That temptation is not unique to South Africa, and it would be unfair to treat it as an African weakness. Every public system has incentives to announce visible projects. Good governance demands an independent evaluation that can conclude a project did not deliver what was promised without threatening the evaluator’s career or budget.

There is no sound reason for physicians to regard machines as rivals in a contest of prestige. A clinician who receives a useful alert about an abnormal image or a possible genetic condition may serve the patient better. The correct arrangement preserves the clinician’s ability to evaluate evidence, communicate uncertainty, and decide what care is appropriate. Nor should the technology be dismissed simply because some institutions want to use it irresponsibly. Shortages of skilled professionals are real. An affordable, validated tool that helps an underserved clinic can be an unusually effective public investment. The burden is on buyers and sellers to demonstrate that it works where they intend to place it.

Training should therefore include more than a simple tutorial. Operators need to know how the system handles poor inputs, which groups were studied, what error rates mean, what a borderline result requires, and how to report an unexpected outcome. Clinicians must learn to challenge fluent errors and recognize when clinical evidence contradicts an algorithm. Administrators must provide time to review near misses rather than punishing staff for slowing down deployment. Patients should have an understandable explanation of what the tool did and whether a human clinician reviewed its output. All of that takes money and institutional discipline. The claim that AI eliminates the need for such work is an advertisement, not a plan.

A sensible outcome framework would follow patient journeys in different environments. One should test a device in a well-equipped urban referral centre, a busy township clinic, and a rural service with more limited resources, adjusting evaluation to the intended indication. Report diagnostic accuracy and the proportion of cases for which the system cannot produce an answer. Then report completed referrals, time to intervention, treatment complications, patient understanding, and total cost. Publish performance by relevant subgroups and investigate disparities. Compare the system with existing care rather than with an imaginary world in which every patient sees a specialist immediately. Only then can decision-makers assess whether the technology reduces or expands inequity.

The most serious long-term opportunity may be cooperation between research institutions and health systems across regions. Shared protocols, independent trials, suitable training data, and local clinical oversight could make useful tools both more generalizable and more affordable. That is better than exporting finished products and demanding trust. Companies may find that designing for imperfect infrastructure also improves resilience in wealthy countries during disasters, outages, or staffing shortages. Innovation does not flow in only one direction. A technology tested carefully under difficult conditions may reveal weaknesses that glossy metropolitan demonstrations missed.

I am optimistic about the power of computation to help identify disease, organize evidence, and accelerate research. I am skeptical of the institutional habit of treating the purchase of technology as proof of public service. The strongest imaging trials deserve attention, the genetics advances deserve disciplined follow-through, and AI-assisted drug development deserves patient, rigorous testing. None deserves exemption from ordinary questions about safety, affordability, and access. The future will not be determined solely by the size of a model trained in California, London, or Beijing. It will also be determined by whether a patient in Johannesburg—or in any clinic far from a great medical centre—can obtain a correct result, receive the necessary care, and go home healthier. That is an outcome worth measuring.

Bismarck, North Dakota Correspondent

Distance changes the meaning of a medical appointment. In a large metropolitan area, a specialist consultation may be inconvenient, expensive, or difficult to arrange. Across much of the northern Plains, it can also require a long drive, a day away from work, an overnight stay during bad weather, and someone to look after children or livestock. The cost is not only what appears on a hospital statement. It is the time and effort needed to reach care. That is why I take the promise of artificial intelligence in medicine seriously. A dependable system that helps a local clinician recognize an urgent condition, interpret an image, or identify a possibility worth referring could make an enormous difference to rural families. But I also know from energy compliance and agricultural lending that a technology promoted from a distant capital often looks better on paper than it does after a North Dakota winter has tested the assumptions.

I am not trained in medicine. My work has taught me to ask whether a procedure functions in the place where it will be used, who is responsible when it fails, and whether the people affected can obtain an answer. In farming and energy, a project can satisfy every milestone on a presentation while never achieving dependable performance in the field. The healthcare equivalent would be an AI screening system that identifies abnormalities but leaves people unable to reach treatment, or a diagnostic assistant that offers uncertain advice with more confidence than the evidence warrants. Rural patients do not need the most fashionable tool. They need one that works despite imperfect connections, limited staffing, unusual hours, and the ordinary unpredictability of life.

Some medical AI applications now have real-world evidence worth considering. Germany’s large PRAIM mammography study found higher cancer-detection rates with AI-supported readings than without them. The Swedish MASAI randomized trial reported better sensitivity and unchanged specificity for an AI-supported mammography workflow, while its interval-cancer outcome met the prespecified noninferiority standard. To me, those findings provide a serious reason to examine whether assisted screening could improve access or capacity in rural programs. If radiologists can review studies more efficiently without losing important findings, clinics may get results back sooner and specialist time may reach more patients. That is better than promises that computers will simply replace experts.

The details matter. Neither study proves that every rural clinic should install whichever product a vendor offers. Performance can change with equipment, images, screening populations, and clinical procedures. Some patients may have characteristics not adequately represented in the development or validation samples. A screening program must maintain a way to investigate uncertain readings and communicate results promptly. Most important, there must be a practical route to additional imaging, biopsy, or treatment when something concerning is found. A system that makes an earlier discovery but cannot connect the patient to help may deliver anxiety rather than meaningful improvement.

Autonomous diabetic-eye screening is another promising example. People with diabetes need appropriate eye assessment because retinal damage may develop before symptoms are obvious. A 2026 study showed that an AI system could produce definitive screening results for most participants in an endocrinology clinic, even with a novice operator. In rural areas, bringing a useful screening capability closer to patients could reduce travel and uncover disease that otherwise goes unnoticed. Yet the same study illustrated the follow-up problem: fewer than half of AI-positive participants completed confirmatory examination at the institution, with outside care not tracked. That is not a reason to discard the screening tool. It is a reason to fund the person who arranges the specialist visit and to measure whether the patient actually attends.

We often speak as though every rural barrier can be solved by telemedicine. It helps, and I would not want to lose it. But a video visit still depends on broadband access, privacy, equipment, and someone who can perform whatever physical assessment is necessary. An algorithm does not draw blood, stabilize an emergency patient, or drive an ambulance over icy roads. It may improve the decision about where and how urgently a patient should be sent, which is valuable in itself. The danger is that policymakers will use the existence of technology as an excuse to neglect the clinical workforce, roads, emergency services, and local facilities that make appropriate care possible.

The sepsis prediction debate makes the point sharply. An updated sepsis model was prospectively evaluated across four American health systems in 2026. Its performance differed by institution, and even at a common sensitivity target, positive predictive values varied substantially. The model can provide information about deteriorating risk, but an alert is not a diagnosis or a treatment. Rural facilities may face difficult choices about which warnings require immediate action and when to transfer a patient. False positives can occupy scarce attention and beds; false negatives can be catastrophic. Local validation and response protocols should be developed with the clinicians who must act, rather than imposed as an administrative metric.

The same is true of general diagnostic chatbots. A 2024 randomized study found that access to a large language model did not significantly improve physician diagnostic reasoning compared with conventional resources, although the model performed well when assessed alone. That result should discourage the notion that a remote clinic can replace its professional judgment with a conversational program. The machine might help a clinician consider a less common possibility or quickly review relevant evidence. It should not be mistaken for a physician who can examine the patient, understand the community, and take responsibility for the next decision. A capable second opinion is useful. An unaccountable first and final opinion is something else.

Bias matters in ways that are not always obvious. Data drawn mainly from large academic centres may omit populations with different disease histories or access to follow-up. A patient who delays care because the nearest specialist is hours away may present differently from someone screened regularly at an urban medical centre. Documentation is sometimes less complete when medical records are divided among independent facilities. An algorithm that treats missing information as evidence of low risk could produce dangerous reassurance. Another that treats every unusual pattern as urgent could overwhelm a clinic. Those possibilities call for testing and monitoring rather than automatic assumptions that the computer is more impartial than a human being.

A randomized study of biased AI predictions found that such advice could make clinicians less accurate even when explanations were attached. This is important for small facilities where professionals must switch tasks constantly. A system may offer a plausible explanation that encourages staff to accept its recommendation when the patient or the clinical record suggests otherwise. Training must teach clinicians how to question the tool, not merely how to operate it. A nurse or physician should feel entirely free to report that a model appears wrong, and the hospital should have a process for examining repeated concerns. If managers treat challenges as resistance to innovation, they will silence the people best positioned to prevent harm.

Genetic diagnosis offers hope to families who have traveled from specialist to specialist without answers. New systems can combine descriptions of symptoms, sequencing results, and medical literature to rank candidate rare diseases. DeepRare and other recent efforts show that AI can identify useful possibilities in difficult cases, sometimes after earlier evaluation was inconclusive. For a family in North Dakota, the ability to reach specialized analytical expertise without repeated cross-country journeys could be a meaningful advance. Yet a promising lead does not remove the need for properly collected samples, expert interpretation, confirmatory testing, or counseling. One incorrect label can affect treatment decisions and family planning. We must be willing to say that a finding is uncertain.

I think the drug-discovery promise is genuine but farther from ordinary care than many headlines suggest. The AI-discovered compound rentosertib reached a randomized phase 2a study for idiopathic pulmonary fibrosis, with encouraging preliminary signals from seventy-one patients over twelve weeks. That is substantial scientific progress. It does not establish the drug’s long-term safety or efficacy, much less an affordable therapy available at the local pharmacy. Rural people who already travel long distances for complex treatment may reasonably wonder whether future innovation will benefit them or only the wealthy centres participating in initial trials. Public discussion should distinguish laboratory milestones from accessible medical products.

Payment systems can strengthen or weaken adoption. A rural practice may not have enough patients to spread software costs across a large volume of visits. The price of devices, training, connectivity, technical support, and necessary staff time may outweigh narrow savings. Conversely, a tool that prevents a costly late diagnosis or unnecessary transfer could offer social benefits greater than the clinic’s immediate financial return. Insurers and public programs should evaluate these possibilities rather than demand that every small facility bear the full cost of innovation whose gains are shared across the system. But public support should come with evidence and oversight, not just a persuasive application for a grant.

Maintenance deserves far more attention. In energy work, the question is never merely whether a system ran well at commissioning. We ask how it behaves after years of use, during unusual conditions, and when the original technicians are gone. Medical AI can change after a software update, a scanner replacement, or a shift in patient population. A rural hospital might lack specialized personnel to audit every change. That makes clear vendor obligations, independent testing arrangements, and regional partnerships especially important. The hospital needs a reliable way to obtain support, contest incorrect output, and revert to a safe process when the technology is unavailable.

Data security is also a concrete concern. Small facilities should not be expected to send sensitive medical and genetic information into poorly understood platforms. Patients must know whether outside companies receive their data, what safeguards exist, and whether information will be reused for model development. Rural communities may be close-knit, which can make privacy particularly important. A diagnosis that is unusual within a small town may be identifying even when obvious personal details are removed. Systems should minimize unnecessary collection and provide serious oversight of vendor access. A person should not have to choose between keeping private information private and obtaining competent healthcare.

The political argument around this technology has become too easy. Some people assume only a large national bureaucracy can establish standards and guarantee fairness. Others believe any standard is an obstacle and the market will sort out the risks. Both positions leave rural patients exposed if they are treated as an afterthought. Government should demand truthful claims, enforce patient privacy, and support independent evidence. Healthcare systems should maintain clinical judgment and measure local outcomes. Vendors should be responsible for defined performance and responsive support. Communities should have the information necessary to decide which applications justify their cost. This is less theatrical than a national AI strategy, but more likely to help a family seeking care.

I would like to see regional evaluations that compare practical alternatives. Does AI-supported screening help a clinic serve more patients than adding a visiting specialist? Can it reduce the time between an abnormal image and appropriate follow-up? Does it help keep care local safely, or does it increase unnecessary transfers? What is the effect on missed diagnoses, staff workload, and total patient travel? Where possible, outcomes should be measured by age, relevant medical history, geography, and other factors that reveal uneven performance. A national average can hide a rural failure just as an average statewide rainfall hides the conditions on a particular farm.

There is a cultural question, too. Many rural people appreciate technological progress and do not need to be lectured about accepting change. Farmers use sophisticated equipment, energy companies deploy advanced analytics, and local businesses depend on modern communication systems. What people resent is being told that a tool they had no role in choosing has solved a problem they continue to live with. Trust grows when developers listen to clinicians and patients before deployment and make adjustments when the system encounters local conditions. It disappears when complaints are dismissed as ignorance or nostalgia.

Ultimately the strongest argument for AI in medicine is that it might extend competent help to people who have too little of it today. The imaging evidence is encouraging. Screening closer to home could make a difference. Genomic interpretation may shorten painful diagnostic journeys. Drug discovery may eventually create therapies that did not exist. I support pursuing those opportunities with enthusiasm disciplined by the facts. But the test on the Plains should remain plain: does this actually help our doctors and nurses get patients to the right care, reliably and without introducing greater risks? If it does, the technology belongs here. If it merely adds another screen while the patient still has to make the same long drive without a clear answer, then we have purchased an expensive way to tell ourselves we are modern.

Cedar Rapids, Iowa Correspondent

When somebody places a percentage in front of me, I want to know what went into the numerator, what went into the denominator, and who chose the date on the report. I spent my working life as a plant accountant, and that habit is hard to break. A production manager can improve the reported cost per unit by excluding rework, assigning expenses to another department, or selecting a particularly flattering week. Eventually the actual bills come due and the books tell the story. The current discussion about artificial intelligence in medicine is full of figures that may be accurate in their narrow context but are being used to support much larger claims. A model can classify medical images accurately; that does not necessarily establish how many patients live longer. An AI-assisted program can find more suspicious lesions; that does not tell us how many unnecessary interventions follow. Before declaring a new medical era, we ought to reconcile the numbers.

My objection is not to technology. I worked in manufacturing long enough to appreciate a well-designed improvement. Machines can perform repetitive checks more consistently than tired people, and good software can find patterns that would be difficult to notice manually. I am also old enough to remember the management programs that promised revolutionary productivity and delivered mostly consultant bills. The difference was always measurable performance. If an investment really made our operation safer, more productive, or more reliable, it appeared in the records over time. Healthcare deserves the same discipline, perhaps more, because the consequences of accounting for the wrong thing are not limited to money.

Take the much-discussed breast-screening evidence. The German PRAIM project involved more than 463,000 women and reported cancer detection of approximately 6.7 per thousand screenings with AI support, compared with 5.7 per thousand without it. That is an additional detected cancer per thousand women screened, and the adjusted relative increase was 17.6 percent. Both statements are true; they simply describe the effect differently. If somebody tells a public audience that cancer detection improved by nearly eighteen percent, people may envision an enormous number of additional lives saved. The actual absolute finding is more modest and should be considered alongside the nature of the cancers found, the subsequent investigations, and long-term outcomes. That does not make the result unimportant. It makes it possible to understand.

The Swedish MASAI randomized trial provides another useful lesson in reading results. Its 2026 analysis found screening sensitivity of 80.5 percent in the AI-supported group versus 73.8 percent with conventional double reading. Specificity was 98.5 percent in both groups, and the trial met its noninferiority criterion for interval-cancer rates. This is substantial evidence that an AI-supported approach can improve certain screening measures without a corresponding loss in specificity under the tested conditions. But notice the limits of the claim. The trial examined screening performance and related outcomes. It did not directly demonstrate that patients ultimately live longer because the software was used. A responsible analyst distinguishes an improved intermediate measure from an established final outcome.

That distinction is not an invitation to postpone every innovation until decades of mortality data accumulate. We often make decisions using the best evidence available, especially when a new approach improves detection and reduces professional workload without an obvious offsetting harm. The decision should simply state its assumptions. If we believe earlier detection will improve health, identify why that belief is reasonable and how it will be tested. If we expect fewer human reading hours to justify the purchase, measure actual labor use and quality afterward. If additional testing will impose cost, include it rather than pretend the algorithm’s output is the end of the transaction. We can act prudently under uncertainty without concealing that uncertainty.

The denominator problem appears repeatedly in screening. Suppose an algorithm identifies a larger fraction of patients who may have a disease. An announcement may emphasize sensitivity, the share of true cases detected. Yet specificity, the share of disease-free people correctly screened negative, matters too. So does disease prevalence. In a population where the condition is uncommon, even a generally accurate system can produce many false-positive alerts relative to true positives. Clinicians understand this, but public discussions often omit it. An absolute count of alarms may be far more relevant to staffing and patient anxiety than a figure describing the model’s rank-order discrimination. I wish more press releases would include a simple example with actual patient counts.

Sepsis prediction provides a particularly clear case. A 2026 multicenter prospective study of an updated Epic model reported meaningful variation across four health systems. At a threshold set to achieve 60 percent sensitivity, the positive predictive value varied from about 13 percent to 26 percent. These figures should not be used to declare the algorithm worthless. Positive predictive value depends on prevalence and the population being tested. But they demonstrate that the hospital receiving an alert may have to investigate a substantial number of patients who do not meet the study’s sepsis definition. A program budget must account for those evaluations and the time they consume. The financial model should also consider the cost of missed sepsis cases and any benefits from earlier treatment. The calculation is a balance, not a slogan.

One vendor may claim its system reduces clinician time. Another may point to increased detection. A hospital administrator may project fewer complications. I would ask whether the claims come from the same patient population and the same period, and whether one improvement generates additional expense elsewhere. If a more sensitive tool leads to extra biopsies, specialists, follow-up visits, and patient transport, those are not necessarily failures. They may be justified by better health. But they must be counted. A manufacturer that reports a new machine’s output while charging all maintenance to another department has not presented a fair comparison. Healthcare programs should not be permitted to do the equivalent merely because the technology is sophisticated.

I would also like to see a careful comparison of AI assistance with sensible alternatives. Perhaps an organization needs more radiologists, better imaging equipment, improved appointment reminders, or an effective system for recalling patients after abnormal results. A new algorithm might be the best investment, but the only way to know is to examine what is actually limiting performance. A clinic that detects disease accurately and cannot obtain confirmatory appointments may gain little from detecting additional abnormalities unless it simultaneously expands its referral capacity. A technology assessment should include the best realistic alternative use of the money. That is how responsible businesses and households make investment decisions.

The autonomous diabetic-eye screening example brings this home. A 2026 clinical implementation study reported definitive AI screening results for most participants, but many patients with positive results did not have documented follow-up within the institution. External visits were not captured, so the figure cannot be interpreted as proof that care was never obtained. It does show that a screening program can be evaluated only partially if it tracks the computer’s output but not the subsequent clinical journey. The most important missing entry in the ledger may be the number of patients who received timely treatment capable of preserving vision. A well-run system should aim to capture that result and describe the limits of its records.

Diagnostic chatbots present a different bookkeeping challenge. A 2024 randomized study gave physicians access to a large language model during diagnostic-reasoning exercises. Their performance did not improve significantly compared with physicians using conventional resources, though the model alone performed well. This creates three separate questions: how well can the software answer a case, how well can a clinician answer it, and what happens when the clinician uses the software? The third is the practical purchase decision. It is a mistake to buy a team-based system on the strength of stand-alone software benchmarks. In manufacturing, the fastest machine in isolation was not always the machine that improved the whole production line.

The danger is not solely disappointing productivity. Another randomized vignette study found that erroneous biased AI advice could lower clinicians’ diagnostic accuracy, and explanatory text did not reliably protect them. In accounting terms, a plausible-looking reconciliation can be worse than an obvious discrepancy because it discourages investigation. If a medical system offers a persuasive wrong answer, the safeguard cannot consist only of telling clinicians the computer may be wrong. The workflow must make contrary evidence visible and encourage professionals to verify important conclusions. Hospitals should record model-related near misses as carefully as they record other quality events, then examine whether corrective measures actually reduce the risk.

Rare-disease genetics illustrates why even high-ranking predictions require careful accounting. AI systems can search literature and compare molecular results at a scale no individual physician can match. Recent studies of DeepRare and other diagnostic reasoning tools report promising performance in identifying candidate conditions. But research statistics may count a diagnosis as correctly ranked when it appears among several suggestions, and that is not the same as a new clinically confirmed diagnosis delivered to a patient. The number of cases tested, the proportion with known answers, and the number of genuinely unresolved cases should all be reported. If a system generates leads for a family, I would want to know which leads were independently confirmed and what changed in treatment or counseling.

Drug discovery attracts investment headlines that can be even less useful than clinical benchmarks. Rentosertib, an AI-discovered compound for idiopathic pulmonary fibrosis, reached a randomized phase 2a trial published in 2025. Seventy-one participants were assigned across several groups, and the twelve-week study reported encouraging preliminary signals. That is a legitimate research milestone. It does not show that a marketed treatment has been created, or that the entire development process will be cheaper, or that other AI-generated drugs will succeed. A phase 2a signal is one entry in a much longer financial and clinical ledger. Investors may reasonably fund the next stage, but they should not treat development risk as already retired.

There is a frequent claim that AI will reduce healthcare costs simply by making work more efficient. It might. Yet improved technology can increase use. Faster interpretation allows more screening. More sensitive screening uncovers conditions that require evaluation. Better molecular understanding creates new therapeutic opportunities, often expensive ones. The result may be better medicine and higher total spending. Economists and accountants should not describe that as a paradox. When an activity becomes more productive, demand can grow, especially when the service is valuable and previously rationed by scarcity. The proper goal is health gained relative to resources consumed, not the simplistic assumption that every new machine must shrink the budget.

Hospitals also need to account for depreciation in an unusual form: software quality can change without the hardware wearing out. An updated algorithm may improve performance or alter it unexpectedly. A change in scanner equipment can affect data inputs. Staffing and patient characteristics can shift. A model that was appropriately validated during purchase may require fresh monitoring. The FDA’s guidance on predetermined change-control plans recognizes that some AI-enabled medical devices may be modified under defined processes. Purchasers should maintain records of version changes, testing, operating costs, and incidents. A plant would not accept a new control setting without verifying production results. A hospital ought to be at least as careful.

A financial assessment must also address liability. Who pays when an incorrect recommendation contributes to harm? A vendor may limit contractual responsibility while a hospital carries the consequences of a patient claim, professional discipline, or reputational damage. Insurance may cover some exposures but not every risk or investigation. These are not reasons to reject all AI. They are reasons to negotiate contracts with understandable warranties, reporting obligations, cybersecurity duties, and remedies. Organizations that cannot explain responsibility before a problem arises generally discover the answer under considerably less favorable conditions later.

Public funds should be monitored with equal care. Legislators may wish to subsidize AI tools at rural clinics or hospitals serving underserved communities. Such support can be justified when access benefits exceed the return any one institution can capture. But I would require clear baseline measures, realistic targets, and independent evaluations. Report the number of eligible patients, the proportion completing screening, confirmed diagnoses, treatment time, adverse consequences, and the complete operating cost. Do not allow a pilot to count the people it contacted as the people it treated. Do not count hoped-for savings as cash already recovered. And publish unfavorable results rather than allowing the next project to repeat the same mistakes.

Patients should receive enough information to make sensible decisions without being buried in technical statistics. A brochure can explain that a system assists with screening, that uncertain results require human assessment, and that an abnormal finding is not automatically a diagnosis. Health systems should be candid about false positives and false negatives. They should avoid claiming certainty when the evidence supports only a probability. People routinely make decisions involving risk when the information is explained clearly. They deserve the same courtesy in medicine that a careful lender should provide in finance: disclose what is known, the important limitations, and who bears the consequences.

My position is that medical AI offers real gains in some settings and plausible opportunities in others, but the story is still early. The imaging studies are substantial. Rare-disease tools merit further independent evaluation. AI-assisted drug research can make valuable candidates possible. General diagnostic systems and alerting tools require caution because combining a model with an actual healthcare organization changes the result. None of this calls for hostility toward technology. It calls for ordinary discipline: compare similar cases, identify the baseline, measure meaningful outcomes, count all costs, and inspect the result after the equipment has been running for a while. That is not the sort of advice that wins a technology conference. It is the sort that keeps the books honest when the conference is over.

Long Island, New York Correspondent

People sometimes ask lawyers and paralegals whether a new invention is legal, as though the answer will be found in one paragraph of a statute. Usually the more revealing question is what the invention does, who uses it, what promises were made, and what happens when somebody is injured. I have spent years around zoning files, assessments, and litigation, including time in a district attorney’s office. That is not the same as practising medical malpractice law, and I do not claim it is. But it does teach an enduring lesson: a disputed decision eventually needs a record, an authority, a responsible party, and a means of review. Artificial intelligence in medicine is advancing quickly enough that healthcare institutions should be answering those questions now, before a patient’s lawyer has to reconstruct them after something goes wrong.

I am not hostile to technology. A radiologist may benefit from a system that highlights a suspicious image. A physician treating a child with an unexplained condition may gain from software that searches rapidly through genetic findings and medical literature. Researchers may save considerable time finding useful drug candidates. The evidence is no longer confined to speculative demonstrations. What concerns me is the way organizations sometimes treat the words ‘AI-assisted’ as though they transfer accountability from a human professional to an impersonal product. They do not. Patients have rights, clinicians have duties, manufacturers make representations, and hospitals control operating conditions. The relationships can be complicated, but that is an argument for clear documentation rather than a reason to leave them vague.

The mammography evidence deserves a fair reading. Germany’s observational PRAIM study included more than 463,000 women and reported improved cancer detection when AI-assisted reading was used. Sweden’s MASAI trial randomized about 106,000 women and found higher screening sensitivity with AI support, unchanged specificity, and a noninferior rate of interval cancers. Those are important outcomes for a specific use of technology. Yet they do not establish that every diagnostic program should be automated, nor do they prove an overall reduction in cancer mortality. The legal and ethical obligation of an institution is to describe what its particular device has been shown to do, not to borrow the strongest result from an unrelated system and treat it as an industry-wide warranty.

The matter of representations can become significant in ordinary commercial law, quite apart from medical negligence. What does a vendor’s sales material promise? Does the user agreement state that the product is only a decision aid while the marketing presentation implies autonomous expertise? Which patient populations were included in validation? Is the hospital made aware of foreseeable false positives and false negatives? Are clinicians told which inputs may produce unreliable results? Organizations should retain those documents and reconcile them before deployment. An attractive demonstration should not be allowed to quietly enlarge the intended use beyond what was tested and authorized.

The FDA reports more than 1,600 AI-enabled medical devices authorized for marketing as of September 2026. That is a remarkable indication of how quickly the category has grown. It is not a blanket legal defense for every use of every product. FDA authorization is linked to applicable requirements and intended use; medical practice, hospital operations, contracts, and other obligations remain relevant. Different software tools may fall under different regulatory frameworks. A system used to sharpen an image, one used to screen for disease autonomously, and a general language model proposed as a diagnostic adviser should not be treated as though they carry identical evidence or duties. Labels matter because they define expectations.

The most difficult issue may be the interaction between a clinician and a recommendation that appears highly authoritative. A randomized clinical-reasoning study found that access to a large language model did not significantly improve physicians’ performance compared with conventional resources, even though the model performed well independently. Another study found that systematically biased AI recommendations could reduce clinician accuracy and that explanations did not reliably mitigate the harm. Taken together, these results should warn healthcare administrators against assuming the phrase ‘human in the loop’ is a complete safeguard. A person who merely approves an algorithm’s answer without a meaningful opportunity to question it is not performing the independent review the public imagines.

In legal work, we encounter a similar distinction between signing a document and exercising judgment about its contents. A person may technically approve a filing that was prepared by someone else. That does not establish the person understood the evidence or had a realistic chance to detect an error. The same principle applies to clinical AI. Hospitals should ensure practitioners have enough information, time, and professional freedom to investigate surprising conclusions. If a physician faces pressure to follow algorithmic recommendations because administrators have made them part of performance scoring, the claim that human judgment remains supreme may be fiction. Meaningful oversight cannot be reduced to an electronic checkbox.

An adverse event will also raise questions about the version of software in use. Did the hospital implement an update? Did the model change after new data were introduced? Was the output retained, together with the input and the clinician’s response? Who has authority to investigate an error reported by staff? The FDA’s 2025 final guidance on predetermined change-control plans recognizes that certain AI-enabled devices may undergo planned modifications under specified processes. But compliance with an external framework does not eliminate the need for records and local monitoring. A hospital should be able to reproduce, as far as reasonably feasible, the circumstances that produced a consequential recommendation. Otherwise, meaningful investigation may become impossible.

Privacy law presents another area where sweeping claims conceal important distinctions. A patient’s medical history, imaging, and genomic information may be legally protected, but protections depend on who receives the information and under what arrangement. A clinician entering a patient narrative into an unauthorized outside chatbot may create risks very different from using a properly contracted and controlled clinical system. Patients should not have to learn the architecture of a hospital’s software stack to know whether sensitive information may be shared. Institutions should establish clear limits, contractual safeguards, access controls, and breach procedures, and they should explain relevant uses in understandable language.

Genetic data deserve additional concern because they can reveal information about relatives who are not directly involved in the care episode. A system analyzing hereditary conditions may identify risks beyond the diagnosis being investigated. Who receives those findings? How are uncertain associations described? When is further testing warranted, and what consent was obtained for secondary findings? The mere ability to generate genetic hypotheses does not answer those questions. Research tools such as DeepRare may help clinicians identify previously missed possibilities, but a ranked disease candidate is not automatically a confirmed diagnosis. Patients should know the difference between a lead, a clinically significant laboratory result, and a finding sufficient to justify a major treatment decision.

Consider the emotional pressure on families of children with unresolved illnesses. Years of unsuccessful visits can produce understandable desperation for an answer. If a program presents a convincing hypothesis, parents may reasonably give it more weight than they would a casual suggestion. Physicians and genetic counselors have a duty to explain uncertainty without dismissing the technology’s potential. A confirmation process is not bureaucratic obstruction when it protects a child from treatment based on a speculative association. Equally, refusing to consider a plausible new lead simply because it emerged from a computer could be difficult to justify if further investigation is clinically appropriate. The law cannot substitute for good judgment, but it can insist that decisions be reasoned and recorded.

There is an uncomfortable possibility that some mistakes will be harder to recognize because AI systems make them consistently. Human clinicians may disagree, prompting further review. A widely installed program could make a systematic error across numerous facilities before anyone notices a pattern. That makes incident aggregation and independent oversight essential. Hospitals and vendors should have channels for reporting suspected model-related harm without suppressing criticism. Regulators and professional organizations should be able to examine repeated failures. We should not have to discover a common defect only after separate patients have suffered the same avoidable outcome in different states.

Sepsis alerts are a practical illustration of the difficulty. A 2026 multicenter validation of an updated model showed performance varied by hospital. At one selected sensitivity threshold, positive predictive values differed substantially among sites. What standard of care follows an alert? It would be unreasonable to assume every warning requires the same intervention regardless of clinical context. It would also be irresponsible to ignore serious warnings merely because the software sometimes produces false positives. A hospital needs a medically sound response protocol, training, documentation, and review of outcomes. The legal question after a problem arises will concern what people reasonably should have done with the information available, not whether the software carried a fashionable label.

There is also a question of disclosure. Patients may want to know when an autonomous system rather than a physician interpreted an image, or when a major treatment recommendation was influenced by a model. I do not think every routine background computational function requires an elaborate consent ceremony. Such an approach would produce forms without understanding. But material uses, important limits, and any reduction in direct professional review deserve transparent explanation. The patient should know how to obtain human review when clinically appropriate. Organizations should avoid lengthy consent language that technically mentions AI while leaving the real nature of the decision obscure.

Drug discovery has its own framework of accountability. The AI-discovered compound rentosertib produced encouraging findings in a small randomized phase 2a trial reported in 2025. The fact that software helped design or identify the molecule does not change the need for rigorous clinical trials, safety monitoring, manufacturing controls, and truthful marketing. A pharmaceutical company should not be held to a lower standard because its scientists used a different method of finding candidates. Nor should AI-derived compounds face a harsher standard simply because they are new. The proper inquiry is the safety and effectiveness of the actual product and the reliability of the evidence supporting it.

I also worry about the practical ability of injured patients to understand what went wrong. Proprietary software may be protected by trade-secret claims, while clinical records reside in another institution and model-development documents in a third. Courts have procedures for managing confidential information, but the cost of obtaining relevant evidence can be substantial. We need workable expectations about audit trails, relevant records, and cooperation in investigations. The objective should not be to force disclosure of every piece of valuable source code to the public. It should be to prevent secrecy from making it impossible to determine whether a safety-critical product contributed to foreseeable harm.

At the same time, public debate should not treat every medical error involving AI as proof that computers are uniquely dangerous. Medicine already suffers from missed diagnoses, delayed referrals, and failures of communication. The appropriate comparison is with the standard care patients would otherwise receive. If AI reduces errors in a defined use, refusing it may have consequences too. Lawyers are often accused of caring only about liability, but legal accountability can support adoption by clarifying responsibilities and encouraging careful implementation. A patient who trusts the process is more likely to accept a useful technological improvement than one who fears an algorithm will make an unreviewable decision.

Hospitals should adopt a written governance structure for meaningful clinical AI, with responsible individuals identified for purchasing, clinical validation, staff training, privacy, security, updates, incident review, and withdrawal of tools that fail. Those responsibilities should be reflected in vendor contracts. Clinicians should have clear channels to disagree with outputs and report concerns. Quality committees should receive information about subgroup performance and actual adverse events, not merely vendor accuracy claims. An institution willing to establish these controls demonstrates confidence in the technology rather than fear of it. Anyone can announce enthusiasm; accountable organizations explain what happens if the prediction is wrong.

The question ultimately concerns trust. Artificial intelligence can improve medical care in selected circumstances, and responsible physicians should be willing to learn from it. The evidence in organized breast screening is strong enough to take seriously. Genomic interpretation and pharmaceutical research offer plausible advances, although many hoped-for outcomes remain unproven. The hazards of biased suggestions and misplaced confidence are also real. My conclusion is that meaningful clinical judgment, documented evidence, and enforceable responsibility must move forward together. If an organization cannot tell a patient what the software was supposed to do, who reviewed its conclusion, and how an error would be investigated, it has not finished developing a safe system. And if the system fails, the answer to whose name goes on the complaint cannot simply be ‘the computer.’

Clermont, Florida Correspondent

Small professional-service businesses learn an uncomfortable truth about technology: a product designed for a large organization may solve a real problem and still be the wrong purchase for a smaller one. The software assumes that someone else handles training, that a second employee checks the reports, that a dedicated technician maintains the system, and that another department follows up when something unusual appears. In a small office, those people may all be the same person. I work with professional-services firms around Clermont, and I have watched a great many ambitious technology projects produce more complexity than the business was equipped to manage. That experience shapes my view of artificial intelligence in medicine. I see potential benefits that deserve investment, but the people marketing the tools must stop treating clinical accuracy and operational suitability as though they were synonyms.

Physicians and nurses have skills that business consultants do not, and I do not pretend otherwise. What is familiar to me is the daily pressure of a practice: appointments to keep, employees to retain, bills to collect, insurance requirements to satisfy, and clients or patients who expect an explanation when something changes. A small medical office may benefit greatly from technology that helps organize information or identify an abnormality. It may also suffer if every new feature creates a separate subscription, another login, an additional training obligation, and uncertain responsibility for errors. When somebody says AI will transform primary care, my first question is whether the physician can still look at the patient instead of at yet another screen.

The case for clinical AI is strongest when the task is narrow and the evidence is concrete. Organized breast-screening programs have now produced significant studies. A large German observational project reported higher rates of breast-cancer detection with AI-supported mammography, while the randomized Swedish MASAI trial found greater sensitivity with the AI-supported workflow, the same specificity, and interval-cancer outcomes that met the trial’s noninferiority standard. Those are serious findings. A practice or imaging centre might reasonably conclude that the evidence warrants considering an appropriate system. But a decision to adopt must also examine integration costs, staffing, local patient populations, and the availability of appropriate follow-up. The price of the license is only one part of the decision.

One of the most common mistakes in small business is buying a tool because it performs impressively in a demonstration. Demonstrations are organized around the functions the vendor wants to show, using examples selected to make the product look good. That is not sinister; it is how products are sold. A practice should insist on seeing the tool operate under ordinary conditions with the data, equipment, and staff it actually has. Can a medical assistant use it safely? What happens when the input is incomplete? Can a clinician review the result quickly? How often will the system return an inconclusive answer? How are errors reported? What is the backup process during an outage? A useful product should survive those questions.

Consider autonomous diabetic-retinopathy screening. For many patients with diabetes, obtaining regular eye assessments can be difficult, and a point-of-care tool may reduce an important barrier. A 2026 study reported definitive results in nearly 96 percent of patients examined in an endocrinology clinic with a non-ophthalmic operator. That is a promising achievement. But only a portion of those screening positive completed a documented confirmatory visit within the institution, and outside visits were not tracked. This is the kind of result a practice owner immediately recognizes. The screening process succeeded technically, while the broader service required additional coordination. Someone must handle the referrals, scheduling, records, and patient reminders. Those tasks cost money and staff time even when a computer performs the initial analysis.

The value proposition may therefore differ between a hospital, a large multispecialty group, and an independent clinic. A large organization may already have referral coordinators and technical support. A small practice may need to create them or contract for assistance. Conversely, a small organization may be nimble enough to train everyone involved and track follow-up personally. Neither structure is automatically better. What matters is whether the full process works. A vendor should be willing to discuss the staffing and service requirements honestly instead of presenting the algorithm as a self-contained replacement for clinical judgment and patient care.

The broader language models generate another form of sales temptation. They are exceptionally good at producing coherent reports and plausible explanations. A physician might use such a tool to help review relevant literature, prepare a differential diagnosis, or summarize complex records. Yet a randomized clinical-reasoning study found no significant diagnostic improvement when physicians were given access to an advanced model compared with conventional resources, despite the model’s impressive performance when used alone. That tells me that an excellent product can fail to deliver its potential when the workflow is wrong. Before using such systems routinely, practices should test how clinicians actually interact with them and determine whether they save time, change decisions appropriately, and improve patient outcomes.

There is an additional danger in how the tool persuades its user. A separate randomized study found that systematically biased AI recommendations reduced clinicians’ diagnostic accuracy, and explanatory language did not adequately counteract the problem. In business consulting, a flawed forecast presented in an attractive spreadsheet is sometimes more dangerous than a rough estimate because it gives clients an unwarranted feeling of precision. The same can happen with medical advice. A detailed paragraph may look like proof when it is only a convincing interpretation of incomplete data. Professionals need to know when to reject a suggestion, seek more information, and explain to patients why a machine’s answer has not been accepted.

The training requirement is not a one-time employee orientation. Staff need to understand the system’s intended use and limitations, the conditions under which results may be unreliable, and their responsibility for escalation. New employees must receive the same instruction. Software updates may require refresher training. A small practice cannot reasonably delegate these matters to a non-existent department, so it should select products whose oversight requirements match its resources. That is not a lower standard of safety. It is a recognition that the same standard may require different operating arrangements depending on the setting.

Genetic interpretation presents a different opportunity. Families dealing with rare unexplained diseases often face a complicated sequence of consultations and inconclusive reports. AI systems can compare genomic variants, clinical findings, and newly published medical literature at a scale few people can manage manually. Studies of platforms such as DeepRare and AI-assisted review of previously unresolved pediatric cases have shown promising diagnostic leads. A family physician might benefit from being alerted to a plausible condition that warrants specialist attention. The benefit could be substantial even if the tool does not make a final diagnosis. But the practice must communicate uncertainty, arrange confirmation, and avoid representing a ranked hypothesis as definitive evidence of a disorder.

There is also a business case for helping clinicians keep up with emerging knowledge. Specialists cannot read every journal article, and generalists are asked to manage an astonishing range of conditions. A carefully controlled literature-retrieval tool could locate clinical guidance or summarize recent studies while linking back to original sources. This may be less glamorous than declaring that AI has diagnosed a disease on its own, but it could be useful in everyday practice. The crucial safeguard is that the tool must not invent citations or conceal uncertainty. The physician needs to know what evidence exists and how closely it applies to the patient, not simply receive an answer that sounds academically polished.

Drug discovery is more remote from the operations of a local practice, but patients have a stake in it. A promising example is rentosertib, an AI-discovered compound investigated for idiopathic pulmonary fibrosis in a 2025 randomized phase 2a trial. The seventy-one-person study showed encouraging preliminary results but did not establish long-term clinical benefit. If AI can help researchers identify promising candidates more efficiently, the industry may improve its ability to search for therapies. Yet that progress will not matter to patients unless subsequent trials confirm safety and efficacy, treatments become available, and costs are manageable. Faster molecular design is not the same thing as faster access to an effective prescription.

Pricing should be examined skeptically. Software vendors sometimes describe their service as an alternative to expensive specialist time, but the cost may include per-use fees, equipment, integration, updates, data hosting, support, and contract escalation. A practice needs to estimate actual volume and total annual spending. If a tool generates more referrals, the practice should calculate the time needed to discuss and coordinate them. If it reduces routine interpretation or documentation time, the practice should measure the hours saved rather than accept optimistic vendor estimates. Most independent businesses cannot survive by celebrating theoretical efficiency while continuing to pay the same bills.

A sensible contract should address what happens when the vendor’s product changes. Who provides support? How quickly can a problem be reported? Will prices rise? Can the practice retrieve its records in a usable format if it changes vendors? Does the system require an internet connection during appointments? What happens if an update modifies clinical behavior or temporarily disables a feature? Who is responsible for reviewing changes and deciding whether the practice can continue to rely on the tool? These details are not glamorous, but they distinguish a useful long-term service from a short-lived technology fashion.

Patient trust may prove just as important as technical quality. Some patients will welcome AI assistance if they believe it adds another careful look. Others will worry that their physician is outsourcing judgment to a computer. A practice should explain, in ordinary language, what role the system plays. If the software helps interpret an image, say so. If the clinician reviews the findings independently, explain that process. If an automated screening result requires further testing, make clear that the initial result is not a final diagnosis. Patients should not be misled into thinking a machine possesses human understanding, but neither should they be denied a useful tool because its operation is unfamiliar.

Privacy is part of that relationship. Medical histories and genetic data are far more sensitive than an ordinary customer’s scheduling preferences. Practices must know where information goes, whether vendors retain it, and who can use it for other purposes. A convenient general-purpose chatbot should not become an informal repository of identifiable patient records without appropriate safeguards. Small practices often have fewer technical staff than hospitals, making clear vendor agreements and employee policies especially important. Information security should be part of the purchase decision, not a repair project after something has gone wrong.

I also want physicians to retain professional discretion. A vendor may promote an AI recommendation as the standard pathway, and an insurer or manager may be tempted to enforce that recommendation to control costs. But real patients do not always fit the categories on which software was trained. A clinician should be able to document disagreement without being punished for refusing a model’s suggestion. In turn, clinicians should be willing to consider evidence that contradicts their first impression. The best relationship is one in which the tool helps reveal possible error, not one in which the machine or the human wins a contest of authority.

Government can encourage this form of responsible adoption by insisting on truthful claims and usable evidence, supporting appropriate validation, and ensuring that patient privacy is protected. It should resist treating the number of installed systems as a public-health achievement. Grants and reimbursement policies should be tied to services actually delivered and useful outcomes wherever practical. I would prefer a modest program that demonstrably improves screening completion and follow-up to an expensive publicity campaign about digital transformation. Patients care about the service, not the label chosen by policymakers.

My clients often ask what kind of technology investment makes sense for the next several years. I tell them to begin with the customer problem and work backward, rather than begin with the product and invent a problem to justify it. Medicine should do the same. If a practice has delayed imaging reports, perhaps AI-supported interpretation is worth assessing. If too many diabetic patients miss eye screening, autonomous screening coupled with referral coordination may help. If clinicians struggle to interpret unusual genetic findings, an evidence-linked decision aid may offer useful leads. Each case deserves its own evidence, budget, and operating plan. A grand strategy announced from a conference stage does not relieve the practice of those decisions.

I believe AI will improve medical care in selected applications and eventually become ordinary infrastructure, much as electronic records and digital imaging have. But the transition will be uneven, and systems built for impressive demonstrations will not always survive contact with daily practice. The imaging evidence is promising; genetics and drug discovery are advancing; general diagnostic models need serious testing and training to avoid over-reliance. What I want from the technology is unremarkable in the best possible sense: clinicians who make sounder decisions, patients who receive timely answers, and practices that can deliver care without creating new confusion or cost. A product that accomplishes those things will not need to claim it revolutionized medicine. Its patients will be able to tell the difference.

Novi, Michigan Correspondent

In automotive supply-chain work, there is a reason manufacturers do not qualify a new component simply because it passed a test in the supplier’s laboratory. We want to know whether the process can reproduce that result, whether the materials change, whether the inspection system catches the defects that matter, and whether anyone can trace a failure to its source. A component that performs perfectly on a test bench but fails after thousands of miles can be a costly embarrassment or a deadly hazard. That is the analogy I use when I hear people argue that artificial intelligence has already proved itself in medicine. Some systems may be excellent, and several clinical studies now provide real evidence of benefit. But the question is not whether a model produced a good score under one set of conditions. It is whether the entire process delivers a dependable result over time, including when something changes unexpectedly.

I am not a medical specialist. I manage relationships and operations across automotive supply chains, where suppliers, production plants, logistics providers, and customers must coordinate work despite having different systems and incentives. Healthcare has its own professional disciplines, but it shares the challenge of transforming information into reliable action. A suspicious image is not a treatment plan. A genetic hypothesis is not a confirmed disease. An automatically generated warning is not a clinical intervention. Each stage depends on someone who knows what to do next. In manufacturing, an inspection camera that spots defective parts is useful only if the line separates them, identifies the cause, and prevents additional failures. In medicine, the consequences of neglecting the next step can be immeasurably greater.

The best evidence for some imaging applications is worth acknowledging. Germany’s PRAIM study examined more than 463,000 breast-screening examinations and found higher cancer-detection rates in AI-supported readings. The randomized Swedish MASAI trial enrolled around 106,000 women and in its 2026 analysis reported higher sensitivity with AI support, unchanged specificity, and a rate of interval cancers that met the noninferiority criterion. The studies were conducted under different designs, and both contribute useful information. This is a stronger basis for adoption than the familiar claim that a model beat doctors on a narrow retrospective dataset. It tells us that a particular human-machine process can perform well in a real clinical environment.

Yet automotive engineers would immediately ask about process capability outside the test sites. Does the model work on scanners made by different manufacturers? What happens after equipment calibration changes? How does performance vary with image quality, patient age, and other clinically relevant factors? How are borderline findings reviewed? Does the tool reduce the kind of missed diagnosis that matters most, or merely increase the number of abnormalities flagged? A procurement team should not confuse a published average with guaranteed local performance. We qualify suppliers because variation is inevitable. Hospitals need an analogous discipline for algorithms that may be deployed across different locations and patient populations.

The distinction between detection and outcomes also resembles what happens on a production line. A new sensor may identify more flaws without reducing the number of defective products reaching customers, perhaps because the downstream correction process is weak. Increased detection can be valuable, but it is not a complete performance measure. In cancer screening, better sensitivity may lead to earlier treatment of dangerous disease, yet the true balance includes follow-up, unnecessary investigations, patient anxiety, and long-term outcomes. In a 2026 diabetic-retinopathy implementation study, an autonomous system generated definitive results for most participants, but only part of the group with positive screens obtained documented confirmatory care at the same institution; care elsewhere was not captured. The screening step had advanced farther than the complete service.

A factory tries to avoid finding defects only at the final inspection station. We prefer processes that prevent them and make correction efficient. Healthcare should similarly focus on where errors enter the diagnostic and treatment process. Some mistakes arise from image interpretation, and AI may help. Others come from missing records, delayed communication, incorrect medication reconciliation, or poor follow-up. An expensive algorithm that addresses the wrong source of error may deliver little improvement. I would like hospital managers to map the entire pathway and identify the largest avoidable failures before selecting technology. A software feature is not valuable because it is clever; it is valuable because it reduces a real failure mode.

The diagnostic-language-model research suggests that the interface between people and software deserves as much attention as the model itself. A randomized study reported that physicians given access to a capable large language model did not significantly improve their diagnostic reasoning compared with colleagues using conventional resources, although the model alone scored strongly. In engineering terms, a high-performing component was inserted into a process without producing a clear improvement in the finished product. That is a signal to study the interface, training, workflow, and division of responsibility. It is not an excuse to declare the component useless, nor proof that any future integration will work automatically.

A second randomized experiment showed an even sharper hazard: systematically biased AI advice lowered clinicians’ diagnostic accuracy, and added explanations did not adequately reverse the effect. That sounds painfully familiar from failures caused by operators trusting a sensor that was miscalibrated. The display may look precise while its underlying assumptions are wrong. If a machine repeatedly gives the same wrong instruction, people may stop questioning it, particularly under pressure. This is why manufacturing plants use independent quality checks, limits on automatic control, and procedures for abnormal conditions. Healthcare teams need equivalent safeguards, designed by clinicians who understand which errors can cause harm.

The public frequently hears that AI systems improve as they acquire more data. Some do, but improvement is not a magic property and a software update is not a promise of better local performance. A model can change because its developers adjust training, thresholds, or inputs. Hospital equipment and patient populations can change independently. The FDA’s 2025 guidance on predetermined change-control plans for AI-enabled medical devices reflects the need to manage certain planned modifications. I would want a controlled process that records what changed, why it changed, how validation was performed, and whether an update created unexpected effects. In automotive manufacturing, a supplier cannot quietly substitute a different material simply because its engineer believes the replacement is superior.

Traceability matters as much as initial accuracy. When a defect appears in a car, manufacturers need to identify affected lots, materials, suppliers, and manufacturing periods. In medical AI, a serious error may require knowing the software version, device configuration, relevant input information, model output, clinician action, and subsequent outcome. Hospitals must preserve the records necessary for lawful quality review while protecting patient confidentiality. Without that traceability, a repeated failure may remain invisible because each affected patient appears to be an isolated event. The goal is not to assign blame reflexively. It is to identify a faulty process while there is time to prevent further harm.

Think about sepsis alerts. The updated Epic model was prospectively evaluated in a 2026 multicenter study, and its performance differed materially among four American health systems. The variation should not shock anyone accustomed to complex operations: different settings create different inputs and baselines. But it should prevent executives from assuming one common threshold will produce equivalent clinical value everywhere. At a threshold selected for 60 percent sensitivity, the proportion of positive alerts corresponding to sepsis varied notably among sites. A hospital must weigh the benefit of earlier warning against unnecessary evaluation and clinician attention. The algorithm should be tuned and governed according to clinical evidence, not administrative convenience.

Reliability includes the people using the tool. An operator who receives too many false alarms may become less attentive. A physician who is pressed to process cases quickly may accept suggestions without verification. A technician may enter information incorrectly because the interface is confusing. In manufacturing, we examine such events as interactions between design and human behavior rather than simply calling the worker careless. Medical organizations should do the same. Training should include realistic examples of wrong AI recommendations, incomplete data, and uncertain findings. The system should make it easy to question results and report anomalies. A culture that penalizes challenges will eventually lose the warnings it most needs.

Rare-disease genetics reveals another important distinction. Modern AI tools can compare immense amounts of literature and genomic information to identify plausible explanations for difficult cases. Recent DeepRare research and other 2026 studies have reported encouraging performance and new leads in previously unresolved records. This resembles using advanced analytics to identify a production defect that conventional inspection missed. But the output is a hypothesis until it is independently confirmed. A patient should not undergo consequential treatment because a model ranked one genetic disorder above several alternatives. Clinical laboratories, specialists, and genetic counselors must verify relevant findings and explain the uncertainty.

The drug-discovery discussion requires similar discipline about stages. AI can help identify biological targets and design compounds. Rentosertib, a candidate associated with AI-enabled discovery, reached a randomized phase 2a trial for idiopathic pulmonary fibrosis, reported in 2025, with encouraging early findings. The trial involved seventy-one participants over twelve weeks. In industrial development, that resembles proving a new design in a limited pilot rather than certifying it for production at scale. The next stages may reveal difficulties in effectiveness, safety, manufacturing, or cost. The achievement is real, but responsible commentary must name its stage accurately. A good prototype is worthy of investment precisely because substantial validation remains.

There is also a supply-chain dimension to AI-enabled medicine that receives too little attention. Clinical systems rely on scanners, sensors, semiconductors, data centres, secure networks, software vendors, and technicians who maintain them. A vendor outage or cybersecurity incident can interrupt care across multiple facilities. Hardware may become obsolete faster than hospitals can replace it. Critical devices may depend on manufacturers operating abroad. Healthcare organizations should understand dependencies and establish contingency plans just as manufacturers do for essential components. A diagnostic pathway cannot be considered robust if one unavailable cloud service causes it to stop functioning without a safe alternative.

Supplier concentration increases the risk. Hospitals may integrate one vendor deeply into electronic health records, imaging, scheduling, and clinical decision support. Switching becomes difficult once staff training, interfaces, and data workflows depend on the product. A supplier can then raise prices, alter terms, or change functionality while the hospital has limited bargaining power. Procurement should consider interoperability, data portability, service-level commitments, cybersecurity duties, and the ability to operate safely if the relationship ends. Those are ordinary questions in industrial supply management. They are even more important when patients depend on the service.

I would insist on measurement that survives the marketing presentation. Choose a defined clinical workflow and collect baseline performance. Introduce the tool in a controlled way, with review by the appropriate professionals. Measure missed findings, false alarms, clinician time, completed follow-up, serious adverse events, and treatment outcomes where feasible. Review performance by relevant patient subgroups, scanners, and facilities. Examine what happens after software updates and staffing changes. Compare results with the original objectives and be willing to withdraw a system that underperforms. A company that keeps shipping defective components because it has already paid for the equipment would not last. A hospital should not cling to a weak algorithm because the license is expensive.

None of this means that technology should be adopted slowly merely to satisfy committees. In manufacturing, a capable new process can create enormous advantages, and organizations that refuse to improve will fall behind. If an AI-supported screening workflow demonstrably identifies important disease more reliably, I want healthcare systems to examine it promptly. If a clinical tool can help a doctor consider a rare diagnosis, it deserves thoughtful use. The issue is controlled adoption, not paralysis. The faster a technology is introduced, the more valuable good validation and continuous monitoring become. Speed without process discipline is not the same thing as progress.

We should also remember that the point of better productivity is to serve people. A factory can produce more goods, but consumers judge their usefulness, quality, and reliability. A hospital can process more images, but patients judge whether the right diagnoses are made, the right actions follow, and avoidable harm is reduced. Using AI to give clinicians more time for difficult decisions would be a worthy achievement. Using it to demand ever-faster patient turnover without monitoring outcomes would be a questionable one. The measure must be quality of care as experienced over the full course of treatment, not simply the number of cases completed per hour.

My conclusion is that AI deserves neither blind acceptance nor an automatic veto. The studies in breast imaging show that carefully designed human-machine workflows can perform very well. Emerging genomics tools can identify leads worth clinical investigation, and AI-assisted research can move drug candidates into meaningful trials. The evidence also shows that biased recommendations, poor integration, and unreliable alerts can undermine human judgment. Anyone familiar with complex industrial systems should recognize the lesson. Quality is not a property of a component advertised in isolation; it emerges from a controlled process that performs reliably under real conditions. Before we celebrate the intelligence of the software, we ought to inspect the production line that turns its output into care.