Dayton, Ohio

When somebody in my line of work says a process has become faster, my first question is not how impressive the demonstration looked. It is what moved through the system, how much came out right, and whether we can reproduce the result on Tuesday with a different crew. The distinction may sound ordinary, but it is precisely the distinction being lost in the excitement over artificial intelligence and scientific discovery. I am an engineer who has spent enough time around shop floors to know the gap between a prototype and a process. I can believe that AI is delivering extraordinary scientific capabilities and still be unwilling to call every clever model output a breakthrough. The evidence available this May gives reasons for enthusiasm, particularly in mathematical reasoning and biology. It also gives reasons to insist that science retain its oldest quality-control department: reality.

Consider the mathematical work. In 2024, AlphaProof and AlphaGeometry together reached the silver-medal standard at the International Mathematical Olympiad. By July 2025, an advanced Gemini Deep Think system solved five of six Olympiad problems, earning a gold-medal-standard score certified by contest officials. That is not a trivial parlor trick. Those problems require structured reasoning, and the movement from specialist formal translation toward answers in ordinary mathematical language represents a real technical advance. Yet an Olympiad problem is designed to have a solution that someone knows how to check. An open research problem can require inventing a useful definition, deciding what is worth proving, and determining whether the result changes anyone’s understanding. The difference resembles an engineering test with a defined acceptance standard versus a factory trying to decide what product it ought to build. Success at the former is meaningful without settling the latter.

A good scientific tool should reduce the time between a promising idea and a reliable answer. Engineers call that cycle time, although scientists understandably prefer more dignified terminology. There are several distinct cycles inside a research project. Somebody must find relevant earlier work, formulate an explanation, design the test, obtain materials, run the experiment, analyze results, challenge assumptions, and communicate the finding. If AI reduces literature-search time by an afternoon but introduces a questionable citation, the time saving can evaporate. If it suggests ten candidate protein structures in minutes but the laboratory can test only one per week, the bottleneck has merely moved. A faster first station on an assembly line does not necessarily mean more finished products leaving the shipping dock.

That is why I find the structural-biology progress unusually credible. Protein structure prediction provides an object that can be compared with experimental evidence. Tools in the AlphaFold family have made predicted structures broadly available, allowing scientists to pursue questions that otherwise might never have reached a practical starting point. Newer approaches can predict interactions among proteins and other molecules, and generative systems can propose novel structures rather than merely catalog existing ones. Those are tangible improvements in a researcher’s working inventory. But the prediction is not the molecule, and the molecule is not its function. A part can look perfect on the engineering drawing while failing because of heat, vibration, impurities, stress, or its interaction with a neighboring component. Biology has its own versions of all of those conditions.

The technical literature itself recognizes this limitation. Researchers comparing AlphaFold models with crystallographic evidence have cautioned that prediction confidence varies and that environmental factors can be missing. Nature’s 2024 discussion of AlphaFold 3 also raised reproducibility questions when complete underlying code was not initially made available. I do not consider either issue an indictment. I consider them familiar signs that a promising technology is entering serious engineering territory. The more consequential a calculation becomes, the more we need traceable inputs, documented assumptions, and independent tests. A salesman can be allowed to describe a machine as revolutionary. The plant manager needs a maintenance manual and a warranty.

The May 19 publications on AI research assistants make the debate more interesting. Google’s Co-Scientist and FutureHouse’s Robin are not just chat interfaces producing attractive summaries. Their developers describe systems that help construct hypotheses, propose experimental work, and analyze actual laboratory results. The Co-Scientist paper reports biomedical validation in areas including drug repurposing and mechanisms of antimicrobial resistance. Robin’s reported workflow connects literature assessment and experimental planning with analysis of results from human-executed laboratory experiments. These are genuine steps toward joining parts of the discovery process. We should take the reports seriously. We should also read the words “in vitro” with respect. A result in cultured cells is a result in cultured cells. It is not proof that a patient will recover or that a treatment has become safe.

If I were asked to evaluate such a system for a manufacturing operation, I would insist upon three separate scorecards. The first would measure task performance: did it find the right paper, generate a valid mathematical argument, or suggest a workable protein? The second would measure workflow performance: did the team complete its next meaningful milestone with fewer hours or less wasted material? The third would measure outcome performance: did the resulting product or discovery do something that matters outside the controlled demonstration? The scores might rise at different rates. That is perfectly acceptable. But when companies move from one scorecard to another in their advertising, customers need to notice. A tool that gets a hypothesis drafted thirty times faster has not thereby demonstrated thirty times as much knowledge.

I would also ask whether the results survive a change in operator. Industrial processes have a habit of looking miraculous while a talented developer is standing beside them and ordinary when that person goes home. An AI scientific assistant may be handled by expert researchers who know which prompt to give, which output to distrust, and when to abandon a false lead. Such expertise is valuable, but it creates a measurement problem. We need to know how a less experienced but reasonably trained group performs, not merely how the inventors perform on favorable examples. An independent replication ought to specify who selected the problem, which tools were permitted, what the system saw, how many unsuccessful attempts occurred, and what counted as success before the trial began.

There is a related problem in science that manufacturing learned painfully: inspection after production cannot guarantee quality if the process itself is unstable. Peer review matters, but a paper’s acceptance is not identical to the validity of its conclusions. The AI Scientist system described in Nature this March produced computer-science experiments and manuscripts, with one generated paper passing the initial review threshold at a workshop. That is an interesting demonstration of automation. Yet publication workflows can be gamed, overloaded, or fooled even when every participant is human. Faster manuscript production without stronger verification could increase defective output as well as useful output. Anyone who has sorted a pallet full of parts knows that higher volume can make a bad defect harder to catch.

I do not think the solution is to deny AI access to research. The possibility of automated mathematical checking could prove as important as automated mathematical generation. If a model suggests an argument, another system may translate essential steps into formal proof obligations. If a model proposes an engineered protein, experiments can check binding, stability, specificity, and behavior under realistic conditions. If it surveys thousands of scientific papers, a retrieval system can attach passages and source identifiers so that the reader can inspect the original evidence. Those safeguards should be designed in from the beginning rather than bolted on when an embarrassing error reaches print. In engineering, verifying a part is not an insult to the machinist. It is respect for the intended use.

An underappreciated opportunity lies in the preparation of experiments, not merely the invention of spectacular hypotheses. A laboratory can lose weeks because a protocol lacks a necessary control, a measurement is improperly calibrated, or a sample label has been handled carelessly. AI systems could help generate checklists, detect inconsistent methods, compare planned tests with established practices, and flag suspicious values before a costly experiment proceeds. Much of that work will never appear in a sensational headline. It may nevertheless deliver more scientific output per dollar than a system that occasionally proposes a dazzling but impractical idea. The tools that transformed factories were often not the dramatic robots visitors photographed. They were improvements in scheduling, gauging, maintenance, and feedback.

There is a workforce question too. If young researchers learn to request a solution before they have learned to interrogate a problem, we could create very fast operators who do not understand their instruments. I have seen versions of that in manufacturing software. A technician knows which button produces the answer but cannot explain why the reading makes sense. The proper response is not nostalgia for slide rules. It is better training. Students should be asked to predict what a model will do, check where it fails, document alternative explanations, and defend conclusions without leaning on a machine’s apparent confidence. An AI result that a scientist cannot understand, validate, or challenge is not a substitute for scientific judgment.

The more mature systems become, the more uncomfortable questions about intellectual property and access will matter. Some research groups can afford compute, proprietary databases, and laboratory automation that others cannot. A public university in Ohio should not be forced to treat a private model’s output as an oracle because the underlying system is unavailable for independent testing. Conversely, businesses that spend heavily to invent useful systems cannot be expected to reveal every trade secret. We will need practical arrangements for test access, publication of methods, preservation of data, and independent auditing. I am less concerned with choosing a fashionable ideology for AI than with knowing whether the equipment can be trusted and whether competition is possible.

There is a temptation to define progress through the most advanced demonstrations. I would define it through the median research operation. Can a small laboratory achieve a credible result it otherwise could not attempt? Can a group studying obscure industrial materials afford to search the literature and model candidates without hiring an entire computational department? Can a public health team identify a plausible mechanism and test it before the problem becomes larger? These are less glamorous questions than whether a model can defeat an elite mathematics examination. They are also closer to the economic meaning of technological progress. A breakthrough confined to an extraordinary laboratory is promising. A tool that improves ordinary practice across thousands of laboratories is transformative.

We should be candid that much of the evidence available today is still early. Reports of experimentally validated AI-generated hypotheses are important, but the samples of problems studied remain limited. The research on AI-augmented scientific publishing published in January raises an awkward counterpoint: individual scientists may produce more influential work while the collective range of research questions becomes narrower. In factory terms, a process can become excellent at producing the items already on the schedule and less capable of discovering what customers will want next. Algorithms trained on the established literature may repeatedly direct talent toward well-described problems, leaving awkward outliers neglected. That would be a loss science could not easily measure from publication counts alone.

For the next several years, I would want a reporting standard that forces a disciplined answer to five questions. What was the baseline without AI? What precise stage was accelerated? How much human selection and correction was required? Did independent replication confirm the result? Did the result alter downstream scientific or practical work? Those are not impossible demands. They are what people routinely ask when a contractor proposes a new manufacturing line. The answers should include unsuccessful trials. A reliable denominator matters as much as a memorable numerator. Thirty useful proposals out of three thousand attempts tell a different story from thirty out of thirty, even if the finished thirty look identical in a press release.

My answer, therefore, is that researchers are seeing a genuine acceleration in selected activities and that some of the biological demonstrations justify the word discovery in a carefully defined sense. The larger claim—that the entire scientific enterprise is advancing at a similarly accelerated pace—has not yet been proven. I expect that stronger integration of models, formal verification, robotics, and reproducible experiments will gradually make that claim more defensible. But any institution that wants to spend public funds or its own shareholders’ money on this promise ought to insist on evidence measured at the point where the work actually creates value.

The scientific method is not endangered because a machine can suggest a hypothesis or outline a proof. It is endangered if the people using the machine forget that an appealing output is not an observed fact. Science should welcome faster tools, remain suspicious of shortcuts, and publish what happened when the proposed idea met an uncooperative world. That is how we teach a technology to serve discovery instead of advertising. In Dayton, we would call it getting the part off the drawing and into service. I see no reason the standard should be lower for something that claims to change our understanding of life itself.

Phoenix, Arizona

The first time somebody explains an artificial-intelligence advance to me as a revolution in medicine, I usually ask what will change in our hospital next Monday morning. Will a patient be diagnosed sooner? Will somebody be able to schedule an appointment that was previously unavailable? Will an experimental treatment eventually become something a physician can prescribe with reasonable confidence? I am not asking because I dislike new technology. I work with complicated technology every day, and there are days when I would welcome a machine that understood our scheduling system better than the people who bought it. I ask because the distance between a laboratory result and a patient outcome is a serious distance. We should celebrate promising research while refusing to confuse one end of that journey with the other.

As of this May, there is substantial reason to believe artificial intelligence is changing what scientists can attempt. Protein prediction tools have changed the preparation for biological research. Systems capable of generating candidate molecules, proposing explanations for microbial behavior, or connecting observations across huge collections of papers offer researchers a range of ideas no individual team could easily assemble by itself. The recently published Co-Scientist and Robin research goes further by joining hypothesis generation to experimental work. If software helps investigators identify an unexplored drug combination and cultured cells respond as predicted, something real has happened. That result deserves attention. It also deserves its proper label: a laboratory finding that may or may not become clinically useful.

In our hospital, the most dangerous phrase is often “the system says.” The phrase can conceal a chain of people, software settings, incomplete data, and assumptions nobody remembers making. Research AI could create a similar habit if laboratories treat an attractive hypothesis as a fact merely because the system cites papers and speaks in a practiced voice. We already know how easily an ordinary literature review can be distorted by selective reporting, poorly designed studies, and publication bias. Putting more computational power behind the selection does not make those problems vanish. It can magnify them. A summary that excludes a crucial negative result can send an entire project down the wrong corridor, just as a clinical chart missing an allergy can turn a routine decision into an emergency.

The mathematics results are fascinating, but I would keep them in their own category. Winning a gold-medal-standard score on International Mathematical Olympiad problems, as Gemini Deep Think did in 2025, is an impressive demonstration of reasoning. A mathematical argument has the considerable advantage of allowing a rigorous check within a defined framework. Medicine is less accommodating. A molecular mechanism can work in a dish, fail in an animal, appear safe in a small clinical trial, and disappoint in ordinary hospitals. Patients bring age, medication interactions, multiple diagnoses, financial barriers, and genetic variation that a narrow experiment cannot fully represent. Successful proof generation does not establish that a machine understands any of those messy realities.

We should also recognize an opportunity on the unglamorous side of research. Investigators spend enormous effort reconciling evidence from thousands of studies. Some of the work is clerical in the best sense of the word: checking inclusion criteria, identifying duplicated populations, comparing outcome definitions, and determining which findings are actually independent. An AI assistant that accurately performs these functions could save months and reduce mistakes. But “accurately” carries a heavy burden. A system that misses a study of adverse events or invents an unsupported citation can harm the quality of an entire review. The appropriate design is one in which every material assertion links to the original source and a human can reproduce the selection of evidence.

I am particularly concerned about the problems that never become glamorous enough to attract an algorithm developer’s attention. Our patients are not a neat collection of textbook diseases. One woman may have poorly controlled diabetes, mobility limitations, insufficient transportation, and no confidence that an insurer will authorize a new prescription. An elderly patient may speak limited English and receive care from several clinics using systems that barely communicate. A research program built around beautiful molecular datasets may be excellent at finding patterns while leaving the circumstances of actual treatment untouched. If AI helps scientists discover ten promising therapies while the price or delivery system keeps most patients from obtaining them, the public-health payoff will be much smaller than the scientific announcement suggests.

The paper from January showing that AI-assisted researchers may become more productive while collective research topics narrow deserves more attention than it has received. It points toward a risk that is easy to overlook in a hospital: technology makes the measurable part of work faster and draws resources away from the less measurable part. An algorithm trained on heavily published diseases may find more attractive directions in those same diseases. A rare condition with poor data may remain overlooked, and problems involving combinations of social circumstances and chronic illness may remain almost invisible. The result could be a larger volume of sophisticated knowledge concentrated where knowledge is already plentiful. More papers would not necessarily mean that medicine became more responsive to the patients who have been ignored.

Some people answer this concern by claiming that scientific AI will eventually become cheap enough for everyone. Perhaps, but cheap access to a screen is not the same as access to a research infrastructure. Developing and testing a treatment can require expensive instruments, patient recruitment, laboratory reagents, data protections, and skilled staff. A program that generates a thousand plausible ideas may overwhelm an underfunded research institution unless it can also prioritize them responsibly. The difference between finding candidates and paying to validate them is critical. If technology companies count every generated candidate as progress, they may shift the financial burden of sorting the useful from the useless to universities, hospitals, and taxpayers.

One standard I would like to see is a patient-centered version of the ordinary scientific scorecard. We should measure time to a validated hypothesis, independent confirmation, time to a plausible intervention, and eventually changes in meaningful outcomes. In medicine those outcomes might include fewer hospital admissions, improved survival, reduced disability, or a reliable reduction in severe side effects. We should also record failures and abandoned candidates. I understand that early research cannot promise an ultimate clinical endpoint; some discoveries will become important only decades later. Still, public communication should make clear where a project sits in that chain. A press release should not be permitted to slide from “identified in a cell culture” to “new treatment for patients” without acknowledging the missing steps.

The May publications about Robin and Co-Scientist offer an opportunity to do this well. Their descriptions include actual experimental work rather than simply evaluating whether the models sound persuasive. That is the right direction. Yet the fact that human scientists executed experiments, interpreted their limits, and chose which hypotheses deserved attention is not an embarrassing qualification. It is part of the success. We should not be so eager to call a system autonomous that we erase the trained people who made the result possible. Our surgical suites are full of extraordinarily capable equipment, and nobody confuses that equipment with a hospital’s obligation to provide competent care.

There is another issue that the public does not see in glossy announcements: the maintenance of knowledge. An AI system may make a finding using a specific version of a model, particular training material, and data that later changes. If investigators cannot recreate the computational conditions, what exactly has been replicated? Laboratory science already fights variation in reagents, samples, and equipment. Now we add model versions, proprietary interfaces, and changing retrieval systems. A convincing replication must say which model did what, which settings were used, which data were accessible, and which portions of the process were selected or rewritten by human experts. None of that undermines scientific freedom. It makes the claims testable.

People often ask me whether AI will replace scientists. That is not the question I would put before a hospital board. I would ask which scientific tasks could be improved, which new competencies our staff will need, and how we protect the ability to recognize a wrong answer. A researcher who lets software search widely and suggest connections may become more productive. A researcher who has never learned to distinguish correlation from mechanism may become more vulnerable to a persuasive machine. Medical training has to preserve the ability to make judgments under uncertainty. It also has to teach how automated systems fail, because even a powerful system may fail in ways that are unfamiliar to the people relying on it.

The incentives of universities and companies are worth considering. Successful research attracts funding, prestige, patents, and partnerships. Those rewards can encourage progress, but they also encourage claims that are difficult to compare. A company might announce an accelerated discovery pipeline without defining whether the improvement reflects fewer laboratory failures, less time reading papers, or simply more proposed experiments. A university may emphasize a high-profile AI publication while the department doing replication struggles to finance equipment. The better response is not to accuse every innovator of bad faith. It is to require reasonable disclosure and finance independent checks. Hospitals understand the cost of adopting technologies whose performance was measured under conditions unlike those in ordinary care.

There is a hopeful side to all this. Some of the most valuable advances may arise where AI allows researchers to connect previously separated fields. A scientist investigating inflammation might discover a useful mechanism in an unrelated branch of biology. A researcher working on a rare condition might identify similarities to a better-studied disease. A system could help design a more discriminating experiment, reducing wasted time and unnecessary animal testing. These possibilities matter. They are especially appealing because they might make research capability more widely available rather than simply making elite laboratories faster. We should design funding programs that test such claims rather than assume the benefits will distribute themselves.

I would also remind policymakers that faster knowledge can create new responsibilities. If researchers are able to identify possible interventions sooner, regulators will need to examine them competently. Reviewers may face a larger volume of submissions, some excellent and some almost entirely machine-produced. Trial designers will need stronger safeguards against poorly justified experiments. Research institutions will need to maintain data security and patient privacy even when external AI tools process information. There is no gain in shortening the discovery phase if later stages become a congested holding area. The improvement needs to be considered as a system, from the first hypothesis to a safe therapy delivered in an actual community.

My position is neither that AI discovery is empty theater nor that the revolution is already finished. There is real acceleration in literature work, structural modeling, problem solving, and selected cycles of biological experimentation. The evidence does not yet justify broad claims about the rate of improvements in patients’ lives. That gap is not proof of failure. It is a reminder of what science must accomplish before its promise becomes medicine. We should be ambitious about the tools, strict about the measurements, and unsentimental about the barriers that remain between an exciting idea and care that works.

I think about a patient leaving a hospital with a stack of instructions and a new treatment plan. She does not need to know which computational system suggested the original molecule. She needs to know the treatment was properly tested, the instructions are understandable, the pharmacy can supply it, and somebody will answer when something goes wrong. If scientific AI helps us arrive at that point sooner, it will deserve its reputation. Until then, the honest description is that researchers are learning to move faster through certain parts of discovery, while the hardest and most human portions of the journey still demand patience. In Phoenix, where heat and distance can turn a small delay into a large problem, I consider that distinction fundamental rather than academic.

The final measure, in other words, should be a combination of intellectual honesty and practical accessibility. I would welcome a national register of independently confirmed AI-assisted discoveries, with clear distinctions among computational predictions, laboratory replications, animal studies, clinical trials, and interventions actually in use. Such a register would not make science boring. It would make progress visible without pretending that every milestone is the finish line. People could then celebrate a finding for what it really established, and researchers could learn from failures instead of hiding them beneath the next announcement. That would be an advance worth paying for.

Savannah, Georgia

A container can be unloaded from a ship in a remarkably short period of time and then sit for days because nobody arranged the next leg of its trip. People outside transportation tend to think movement at the dock means the cargo is on its way. People inside transportation know better. We ask who has the chassis, whether the rail connection is available, whether customs information matches the shipment, and whether a warehouse is prepared to receive it. I think about that whenever I read a headline promising that artificial intelligence has accelerated scientific discovery. Which part has accelerated? A mathematical proof? The search through scientific literature? The proposal of a new molecule? And what is ready to receive the result on the other side? The answer matters more than the speed of the first handoff.

The research developments visible by late May 2026 are impressive enough that dismissing them as tricks would be unserious. Google’s Gemini Deep Think reached a gold-medal level on the 2025 International Mathematical Olympiad. Its February 2026 research update described extending this kind of problem solving toward more advanced mathematical and scientific questions. In biology, increasingly capable prediction and design tools have changed how scientists study proteins, and newly published systems such as Co-Scientist and Robin connect hypothesis generation to experimental planning and data analysis. Even someone who has never worked in a laboratory can understand why that could matter. A team that searches more efficiently, identifies a better candidate, and chooses the next experiment wisely should move closer to a useful finding.

I would still like to know how often that full sequence works. Port operations are full of performance measures that become meaningless when separated from the journey. A terminal may advertise crane productivity while trucks wait at the gate. A carrier may meet its sailing schedule while deliveries arrive late. Likewise, an AI system may process a million documents or generate thousands of proposals without shortening the time to a replicated discovery. That does not make it useless. It means researchers and funders need a complete map of the process. An impressive local speed improvement cannot be assumed to improve the entire route.

The reason literature analysis attracts me is that it resembles the documentation problem familiar to transportation. A shipment may depend on bills of lading, certificates, regulatory notices, inspection records, and contracts created by people who do not know one another. Scientists face their own ocean of documents, with studies written in different languages, disciplines, and conventions. An AI assistant capable of locating a neglected result, checking how a claim has been cited, or highlighting contradictory findings could spare researchers considerable time. That benefit is credible even if no machine ever makes an original discovery. Better information routing is itself productive infrastructure. But when information systems make mistakes, their smooth appearance is a danger. A document that looks complete may still describe the wrong cargo.

This is why I am skeptical of systems that provide beautiful scientific summaries without a way to inspect the underlying papers. A claim should travel with its source, its limitations, and an indication of whether it comes from direct observation or interpretation. Researchers should be able to open the original study and find the relevant evidence, not simply receive a confident paraphrase. The danger is not limited to invented citations. An automated summary may treat several papers drawn from the same dataset as independent confirmation. It may overlook a negative experiment because that paper is less frequently cited. It may combine incompatible definitions because they use the same word. None of those errors is necessarily dramatic. Any one of them can distort the decision about what experiment to run next.

There is also a strong argument for embracing these tools precisely because scientists are overwhelmed by available information. A researcher cannot read everything published in fields adjacent to her own. Useful connections may be lost because vocabulary differs between disciplines or because a paper appears in a journal outside someone’s usual attention. Computer-assisted retrieval can make discovery less dependent on whom a scientist happens to know. I see the appeal in that. One of the most frustrating things about port logistics is learning that somebody elsewhere in the chain possessed the correct information all along but had no reliable way to share it. A scientific system that reduces such missed connections could have a lasting impact even if its most impressive achievement is organizing knowledge rather than creating it.

The experimental studies are more difficult to interpret. The May publications concerning Co-Scientist and Robin provide stronger evidence than a demonstration consisting solely of generated text. Their teams report hypotheses linked to laboratory tests, including work involving candidate treatments and mechanisms of disease. That is important because a physical result does not owe anything to a model’s rhetorical confidence. The cells respond or they do not. The assay detects an effect or it does not. Yet I would ask whether the experiments were independently repeated and what happened to the suggestions that failed. If a research system offers a great many ideas and scientists publish the few successful tests, readers need the denominator to judge how efficiently the system operated.

I also wonder whether the tools are making laboratory work genuinely easier or merely moving costs downstream. A better literature survey may give a scientist fifty appealing targets rather than five. Somebody must supply the reagents, reserve the equipment, prepare the samples, hire the technicians, dispose of the waste, and review the data. Scientific research has a supply chain too. Instruments contain specialized parts and often require imported materials. Research facilities depend on utilities, shipping, insurance, maintenance, and trained labor. We should not treat those constraints as afterthoughts because software moved part of the process into the cloud. An abundant list of ideas can become a backlog if nothing expands the capacity to test them.

The financing implications are considerable. Universities and private laboratories may buy AI services that promise lower costs, while discovering that better candidate generation requires larger experimental budgets. For some projects, that would be an excellent trade: testing more promising ideas rather than wasting resources on poor ones. For others, it might become an expensive process of sorting machine-generated possibilities. The decision should be made with actual numbers, not enthusiasm. I would like institutions to report total cost per verified finding and time from a proposed hypothesis to a reproducible result. They should compare teams using AI with teams performing similar work without it. The difference between improved allocation and mere increased volume ought to be visible.

I am mindful of how much scientific progress depends on trust among specialists. Port workers, dispatchers, customs officers, warehouse managers, and truck drivers each see a different part of the same operation. We accomplish nothing when one group dismisses the expertise of the others. Research AI will work best when the computational team and the experimental team understand one another’s limits. Software developers should not assume a biologically interesting pattern is useful because it scores well on a benchmark. Laboratory scientists should not reject unfamiliar computational methods simply because they were not part of their training. The intelligent approach is to make each team’s claims legible to the other and insist on independent checks where their responsibilities meet.

The question of accountability is not academic. Imagine that a model ranks possible compounds based partly on a database containing flawed measurements. A research group commits a year to testing the highest-ranked candidates. When the results disappoint, who examines the chain of assumptions? The vendor may say the model merely offered suggestions. The laboratory may blame a lack of transparency. The funder may blame poor project management. Everybody can be partly correct while nobody repairs the process. Agreements should specify access to records, the version of the model used, the source of data, and the responsibility to investigate systematic errors. We should design the paperwork before the expensive cargo leaves the dock.

There are positive possibilities that reach beyond elite institutions. Researchers in smaller cities may have the expertise to study a practical problem but not the staff to survey a fast-moving international literature. A trustworthy AI assistant could extend their reach. Coastal communities, for example, face questions about fisheries, changing habitats, invasive species, saltwater intrusion, and disease conditions that may require knowledge scattered across oceanography, engineering, chemistry, and public health. A system that helps local researchers assemble evidence and design practical tests could deliver benefits not captured by a spectacular demonstration of mathematical reasoning. We should judge the technology in part by whether it strengthens work conducted outside the best-funded institutions.

On that point, the January 2026 research about AI’s influence on scientific careers should make us cautious. The analysis suggested that scientists using AI methods can produce more work and gain more citations, yet the research portfolio as a whole may concentrate on a narrower collection of subjects. That sounds familiar to anyone who has watched businesses optimize what they already know how to measure. The most profitable freight route attracts more capacity; the difficult rural connection remains neglected. Efficiency can reproduce an existing map of priorities, and the map may be incomplete. If AI repeatedly directs scientists toward rich datasets, it might make the well-studied even better studied while less tractable but socially valuable questions receive less attention.

Some enthusiasts would answer that the market will sort everything out. I would not dismiss private investment, which has plainly financed much of the computing progress. But scientific knowledge contains benefits that are difficult for one business to capture. A better understanding of an ecological process or a rare disease may matter enormously to a community without producing a lucrative immediate market. If public research agencies use AI to improve productivity, their funding decisions should preserve a place for neglected problems and long-term inquiry. A system that produces faster answers to fashionable questions does not eliminate the need for people to ask unfashionable ones.

Mathematics illustrates a different dimension of trust. The Olympiad achievement was strengthened by official grading of the solutions. An independent standard made the result easier to interpret than an internal declaration of success. Researchers exploring AI-assisted proofs should seek similarly clear checks where possible, including formal verification and independent expert review. More difficult research claims may lack such immediate criteria, but the lesson survives: define success before celebrating it. Do not let the same organization choose the challenge, mark the performance, and announce the historical meaning without scrutiny. A tool can be excellent and still benefit from outside examination.

One of my practical concerns is that an institution might neglect ordinary knowledge-management improvements while paying for sophisticated AI. I have watched organizations spend money on predictive dashboards while basic records remain inconsistent. The dashboard becomes a more attractive way to display confusion. Science has the same vulnerability. If laboratories do not publish methods clearly, retain negative findings, or maintain usable data, a language model may assemble an elaborate picture from an unreliable record. Better scientific data stewardship is not glamorous, but it is a prerequisite for trustworthy automation. The best model cannot repair an observation that was never recorded or a result that was discarded because it was inconvenient.

The pace of change makes these issues urgent. As models improve, laboratories will face competitive pressure to adopt them simply to avoid appearing old-fashioned. That can be productive when tools are ready and dangerous when the evaluation is incomplete. I would encourage early adoption coupled with measured pilots. Give two teams comparable tasks. Record the baseline cost and time. Let the AI-assisted team use its tools. Require both to produce source-backed findings and replicable experiments. Then publish not only their successes but their extra work, delays, and corrections. A business would recognize that as an operational test. Science ought to find the idea compatible with its commitment to evidence.

My answer is that the gains are already real at several points in the research journey. A model that completes demanding mathematical work, makes protein hypotheses more accessible, or helps design a successful laboratory experiment has done something consequential. But the extent of economy-wide or science-wide acceleration remains uncertain. We need to know whether the whole route becomes faster, what the new bottlenecks are, and whether independent operators achieve similar results. A system that moves information faster while leaving the next station overwhelmed has improved part of the process, not completed a transformation.

I would welcome the day when a research team in Savannah can use an AI system to identify a sound explanation for an environmental problem, test that explanation with local samples, and share a reproducible result with other coastal communities. That would not require announcing that machines had replaced science. It would require a trustworthy chain of evidence, real experimental capacity, and people willing to take responsibility for every handoff. We already know how to recognize a shipment that arrived safely. We should insist on similar clarity before declaring that an AI-assisted discovery has reached its destination.

Clermont, Florida

I consult with smaller professional-service businesses, and I have learned to distrust the phrase “pays for itself” when nobody has defined either side of the calculation. A software vendor can demonstrate that a task takes fifteen minutes instead of two hours. The owner buys the system, discovers that employees must check the output, learns that integrations cost extra, and notices that the clients have not agreed to pay more for the work. None of this means the software was a mistake. It means technological capability, productivity, and financial return are three different things. The same distinction is fundamental to the debate about whether artificial intelligence is accelerating scientific discovery. The technical milestones are genuine. Their economic and social consequences must still be established rather than assumed.

This May offers more evidence than we had even a year ago. Mathematical systems have progressed from specialized competition demonstrations to gold-medal-standard work on the International Mathematical Olympiad. DeepMind has described systems attempting more advanced research-level questions, although claims about internal benchmarks deserve independent scrutiny. Protein-structure tools have changed how researchers prepare experiments. The newly published Co-Scientist and Robin projects show how artificial intelligence can organize literature, suggest hypotheses, and contribute to laboratory-tested findings. These developments are too substantial to dismiss as marketing. Yet they do not answer the ordinary business question of whether an institution purchasing the technology can now produce more validated knowledge with the same resources.

A service firm purchases technology to improve one of several things: speed, accuracy, capacity, customer experience, or the cost of delivery. Those categories apply to scientific institutions surprisingly well. A literature-review tool might reduce hours spent assembling background research. A proof assistant might help a mathematician explore logical possibilities. A protein designer might produce a shortlist that saves unnecessary laboratory work. A research agent might connect experimental results to a useful follow-up idea. Each could be valuable, but the value is not interchangeable. If the underlying limit is access to laboratory instruments or experienced researchers, additional machine-generated hypotheses may add little to overall output. A system should be evaluated against the bottleneck it actually relieves.

A simple hypothetical illustrates the issue. Imagine a research group spends one hundred working days moving from a question to a validated result. Twenty days involve gathering literature, another twenty involve developing hypotheses and planning tests, and sixty involve experiments and confirmation. Suppose AI reduces the first forty days to ten. The result is an excellent thirty-day saving, but it is not a fourfold improvement in the full process. If faster ideation produces twice as many experiments and the laboratory cannot expand, the waiting line may lengthen. That example is not a claim about measured science. It is a reminder of the mathematics of workflow. Faster tasks do not automatically create faster projects.

There is another complication in research that ordinary consulting firms recognize: the customer may not be the purchaser. Universities buy tools, scientists use them, journals evaluate the resulting manuscripts, companies license discoveries, and the public may ultimately fund or benefit from the results. Each participant has different incentives. The university may prize grant revenue and publication count. The researcher may want genuine knowledge but also needs tenure or continued employment. The software vendor wants recurring subscriptions and evidence that its product changes the field. The public wants useful science conducted responsibly. We should expect these interests sometimes to align and sometimes to diverge. Good institutional design takes the divergences seriously.

The strongest evidence of acceleration would compare the total cost and time required to produce a finding that independent investigators can reproduce. This is harder than counting submitted papers or modeled molecules, but that is the point. A scientific result with a beautiful abstract and weak evidence is not equivalent to one that survives multiple tests. A model may generate an apparently coherent argument faster than a researcher, but if the researcher spends the saved time locating a subtle error, the net benefit changes. In a consulting office, we do not call a job profitable because the first draft was fast. We count the revisions, mistakes, client calls, and time needed to finish it correctly.

Publication itself is a questionable output measure. This March, Nature described an AI Scientist pipeline able to prepare machine-learning research manuscripts, including work that met a workshop’s initial review threshold. That is a meaningful proof of capability. It could also contribute to an already difficult problem if automated paper production grows faster than qualified review. A scientific institution that generates thousands of manuscripts may increase the work demanded from referees without adding comparable value. I would want to see the proportion of findings that other groups use, extend, or reproduce. Faster document production can be a benefit, but the publishing system ought not to become the dumping ground for cheaply manufactured plausibility.

The January Nature study that found a tension between individual researcher productivity and the collective breadth of scientific inquiry is especially relevant to incentives. People naturally follow rewards. If AI works best where structured datasets and plentiful prior papers exist, it may make those topics even more attractive. Researchers may choose projects they know the models can help complete, even when a less fashionable problem is socially more important. Institutions could record strong metrics for each scientist while allowing the overall research portfolio to become less diverse. Small businesses face a version of that trap when software makes their most profitable existing service easy to sell and they neglect the next service customers will need.

I am inclined to defend experimentation by private companies, because the cost of a mistaken bet should not always be imposed on taxpayers. If a biotechnology firm believes AI will shorten its discovery process, let it spend money, publish appropriate evidence, and compete. But public research agencies have a different obligation. They must support foundational knowledge, independent testing, and questions without an immediate paying customer. One of the great promises of AI is that it could reduce the cost of exploring neglected fields. We should structure public funding to test whether it does so. An algorithm that merely helps well-funded laboratories publish more on subjects already attracting money is not the same public achievement as making reliable tools available to small institutions.

Access pricing deserves closer attention than it usually receives. The underlying systems may involve expensive computing, proprietary models, research databases, specialized services, and usage limits. A university might become dependent on a platform that later changes its terms. A laboratory could build a workflow around a model only to discover that a new version behaves differently. Scientists need predictable access to methods if their findings are to remain reproducible. Private property rights are real and important, but so is the public interest in research that can be checked. Licensing agreements could include provisions for preserving the software version, retaining audit logs, and enabling independent scientific inspection under reasonable protections.

A market also needs honest comparisons. I would like independent evaluations in which several groups tackle equivalent research problems under clearly described conditions. Some groups would use particular AI tools; others would use conventional methods; still others might combine them in different ways. Evaluators would count human labor, computing cost, experimental materials, error correction, and outside validation. They would also record whether the AI tools widened or narrowed the range of approaches researchers considered. Such tests would not establish every possible scientific benefit, but they would give buyers and funders a much firmer basis for decisions than vendor-selected stories. A small business owner who has been burned by an exaggerated technology claim would recognize the value immediately.

We should not forget that some discoveries are valuable precisely because nobody knew how to look for them. An AI system trained to detect patterns in accumulated literature may be exceptionally good at extending familiar lines of inquiry. It may be less capable when the important step is rejecting a seemingly reasonable premise. We should reward researchers who test suggestions that contradict model consensus and investigate problems for which datasets are weak. Otherwise, we risk turning scientific agendas into the equivalent of search-engine recommendations: increasingly efficient, increasingly personalized, and less likely to surprise us. Creativity has an economic value even when it cannot be neatly scheduled in a research budget.

There is, however, a practical benefit that enthusiasts sometimes understate by focusing on dramatic discoveries. Scientists often spend years learning enough adjacent disciplines to work effectively on a new problem. A careful AI assistant could shorten that preparation. It might explain unfamiliar terminology, retrieve key foundational papers, or help a physicist understand a biological technique well enough to collaborate productively. It could reduce the transaction cost of interdisciplinary work. In business, a tool that lets competent people coordinate across specialties often creates more value than one that does a narrow expert task spectacularly well. I can imagine a similar effect in science as long as the explanations remain linked to verifiable sources.

The biological examples deserve both praise and disciplined accounting. The new research assistants have reportedly contributed to hypotheses that were tested in cells, including work relevant to drug repurposing. That provides stronger evidence than an unsupported claim of theoretical potential. But an in-vitro result is early in the value chain for a medical therapy. Pharmaceutical development includes toxicology, dosing, clinical trials, manufacturing, regulation, and the question of whether a finished treatment improves outcomes. Even if AI doubles the speed of identifying plausible compounds, it may have little effect on the largest costs later in development. It could still be extremely valuable if it reduces failure at those later stages. We need evidence of which effect is occurring.

One reason I am optimistic is that knowledge work has been difficult to measure and improve systematically. Scientists have long relied on ingenuity, experience, and a degree of luck, with information scattered across repositories and publications. Computational tools can reduce the friction of searching and organizing that information. The question is not whether every discovery needs a human touch at each step. There is no virtue in forcing skilled people to perform work machines can do reliably. The question is whether the time saved is used for stronger experiments and better decisions or simply converted into additional output that nobody has time to examine. The difference will be determined partly by incentives, not by algorithms alone.

An institution considering major investment should therefore prepare something like a business case. It should identify the research activity being improved, estimate the current cost and rate of reliable success, specify the new capabilities expected, and describe how the claims will be tested. It should budget for scientific validation, security, maintenance, training, and model changes. It should name the person responsible for evaluating whether the technology continues to perform. It should also plan an exit path. None of this is romantic, but there is plenty of romance in scientific inquiry already. The purchase order can afford a little discipline.

I do not want the debate reduced to believers and skeptics. The mathematical demonstrations are genuine achievements in defined settings. The protein advances can make new experiments practical. The recent multi-agent systems present credible examples of AI contributing to laboratory research. At the same time, the full economic return remains uncertain, and the danger of flooding science with attractive but weak work is not hypothetical. The sensible conclusion is differentiated: some tasks are already much faster, some research projects are meaningfully improved, and broad acceleration across all of science has not yet been established.

Here in Clermont, growth has taught us to distinguish a promising development announcement from roads, schools, and utilities able to support the people who arrive. Scientific AI faces the same test. The announcement is that ideas can be produced faster. The investment decision is whether the institutions needed to verify and use those ideas can grow with them. When an AI tool helps researchers reach reliable, important results more efficiently, we should adopt it enthusiastically. When it primarily produces more claims at greater speed, we should stop confusing traffic with progress. The future belongs not to the organization generating the most possibilities, but to the one capable of turning good possibilities into knowledge that survives examination.

Novi, Michigan

If you ask someone outside the automotive business what makes a vehicle, you may get an answer involving design, horsepower, and the badge on the front. Ask someone who manages supply chains, and we will talk about suppliers you’ve never heard of, components that must arrive in the right sequence, software revisions, tooling, inspection, and the problem caused when a tiny part from one plant becomes unavailable. Scientific discovery has a supply chain too. It includes ideas, publications, data, computing capacity, laboratory materials, equipment, people, and institutions able to check the finished claim. Artificial intelligence is starting to transform some of those connections. I believe the gains are real. I also believe the public discussion neglects the constraints that determine whether faster ideas become useful results.

Start with the mathematical examples because they are exceptionally clear. A model reached gold-medal-standard performance on the 2025 International Mathematical Olympiad. That success has external grading and defined problems, which make it much easier to evaluate than a vendor’s claim that its system is “creative.” More recent work describes the use of advanced reasoning models on research-level questions. I take these developments seriously. Yet solving six well-defined problems on a schedule differs from producing a sustained stream of new mathematics that experts consider meaningful. A high-performance test engine does not automatically become a production vehicle. The demonstration shows what is possible under specified conditions. The system’s routine contribution to scientific work still needs to be measured.

In biology, the relationship between computational advances and real output is more complicated. AlphaFold models can provide structural predictions that researchers use to plan experiments. Generative methods can propose proteins that have never existed in nature. Those tools may narrow an enormous design space, much as optimization software can suggest a component geometry that a human designer might never have considered. But predicted geometry alone does not establish strength, manufacturability, durability, or function. In biology, it does not establish whether the protein folds correctly in a real cell, binds only the intended target, remains stable, or produces unwanted effects. At some point, somebody has to make the thing and test it.

This is where I see the largest potential productivity gain. Many scientific projects fail not because researchers cannot imagine an idea but because the number of possible experiments dwarfs the capacity to conduct them. If an AI system can select a much smaller and better-ranked group of candidates, the result may be a major improvement in laboratory efficiency. It reduces the experimental equivalent of ordering parts that will not fit. The recently reported AI-assisted biological research suggests that such ranking and planning could be practical in selected settings. Yet any meaningful assessment has to count how many candidates were generated, how many were tested, how many succeeded, and how independent laboratories performed. Success stories selected after the fact do not reveal the full yield.

A qualified automotive supplier can spend months demonstrating that a component repeatedly meets requirements. The customer wants to know what happens across temperature cycles, vibration, wear, and ordinary variation. Scientists need a parallel mentality when designing new proteins or other biological structures. A candidate should be tested under conditions relevant to its intended use, not merely under a convenient assay. Researchers should measure specificity, stability, reproducibility, and unexpected behavior. A model may recommend a promising molecule, but a single favorable measurement is only the beginning of qualification. The discipline is familiar: document the specification, validate the process, test the product, and examine the cases where it failed.

The software version becomes part of the supply chain as well. Our industry has learned that an apparently small change to a supplier’s process can create defects in a downstream product. AI research workflows must account for model updates, changed training sources, revised retrieval services, and different computational settings. Two laboratories might believe they are repeating an experiment with the same AI tool while using different underlying models. The gap is easy to miss when software is accessed through a simple online interface. Reproducible research requires more than retaining the final prompt. It requires a record of the relevant system, the information made available, the choices made by humans, and the tests that followed.

We should be careful about the temptation to assign credit to a single spectacular technology. An automobile reaches the showroom because thousands of improvements in materials, controls, machining, testing, and coordination become part of one working system. Scientific progress often has the same character. A new biological finding attributed to AI may also depend on decades of experimental research, public databases, earlier computational methods, and the judgment of investigators who designed the study. That does not diminish the AI contribution. It locates it accurately. When we identify which stage made the difference, we can improve the entire system instead of chasing the most photogenic component.

The growth in machine-generated literature could become a serious supply-chain defect. Research assistants are supposed to extract useful knowledge from publications. What happens when publications increasingly contain unverified machine-generated claims? The input inventory becomes contaminated. A model may cite an attractive article; another model may summarize it; a third may propose an experiment based on both; and the repeated mention may create the impression of independent support. This is a variation on a counterfeit-part problem. A defective component can pass through multiple distributors and acquire the appearance of legitimacy because it arrives with respectable paperwork. The remedy is provenance: verified original sources, independent tests, clear authorship, and a route back to primary evidence.

I find the January 2026 study on scientific productivity particularly sobering. It suggested that AI use may increase researchers’ individual output while concentrating inquiry in fewer areas. Businesses face the same pressure when data systems encourage managers to optimize products with plentiful performance information. The organization gets more efficient at what it already does and may neglect unusual opportunities. A supplier whose forecasting software always favors familiar customers can miss the growth of an entirely new industry. Science needs its equivalent of exploratory capacity: funding and time for questions with poor data, unfamiliar methods, and uncertain returns. Otherwise, a faster research machine may be pointed toward an increasingly narrow destination.

Some commentators imply that AI itself eliminates scarcity. It does not. Research depends on physical infrastructure. Laboratories need reliable electricity, cooled equipment, chemicals, biological samples, precision instruments, and trained technicians. High-end computing requires chips, networking equipment, cooling systems, and power supplies. Many of these components involve global supply chains that can be disrupted by trade restrictions or manufacturing delays. The field may be able to produce candidate designs at extraordinary speed while remaining dependent on a small number of hardware providers or specialized research facilities. Strategic planners should consider those dependencies before describing a future in which all scientific exploration becomes effortless.

The international aspect matters to a manufacturing region like ours. American universities and firms may develop excellent AI research systems while relying on imported equipment or specialized materials for experimental validation. Other countries may have powerful laboratories but insufficient computing access. Scientific collaboration can ease these constraints, while excessive restrictions or concentration could deepen them. Protecting sensitive technologies has legitimate purposes, particularly where national security is involved. But broadly shared scientific knowledge is also a source of innovation. We need to distinguish tools that require safeguards from ordinary research cooperation that benefits participants on all sides.

A concern of mine is that software vendors will sell research institutions “integrated discovery” before the integration is mature. That phrase can mean everything from a helpful literature assistant to an agent that proposes experiments and analyzes results. Those capabilities have different requirements and risks. A buyer should insist on demonstrations using unfamiliar questions, with appropriate control groups and clear human oversight. The system should not be evaluated solely on a handful of previously publicized examples. Supplier qualification in our business would be absurd if we accepted only the specimens the supplier had selected for display. A scientific purchasing department should be no less demanding when a platform’s recommendations will determine substantial research spending.

I do not want to underplay the promise for smaller laboratories. One reason manufacturing has become more sophisticated is that powerful design and simulation tools have spread beyond the largest corporations. A capable small supplier can now model a part, share detailed specifications, and collaborate across distance in ways that once required far more resources. AI may do something similar in research. An investigator at a modest university could search specialized literature, explore a mathematical formulation, or generate a set of protein hypotheses without possessing the full computational staff of an elite institution. If the results are transparent and affordable to verify, that diffusion could have more durable value than a series of record-setting demonstrations.

But diffusion requires training, not just licenses. Engineers who rely blindly on simulation can make expensive mistakes. Researchers who accept AI-generated hypotheses without understanding their assumptions may do the same. Scientific education should teach what models represent, what they omit, how datasets were assembled, and why a plausible result may fail experimentally. Students should learn to challenge outputs and preserve alternative explanations. Their instructors should encourage the use of AI where it expands understanding while requiring independent work on the foundations. An effective tool should make an informed worker more capable, not turn the work into a black box.

Another issue is how organizations measure a return on research investments. A technology that reduces the cost of one experiment by half is valuable even if it does not immediately produce an important discovery. A technology that increases the number of experimental candidates might be valuable if it improves the odds of success. But these benefits should be measured in terms that reflect the actual project. Counting proposals, publications, or model tokens is like counting parts shipped to a factory without checking whether the vehicles were completed. The more meaningful indicators include successful independent replications, reduced wasted experiments, shorter time to verified insights, and new lines of inquiry that become practical.

There is a legal and commercial layer we should not ignore. Scientific discoveries can become intellectual property, and the ownership of AI-assisted work may be disputed. A company might contribute proprietary data to an external service without understanding how that data is used. A laboratory may later struggle to show exactly which method produced an invention or whether the underlying model may be used in commercial development. These are practical barriers to adoption. Institutions should negotiate clear terms for confidentiality, use rights, auditability, and retention of records. The absence of such rules could deter collaboration or create expensive disputes long after the original research team has moved on.

I am optimistic because the basic direction is right. Better tools for reasoning, literature analysis, and biological design can make expert scientists more productive. The evidence through this May includes demonstrations that go beyond generating polished text, and that deserves recognition. The optimistic case becomes credible when the tools enable a consistent sequence: choose an important question, identify sound evidence, generate a useful candidate, test it, replicate it, and explain what changed. The pessimistic case becomes plausible when organizations skip the final stages and celebrate volume. Both futures are possible because institutions, not only algorithms, determine how research is conducted.

My conclusion is that the acceleration is real but uneven, like the modernization of a complex manufacturing supply chain. Some stations have become dramatically faster. Others remain constrained by physical testing, human judgment, money, and the ability to verify results. We should stop asking whether AI has already replaced the scientist and ask whether it has improved the whole production of knowledge. In Novi, I have watched companies succeed by measuring the complete process instead of applauding one fast machine. Science will achieve more by doing the same. The breakthrough becomes meaningful when it arrives in usable form, meets its specifications, and keeps doing so when the original team is no longer standing nearby.

That standard also prevents us from missing progress that is not yet glamorous. If one laboratory now asks a better question and avoids fifty unproductive tests, the improvement deserves attention even if nobody can give it a dramatic name. If another laboratory uses a model to produce a striking diagram but cannot reproduce the proposed effect, the record should say so with equal clarity. A useful scientific culture keeps both stories visible. It learns from a failed experiment rather than burying it, because the negative result may prevent the next group from wasting the same resources. Faster science must include faster learning from mistakes, not merely faster announcements of success.

Cedar Rapids, Iowa

In accounting, the first thing I want to know about an exciting number is what is included in it. A company announces record sales; I ask what it collected. An economic-development committee announces new jobs; I ask when those positions were filled and how long they lasted. Somebody promises the city a magnificent return on an investment; I ask whether maintenance, debt service, and the cost of land are in the calculation. This habit can make me less welcome at celebratory meetings, but it has saved people from paying for stories that did not survive contact with the ledger. Artificial intelligence is now producing a marvelous collection of scientific headlines. I would like to see the accounting behind them.

We know enough by May 2026 to say that several technical advances are real. Artificial intelligence has achieved gold-medal-standard performance on the International Mathematical Olympiad, and the grading of those solutions by officials gives the claim substance. Protein prediction systems have made information available to researchers on a scale that would have been difficult to imagine not long ago. New research assistants have contributed to hypotheses and experiments that can be examined outside their own software environment. I have no interest in pretending these events are meaningless. I want to avoid replacing a cautious assessment of productivity with a contest to see which institution can attach the biggest adjective to its latest announcement.

A simple accounting principle would help. Every reported improvement ought to have a baseline, an adjustment for resources used, and a clear description of the resulting product. In mathematics, a valid proof is a meaningful output, although a competition result should not be confused with a research career. In biological experimentation, a promising candidate is an intermediate product; a replicated laboratory finding is another milestone; a therapy that demonstrably benefits patients is something more. In literature analysis, the mere number of articles read by a program does not establish that its summary is correct. We should track each accomplishment separately, rather than add them together as though every stage were the same item in inventory.

The labor costs are especially easy to misstate. An AI-assisted researcher may produce an initial proposal in hours rather than days. But the team may need to verify citations, repair calculations, repeat experiments, and train staff to understand the tool. A bill for computing services may replace a portion of payroll without reducing total cost. That can still be worth doing if the finished work is more reliable or if the same people can undertake additional valuable research. Yet nobody should claim a productivity dividend solely by measuring the time before the first draft appeared. In plant accounting, we calculate the full cost of the unit that leaves the factory, not just the time required to start the machine.

One of the research reports that interests me most appeared in January. It suggested that scientists who adopt AI methods can improve individual publication and citation outcomes while the range of scientific subjects receiving attention becomes more concentrated. In financial terms, the institution may improve the returns reported by each department while increasing its exposure to the same underlying risk. Everyone becomes more productive within familiar categories, and the organization ends up with a narrower portfolio. That is a problem for scientific discovery because the great result may come from the neglected problem that conventional measures do not reward. A well-run portfolio includes some investments that are difficult to model and may never pay off. Without that tolerance, an institution can optimize itself into irrelevance.

The emerging systems that produce scientific papers illustrate the need to distinguish activities from outcomes. A research pipeline reported in Nature this March can carry out portions of the computational science process and assemble manuscripts. One paper in the report reached the review threshold of a workshop. That is evidence of a real new capability. It is not evidence that the system can fill a journal with discoveries of the same significance as independently confirmed advances in biology or physics. Workshops, journals, and fields have different standards. Even when a paper is accepted, its conclusions may later prove wrong. The number of manuscripts prepared at low cost is an efficiency measure for writing and experimentation, not an automatic entry in the ledger of new human knowledge.

A related concern is that the cost of reviewing the output may fall on people outside the organization generating it. Academic referees often perform unpaid or lightly rewarded work. Scientific databases and archives require funding. Independent laboratories may need to replicate a finding before it can be used safely. If AI systems make low-quality submissions much cheaper to produce, the external cost could rise dramatically. The private institution could report increased productivity while the scientific community absorbs a larger verification bill. Economists have a name for effects of that kind, but it does not take an economist to understand the unfairness. The operation is not truly efficient if somebody else has to clean up the waste without being counted.

I would insist on measuring rejected hypotheses as well as successful ones. Imagine one system proposes two thousand protein structures and ten are found useful. Another proposes fifty and ten are useful. Both can advertise ten successes; only the second has provided clear evidence of superior selection if costs and conditions are comparable. Now suppose the first system runs so cheaply that the total bill remains lower. The comparison changes again. This is why denominators, laboratory costs, and time to validation matter. A responsible report should provide enough detail to reconstruct the comparison. A few dramatic examples might be good publicity, but they are not management information.

I do not necessarily agree with those who say government should avoid the field because private companies will make the necessary investments. Basic science does not always generate returns that a single investor can capture. A project mapping a poorly understood environmental process might be valuable to many communities without supporting an obvious commercial product. The federal government already supports universities and laboratories because the long-term public benefits can exceed private incentives. The right question is how to make that spending accountable without requiring every project to promise immediate profit. AI might make basic research more efficient, but public agencies should pay for independent evaluations of that claim rather than treating a vendor’s forecast as a fact.

At the same time, I would not allocate enormous sums merely because a new tool is fashionable. Research grants should state what new capability is needed, why existing methods are insufficient, and how success will be assessed. Investments in high-end computing may be justified, but so might investments in instruments, laboratories, data stewardship, or technicians needed to test the resulting predictions. An institution that builds a giant computing facility without budgeting for the scientific work around it resembles a town that borrows for a new factory and discovers too late that the road and water system cannot support it. Capital expenditure is only one part of the commitment.

The issue of private platforms concerns me because it makes long-term costs harder to estimate. A public university may train researchers around one company’s model, only to face higher prices or changes in service. That is a dependency, and dependencies can be expensive even when the initial purchase looks attractive. Scientific reproducibility adds a special concern: a research group must be able to describe the conditions under which a claim was produced. Institutions should negotiate for version records, data retention, appropriate inspection rights, and the ability to switch providers where practical. Those provisions are not hostility toward business. They are normal precautions when the institution cannot afford a sudden loss of essential services.

I also think the discussion of mathematics deserves more careful language. A gold-medal-standard score on a hard examination is not the same as an award bestowed on a machine as a human competitor, but the underlying result is impressive without exaggeration. The ability to produce solutions that official graders recognize as correct is a substantial demonstration. Yet science involves choosing worthwhile questions and interpreting their importance as well as answering them. A mathematical proof may be valid and still add little to a field; a difficult conjecture may remain important even when a system has no path to solve it. A reasonable evaluation would separate problem-solving competence from a sustained contribution to research programs.

Biology provides a strong reason not to insist on a single metric. A computational structure prediction can be useful even if it requires experimental confirmation, because it may substantially narrow the search for a workable explanation. A new model can make a previously inaccessible class of molecular designs easier to explore. A research agent can suggest combinations that humans did not notice, as recent laboratory-linked work suggests. These tools may create options whose benefits do not show up immediately in a count of completed discoveries. A balanced accounting should record intermediate assets and their probability of leading to useful work while resisting the temptation to treat every asset as already realized income.

Our experience in manufacturing also tells us that technology shifts can take years to affect overall productivity. Plants may install new machines while learning to reorganize maintenance, staffing, materials, and quality control. Initial results can be disappointing even when the technology eventually changes the economics of production. AI-assisted science may follow a similar path. Institutions must revise training, data practices, publication standards, and experimental planning before they receive the full benefit. Early measures could understate the potential. That is not a reason to abandon measurement. It is a reason to take measurements repeatedly and explain the transition costs rather than announce victory after the first favorable quarter.

I would support a public reporting framework for major AI-assisted research projects. Each project could state the scientific question, the baseline workflow, the precise AI contribution, the cost of computing and validation, the independent evidence obtained, and the unresolved limitations. Positive and negative results would both be included. This would make it easier for other institutions to assess whether a tool was relevant to their work. It would also reduce the ability of a project to redefine success after seeing the outcome. I am not proposing an accountant’s veto over imagination. Researchers should remain free to pursue uncertain ideas. I am proposing that when they claim a measurable improvement, they should use numbers capable of being checked.

A concern that gets less public attention is the possibility of duplicated computational effort. If hundreds of laboratories employ similar models to investigate the same heavily publicized problems, aggregate spending could rise without a comparable increase in unique findings. Better tools for sharing negative results, identifying ongoing work, and coordinating open research may create more value than simply buying more computational capacity. The academic incentive to be first can discourage openness, so policies and funding structures would need careful design. Science is a competitive enterprise, but its knowledge is cumulative. The public has an interest in avoiding needless repetition while preserving independent replication where it is scientifically necessary.

There is a distinction between redundant work and replication that administrators frequently misunderstand. A repetition designed to check a result is valuable, especially where it could expose an error. A repetition conducted because nobody can find a prior result is often wasteful. AI literature tools could help distinguish the two by locating research and explaining methodological differences. That would be a particularly useful kind of acceleration. Instead of merely making individual scientists quicker, it could reduce waste across institutions. The gain would appear in avoided expenditure and stronger confidence, not necessarily in a large new number of publications. These are precisely the improvements conventional publicity measures tend to miss.

My answer is that AI is already accelerating particular scientific tasks, and the evidence includes more than convenient demonstrations. The May biological research and the mathematics achievements show real capabilities. But we have not established a simple multiplier for discovery across all sciences, and anyone offering one should explain the baseline and the denominator. Productivity can improve while important questions remain neglected. Output can grow while verification costs rise. An institution can publish more papers without generating proportionately more reliable knowledge. Those are not reasons to reject the tools. They are reasons to maintain honest books.

I would love to see more scientific discoveries, especially ones that eventually improve agriculture, manufacturing, health, and the ordinary communities that often receive little attention from research announcements. Faster and cheaper methods could help accomplish that. But until a claimed advance survives independent examination, it belongs in the column for promising work, not completed accomplishment. The people who pay for research deserve both ambition and a faithful accounting. We should applaud a breakthrough when it is real, record a failed experiment when it is not, and stop treating promotional output as a substitute for an audited result. That approach served me in a plant office for decades. I expect it would serve scientific institutions just as well.

Las Vegas, Nevada

I spent enough years in casinos to recognize a display designed to hold an audience’s attention. The bright lights are real, the game is real, and the payout may be real. None of that tells you what the player should expect over a thousand visits. The odds reside in the full record, not in the one win everybody remembers. I approach artificial intelligence in scientific research with similar instincts. A machine that solves a difficult mathematical problem is impressive. A research assistant that suggests an experimentally useful hypothesis is exciting. But institutions investing money and reputations in these systems need the equivalent of a transparent rulebook, a tested process, and an honest accounting of outcomes. Otherwise, the public could be watching a series of extraordinary demonstrations and mistaking them for normal performance.

The achievements are not imaginary. The International Mathematical Olympiad result from 2025 deserves credit because outside graders evaluated the work. The advances in protein structure prediction have given biologists useful tools. More recently, systems described in scientific journals can combine literature analysis with hypothesis generation and experimental planning. These are developments we should welcome, especially if they spare researchers months of repetitive work. But I hear some people discussing them as though science is about to become a fully automated entertainment program: ask a question, watch the machine think, receive a miracle. Anyone who has tried to operate a real enterprise knows that dependable service is less theatrical and more demanding than the opening-night demonstration.

In hospitality, the customer sees an event; managers see labor schedules, suppliers, safety inspections, insurance, training, contracts, and the equipment that fails at exactly the wrong hour. A successful performance requires all those systems to cooperate. Scientific discovery likewise depends on a chain of support. An AI system may suggest a molecule, but researchers need the ability to produce it, measure it, evaluate its safety, and reproduce results. A mathematical model may propose a proof, but qualified reviewers must determine its validity and significance. A literature assistant may save time, but its sources must be accurate. If the output cannot survive those later steps, the speed of generation is merely a feature of the presentation, not an improvement in the finished service.

The field should adopt a distinction familiar to operators: uptime versus peak performance. A machine may perform spectacularly when conditions are ideal but fail frequently in ordinary use. A research AI system may produce excellent suggestions on selected benchmark questions yet behave unpredictably on messy, poorly documented scientific problems. I would want to know its rate of useful answers across a representative collection of tasks, how much expert correction it requires, and whether performance declines when datasets are incomplete. I would also ask what happens when the vendor changes the software or the institution loses access. A laboratory needs to know whether the tool will function reliably enough to justify building its routine operations around it.

The January Nature research about AI’s influence on scientific agendas strikes me as a warning about incentives. If researchers using AI publish more and receive more citations but collectively investigate a narrower range of questions, the visible scoreboards may be improving while the underlying game becomes less interesting. That can happen in business too. Managers learn to maximize the metrics that owners review. They may succeed at the measured objective while neglecting services that are less easy to count. Science cannot afford a system that rewards only the work already compatible with automated tools. Some of its most important contributions come from people who notice that the established story does not explain an inconvenient observation.

I am also struck by the potential for scientific publishing to become overcrowded. The AI Scientist study published in March demonstrated that automated systems can assemble computational experiments and manuscripts, and one of its papers reached a workshop’s review threshold. That achievement is real, but it points toward an operational problem. If generating a paper becomes cheap and evaluating a paper remains expensive, journals could be flooded. Reviewers might spend their energy filtering repetitive, weak, or inaccurate submissions. The institution generating the papers would count productivity, while the rest of the scientific community inherited the screening burden. In hospitality we call that shifting work onto somebody else’s crew. It usually ends with angry customers and a strained operation.

There is nothing automatically virtuous about human-only research, either. Scientists have always made mistakes, chosen flattering results, overlooked literature, and defended favored explanations longer than they should. A well-designed AI tool could reduce those failures. It might identify contradictory studies, suggest appropriate controls, or flag a calculation that does not reconcile. It could test assumptions with more patience than a tired graduate student working after midnight. The question is how to structure the system so that its mistakes are detected and its strengths are used. A good safety program does not begin by declaring workers perfect or machines dangerous. It asks where failures can occur and how the operation will catch them before they cause damage.

A research organization should know who owns each decision in an AI-assisted workflow. Who approves the scientific question? Who checks a model-generated citation? Who decides that a proposed experiment is ethically appropriate? Who interprets an ambiguous result? Who signs the final paper, and what precisely are they asserting? Those are elementary questions of management, not philosophical puzzles. A system with impressive capabilities can still fail when responsibility is diffuse. I have managed events where everybody assumed someone else had confirmed a vendor or inspected equipment. The result was a very ordinary problem with a very expensive resolution. An AI lab without clear accountability invites a more sophisticated version of the same mistake.

The recent Co-Scientist and Robin reports offer a constructive model because they connect computational work to laboratory evidence. Their teams do not merely ask whether the system can produce plausible paragraphs. They describe specific proposals tested in biological settings, including cultured cells. That matters. Still, an institution should distinguish its internal research result from broad outside confirmation. The experimental setting may have been selected because it suited the method, and the system may have relied on human expertise to recognize which suggestions were worth testing. These qualifications do not negate the finding. They tell us how to interpret it. Scientific progress is strengthened when a result arrives with an honest description of the conditions under which it occurred.

One practical issue that deserves more attention is the cost of computing. Research AI is not free merely because the researcher sees a text box. The infrastructure requires data centers, chips, electricity, cooling, and engineers. The most advanced systems may impose usage limits or require expensive institutional arrangements. A laboratory can become dependent on a tool whose pricing or availability it does not control. Hospitality businesses have learned the risks of building their operation around platforms that can change the terms of access overnight. Scientific institutions should negotiate records, continuity provisions, and alternatives rather than assume a vendor’s current generosity will last indefinitely.

My skepticism is not an argument for a government bureaucracy to approve every scientific use of AI. That would be a fine way to make ordinary research slower without making it better. I would favor practical standards of disclosure and independent testing, especially when public funds support a major project. If a company claims its tool can shorten a discovery process, let outside investigators evaluate it using meaningful benchmarks. If a university publishes an AI-assisted finding, provide sufficient methods and records for replication. If a system changes, say so. These conditions would make the market more credible, not less competitive. Investors and researchers would be able to distinguish tools that work from tools with attractive promotional departments.

We also need to separate scientific significance from commercial value. A new mathematical proof may be important even if no product uses it for decades. A protein-design method may save an enormous amount of work without producing an approved treatment immediately. A small improvement in literature analysis could benefit thousands of researchers while making no headlines. Conversely, a public demonstration that attracts enormous attention may have little impact on real scientific questions. Those distinctions resist a single scoreboard. The industry will have to live with more than one measure of success. Science is not a slot machine that pays out on a published schedule.

I worry about younger researchers becoming dependent on polished output before they learn how to recognize a bad answer. That is a training problem. Employees who know only the automated reservation system cannot solve a problem when the network goes down. Scientists who know only how to request a machine-generated explanation may struggle when an observation contradicts the model. Educators should use AI to expand the range of problems students can study while requiring them to inspect sources, defend assumptions, and understand experimental design. A useful assistant should increase the scientist’s capacity for judgment, not reduce the profession to selecting from menus of suggestions.

The technology also makes scientific collaboration potentially more inclusive. Researchers without large teams may be able to search unfamiliar fields, compare methods, and explore ideas that were previously beyond their immediate expertise. I hope that advantage reaches smaller universities and institutions outside the established research centers. But access to software is not the same as access to laboratories and instruments. If only a few wealthy organizations can afford to validate the generated ideas, the concentration of power could remain. Public agencies and foundations should consider supporting shared experimental resources alongside open or affordable computational tools. Otherwise, we may create a democratized market for hypotheses and an exclusive market for proof.

There is a reason casinos account carefully for unusual outcomes. A rare win tells an exciting story, but responsible management depends on long-run patterns. Research organizations should publish something resembling a full record of AI-assisted projects: questions attempted, suggestions tested, time spent, resources used, and results replicated. They need not reveal confidential commercial details in every instance, but the public claims should be specific enough to examine. A company can legitimately boast that a new tool helped identify a biological candidate. It should not imply a treatment is available when the work has not progressed beyond a preliminary experiment. The distinction protects both the public and serious innovators from disappointment.

I see genuine acceleration in pieces of the scientific process. Literature can be searched more effectively. Mathematical systems have reached meaningful milestones. Biological tools can suggest structures and pathways that scientists might not have found efficiently on their own. Some of the recent integrated research demonstrations provide evidence of useful experimental contributions. We should not minimize those achievements simply because companies also have incentives to exaggerate them. But the claim that all of science is now accelerating at a revolutionary pace remains much broader than the proof available this May. It is reasonable to invest while continuing to measure whether the outcomes justify the excitement.

One day we may discover that the largest contribution of AI was not a famous machine-designed molecule but a thousand uncelebrated improvements to how scientists organize evidence, avoid mistakes, and choose experiments. I would not be disappointed. In operations, the best innovations often remove the little failures that make life harder for everyone else. Reliable systems seldom look glamorous once they are doing their jobs. The public remembers opening night; the people running the business remember whether everything worked the following Thursday. That is the standard I would apply to scientific AI. Enjoy the astonishing demonstration, ask to see the full operating record, and do not mistake the highlight reel for the house rules.

There is one more principle that experienced operators understand: a successful emergency response should not be confused with a sound routine. If a research system occasionally discovers an extraordinary solution after enormous expert intervention, publish it and learn from it. But do not budget the next five years as though such miracles are ordinary daily production. Build budgets and expectations around repeatable performance, then let surprising achievements be the upside. This is particularly important when students, patients, and taxpayers may make decisions based on claims about how quickly tomorrow’s scientific problems will be solved. People can tolerate uncertainty when institutions state it plainly. They lose confidence when confidence is sold as a substitute for evidence.

Johannesburg, South Africa

Where I live, officials have occasionally announced a technical solution to a problem that continued to exist long after the announcement. A generating plant may have been designed, a financing plan drafted, and a contract celebrated, while households and businesses still experienced interruptions in supply. I mention this because scientific artificial intelligence inspires similar confusion between capability and delivery. A mathematical model can perform at an extraordinary level. A research agent can recommend promising experiments. A powerful computer can inspect data faster than a human team. None of those accomplishments means that the institutions required to produce reliable scientific knowledge have suddenly appeared wherever they are needed. Intelligence is a resource. It is not by itself the infrastructure that turns information into dependable outcomes.

I am impressed by the progress demonstrated before this May. The 2025 International Mathematical Olympiad result showed that an advanced AI system could produce difficult solutions that external graders accepted. Protein prediction methods have made it easier for researchers to examine biological structures and formulate testable ideas. And the May 19 Nature papers concerning Co-Scientist and Robin describe computational systems contributing to laboratory research, not merely writing fluent explanations. These are meaningful developments. Yet the question of whether science as a whole is accelerating should be answered with evidence from many institutions and contexts, not only from a few well-resourced teams. A technical achievement at the frontier is a demonstration of possibility; broad development requires distribution, reliability, and maintenance.

Science is sometimes discussed as though it consists of one brilliant person having an insight. In practice, a modern discovery may require high-quality data, a well-designed instrument, stable electrical power, laboratory technicians, ethical oversight, secure computing, and funding that survives a failed experiment. Many countries possess capable scientists while lacking consistent access to some of these essentials. An AI tool that proposes a thousand new research directions does little for a laboratory that cannot purchase the relevant reagents or maintain its equipment. It may even increase frustration by making the distance between what researchers know how to attempt and what they can afford to test more visible. Before celebrating a global revolution, we should ask who has the means to participate.

The global concentration of advanced computing is a strategic issue, not only a commercial one. Powerful research systems depend on chips, data centers, cloud access, specialized software, and sometimes proprietary datasets. Much of that infrastructure is controlled by relatively few organizations operating from wealthy countries. Their investment has produced genuine innovation, but access is not equal. Institutions outside those regions may face higher costs, weaker connectivity, currency constraints, or restrictions on data use. They could become customers for scientific intelligence developed elsewhere without developing comparable capacity to validate or improve it. That would reproduce an old pattern of dependency in a new technical form.

I do not argue that every country must build its own frontier model. That would be financially absurd. The sensible objective is scientific agency: the ability to select questions relevant to local conditions, inspect a tool’s assumptions, control sensitive data, and reproduce results. Shared computing infrastructure, collaborative research agreements, open methods, and regionally accessible laboratories can contribute to that objective. Governments that cannot keep public services functioning should be especially cautious about announcing vast AI ambitions without realistic operational plans. A successful national research policy requires trained people and working institutions, not merely a brand and a launch event.

The biological examples reveal another challenge of generalization. A model trained extensively on data from certain organisms, laboratory conditions, or patient populations may perform less reliably when asked to address poorly represented environments. Researchers must not assume that a result obtained in a highly controlled laboratory will apply in a clinic with different resources or in a population whose characteristics were scarcely included in the underlying data. The issue is not that scientific findings should be made local by decree. It is that claims of universality demand diverse evidence. When laboratory experiments are limited, researchers should say precisely what was tested and leave other conditions open for investigation.

The literature-analysis tools are particularly promising for researchers who work across borders. Scientific knowledge is dispersed among journals, repositories, languages, and disciplines. A capable assistant might identify an overlooked connection or allow a small research team to examine work it could never manually survey. This could reduce a genuine disadvantage. But a summary is useful only when its provenance is secure. Systems that invent references or treat uncertain claims as established can undermine research faster than they accelerate it. I have seen enough reports in finance and public administration to know that polished language often conceals weak evidence. A scientific assistant should make sources easier to inspect, not easier to avoid.

There is also the matter of research priorities. The January 2026 Nature study reported an apparent paradox: researchers using AI methods could achieve greater individual impact while the collective focus of science became narrower. This is concerning for places whose practical problems are not prominent in mainstream international datasets. Local infectious diseases, agricultural constraints, environmental conditions, and infrastructure challenges may receive less attention because the published record is sparse or the commercial return uncertain. If algorithms direct researchers toward subjects with abundant data, neglected questions may fall further behind. We should not confuse a higher number of papers with a more equitable distribution of knowledge.

Some advocates answer by saying that the leading systems will soon become capable of solving almost anything. That is a forecast, not a record of accomplishment. We should resist placing public budgets or research institutions at the mercy of expectations that may not materialize on schedule. Even if mathematical reasoning advances quickly, the material costs of experiments remain. A model cannot determine the behavior of every unknown physical system without reliable observations. It cannot manufacture a compound merely by writing its formula. Nor can it solve disputes over access to samples or consent from communities affected by research. A technology can reduce bottlenecks, but the remaining bottlenecks do not disappear out of courtesy.

The May research on Co-Scientist and Robin deserves careful attention precisely because it addresses that divide. Their published reports link computational proposals to actual laboratory activity. Some hypotheses received experimental support, showing that the systems contributed to more than verbal plausibility. But those experiments were conducted within specific projects by teams with relevant expertise and resources. We need independent replications and assessments of how often such systems improve the entire research process under different circumstances. The developers have demonstrated something worth building upon. They have not shown that any reasonably equipped laboratory can reproduce the same acceleration across unrelated scientific problems.

My background in financial risk makes me attentive to the direction of accountability. Who is responsible when a research AI system recommends a costly experiment based on unreliable data? Who can inspect the system if the result cannot be reproduced? What happens when a model version changes and the previous methodology becomes inaccessible? A university could find itself unable to defend a published finding because the underlying service was updated or withdrawn. The solution is not necessarily more regulation in every area. It is documented procedure, contractual rights, independent evaluation, and a clear allocation of responsibility. Scientific integrity should not depend on whether a vendor continues to support an old interface.

A less obvious risk concerns the quality of scientific institutions themselves. If senior officials measure success through AI announcements, they may reduce funding for the ordinary capabilities researchers need. Equipment maintenance is not glamorous. Laboratory safety attracts little public celebration until something goes wrong. Data management is expensive and tedious. Yet these functions make reliable results possible. Science can become distorted when a new technology receives capital expenditure while the supporting operation deteriorates. Anyone familiar with the gap between a promised infrastructure project and its eventual service level should understand the danger. One cannot repair institutional weakness with better publicity.

I do not want to sound gloomy. There is an opportunity to build research capacity in ways previously impractical. AI assistants might help investigators write better experimental protocols, detect inconsistencies, translate technical literature, and share methods across institutions. Young scientists could reach sophisticated tools without years of dependence on a narrow circle of specialists. Regional networks could pool experimental facilities while sharing computational services. Better evidence synthesis might help public health agencies identify patterns more quickly. These gains would be substantial even if no machine ever independently established a major new law of nature. Development often consists of making an important capability routine, not of staging an extraordinary demonstration.

The financing model for this development deserves scrutiny. Public institutions should not enter open-ended contracts with private providers based only on promotional claims. They should specify measurable services, ensure data protection, retain access to research records, and preserve the ability to change tools. Private firms should be encouraged to compete on performance and transparency. International collaboration could reduce duplication and support independent evaluation, provided institutions maintain appropriate control over sensitive information. The aim should be to spread the ability to conduct reliable research, not to distribute dependence on a small number of computational gatekeepers.

I would favor practical pilot studies in different types of research organizations. Give comparable teams similar questions; record baseline effort, available equipment, computing expenditure, and success rates; and measure results after independent verification. Include institutions in countries with weaker infrastructure, not only the best laboratories. Evaluate whether the tools actually reduce disparities or simply increase the output of groups already capable of performing complex experiments. Researchers should publish negative findings about the tools, because those failures can prevent others from making the same investment. A revolution that cannot tolerate an audit is not one I would finance.

Mathematical proof, literature synthesis, and biological design also ought to be evaluated separately. Proofs admit forms of formal verification that most biological hypotheses do not. Literature systems need comprehensive retrieval and accurate attribution. Biological design requires physical testing under appropriate conditions. A general declaration that “AI accelerates science” can conceal these distinctions. Some applications may already be mature enough for routine use; others remain exploratory. The responsible investor, scientist, or policymaker should ask which stage is being improved and what evidence establishes the improvement. There is no dishonor in saying that an important capability remains unproved in a particular setting.

My conclusion is that AI has accelerated specific scientific activities and made some forms of research newly practical. The record through May includes credible technical demonstrations and preliminary laboratory validations. What has not been established is a universal acceleration in reliable scientific discoveries, especially outside the institutions that developed and tested these systems. I suspect the long-term contribution will be considerable, but my suspicion is not evidence. It should guide experiments and investments rather than substitute for their results. We can be ambitious without allowing our ambitions to rewrite the measurement of what has happened.

Johannesburg has taught me to distinguish infrastructure that exists on a presentation slide from infrastructure that works when people need it. Scientific AI should face the same test. Show me the replicated result, the accessible method, the trained research team, and the capacity to follow through on the next experiment. Show me how an institution with ordinary constraints performs, not only what an elite laboratory achieved under ideal conditions. When those answers improve, we can speak confidently about accelerated discovery. Until then, celebrate the genuine breakthroughs while remembering that intelligence is not infrastructure—and a faster idea is not necessarily a finished piece of knowledge.

There is a final point about ownership of knowledge. When a publicly supported laboratory uses a proprietary system to make a discovery, the researchers should be able to explain enough of the process for the public to understand what its investment produced. That does not require publication of every confidential algorithm. It does require access to the relevant evidence, an account of limitations, and opportunities for independent examination. Without those arrangements, the public may finance work whose scientific foundations become inaccessible when a contract expires. That would be a peculiar form of progress: more apparent discoveries accompanied by less ability to check how they were reached. Real capacity means keeping the means of verification in working order.

Tyler, Texas

My business is insurance, which means I make a living reminding people that a favorable prediction does not eliminate uncertainty. A contractor can have a fine safety record and still suffer a serious accident. A clinic can purchase expensive equipment and discover that its greatest exposure lies in a mundane documentation problem. A church can believe it is covered for an unusual event until someone reads the exclusions. I approach scientific artificial intelligence through the same habit of mind. It is capable of impressive things, and the record available this May contains genuine advances. But the language surrounding those advances often slides from “this system performed well under test conditions” to “this system will transform research” without pricing the risk that separates one claim from the other.

We have reached a point where serious people can no longer dismiss AI as a toy that writes plausible paragraphs. At the 2025 International Mathematical Olympiad, Gemini Deep Think produced solutions achieving the gold-medal standard and receiving outside grading. In biology, computational predictions and protein-design tools have become useful components of research. Recent peer-reviewed work has described AI research agents proposing hypotheses and assisting with the analysis of experiments. Those are accomplishments. I have no interest in denying them to preserve a comfortable opinion about human superiority. My concern is that the claims are being sold into institutions whose leaders may not understand the limits well enough to manage their exposure.

Consider the risk of a confident answer that is wrong in a subtle way. A visibly nonsensical output is easy to reject. A plausible scientific argument with one invalid inference is more dangerous, especially when it is presented fluently and accompanied by references. A literature-analysis system could misstate a study, overlook a relevant contradiction, or produce a citation that does not support its conclusion. Scientists are trained to check such things, but workload and familiarity can make people careless. Organizations adopting these systems should define who verifies the sources, what level of uncertainty requires escalation, and how errors will be recorded. Without those controls, the apparent savings from automation may be purchased with a hidden accumulation of scientific risk.

The distinction between a mathematical proof and a biological hypothesis matters enormously. In mathematics, a valid proof can in principle be checked against explicit logical standards. A contest with known questions and independent graders creates a useful reference. Biology deals with systems whose behavior changes under conditions no model can describe completely. A protein that appears stable in a prediction may behave differently in an actual cell. A treatment hypothesis that succeeds in cultured cells may fail in patients. We should not let the excitement from one kind of demonstration imply equivalent certainty in another. An insurer would never treat an engineering simulation and a long operational safety record as interchangeable evidence. Scientific institutions should resist that confusion too.

The recently published Co-Scientist and Robin research is encouraging because it moves closer to actual experimental results. These systems are described as combining literature searches, hypothesis generation, experiment planning, and data analysis in ways that contributed to laboratory work. Some predictions received empirical support. That strengthens the case for their use. But there are practical questions about the selection of examples, the role of human experts, and the repeatability of success in other laboratories. A credible risk assessment should record not just the successful tests but how many proposals were discarded or failed. Otherwise, we may be looking at a selection of excellent claims without a clear picture of ordinary performance.

People sometimes tell me that a technology should be judged by what it enables, not by every error it makes. I agree up to a point. We do not close the roads because vehicles can crash. We set standards for drivers, equipment, maintenance, and liability. Scientific AI should receive similar treatment. Permit broad experimentation where the risks are limited, encourage innovation, and require stronger safeguards when an output determines expensive or consequential actions. The same literature assistant might be harmless when helping a student identify readings and dangerous when its unsupported conclusion shapes a clinical trial. Risk follows the intended use and the consequences of being wrong, not the glamour of the technology.

I would like institutions to develop a clear record of model performance. When a system proposes a research direction, log the source materials, version, date, and significant human edits. When scientists test the proposal, record success, failure, and uncertainty. When a result cannot be reproduced, preserve the case so others can learn from it. A record of adverse events is indispensable in many safety-sensitive occupations, and scientific automation needs its own version. The goal is not to punish every failed hypothesis. Research requires failure. The goal is to distinguish useful exploration from systematic defects that repeat because nobody is tracking them.

The financial risks extend beyond wasted laboratory supplies. A research institution could become dependent on a proprietary model that changes its terms or disappears. A scientist might unknowingly expose confidential patient data or intellectual property through a service that is poorly governed. A published finding could become difficult to reproduce because the underlying system was updated. A company might build a development strategy around AI-generated candidates whose performance collapses outside the original test. None of those risks proves that an institution should reject the tool. They prove that contracting, recordkeeping, security, and independent testing must accompany adoption. If a customer tells me a new process has zero risk, my suspicion rises rather than falls.

The academic incentives complicate matters. Research organizations are rewarded for novelty, publication, and visibility. Technology companies are rewarded for customers and valuations. Journalists need interesting stories. Those incentives can cooperate to produce exaggerated certainty even when nobody intends to deceive. A company publicizes a dramatic model performance; a university emphasizes its role; news coverage describes the discovery as imminent; investors and officials begin speaking as though the practical benefit is established. By the time cautious experts explain the missing experimental steps, the public may have moved on. Scientific communication needs the discipline to say exactly what was demonstrated, what remains a prediction, and when replication is still pending.

The January 2026 Nature study showing a possible narrowing of scientific topics under AI adoption raises a different kind of exposure. Institutions may become highly productive while making themselves more dependent on a limited set of assumptions or favored research directions. That resembles a business sector where every insurer has quietly taken the same concentrated risk. The individual policy looks reasonable, but the whole portfolio becomes vulnerable to one unexpected event. In science, the unexpected event may be an observation that contradicts a widely used model or a neglected line of inquiry proving much more important than the fashionable one. Diversity in research questions is a form of intellectual risk management.

I am not suggesting that government should attempt to schedule curiosity. That would fail spectacularly and probably cost a fortune. I am suggesting that public funders preserve avenues for independent research, replication, and work on problems unlikely to attract immediate private investment. If every laboratory buys the same proprietary model and pursues the same high-scoring topics, scientific independence will become thin. Healthy competition requires alternatives, including researchers capable of checking popular assumptions. Taxpayers should be able to see what public programs accomplished without demanding that every experiment guarantee success. The correct use of accountability is to expose performance and learning, not suppress uncertainty.

There is a workforce aspect often overlooked. A young scientist may grow accustomed to asking a system for the next promising hypothesis. That could make her far more productive, especially if the system locates research outside her experience. But if the training does not include independent checking, she may become unusually vulnerable to persuasive errors. In insurance we train people to read the actual policy rather than rely on the sales summary. Researchers need to read the actual paper, inspect the original data where appropriate, and understand the experimental assumptions. The machine should provide speed and range. It should not become the final authority merely because it produces an answer before a person can.

At the same time, the technology may make high-quality research more available to smaller institutions. A laboratory at a regional university could use reliable tools to identify relevant literature, compare methods, and design feasible experiments. That could be a genuine public good if the systems are affordable and if the laboratory has the resources to validate the results. I would be glad to see institutions in East Texas gain access to sophisticated research capabilities without relocating their best people to a handful of coastal campuses. Such diffusion could expand the range of practical questions studied. It would also create a larger market for transparent, dependable scientific tools rather than leaving advanced research concentrated among a few wealthy organizations.

The industry needs sensible standards, not impossible guarantees. Developers could disclose what tasks their systems were tested on, how many attempts were required, the kinds of errors observed, and the conditions under which performance changes. Research institutions could preserve records and identify where human judgment altered the result. Publishers could require clear descriptions of AI’s role in generating and validating scientific claims. Independent groups could evaluate reproducibility using questions chosen outside the developer’s promotional examples. These practices would reduce uncertainty without freezing innovation. The more reliable the product, the easier it should be to demonstrate its value under fair tests.

I find the word “autonomous” particularly prone to misunderstanding. A system may autonomously search literature, generate hypotheses, and draft conclusions while relying on humans to choose questions, approve experiments, execute laboratory procedures, and interpret the practical importance of results. Those human contributions are not disqualifications. They may be precisely why the process works. But if an institution advertises autonomous discovery, readers should understand which stages were genuinely automated and which depended on expertise. The issue is not who receives the applause. It is where responsibility resides when something fails, and how other investigators can repeat the work.

No industry benefits indefinitely from hiding defects. A tool that produces weak proposals in certain areas should be assessed frankly so users know when to avoid it. A system that performs exceptionally well on mathematical reasoning but poorly on specialized chemical questions should be marketed according to those strengths and limits. A service that generates accurate summaries for well-structured papers may still require substantial checks when the evidence is contradictory. Risk-based use permits institutions to capture real benefits without pretending that general intelligence means uniformly dependable performance. The best advertisement for an effective tool is a record of transparent performance under conditions that matter to the customer.

I believe researchers are already seeing acceleration in certain tasks and, in selected biological cases, in producing useful laboratory findings. The evidence does not yet justify saying that every field has entered a new era of uniformly rapid discovery. The full rate of validated progress remains hard to measure because institutions use different methods, projects face different constraints, and many outcomes take years to establish. That answer may sound insufficiently dramatic to people looking for a verdict. But a careful distinction between what is known, what is promising, and what is speculative is exactly what allows investors and scientists to make responsible decisions.

When somebody offers me a proposal in Tyler, I ask about the worst credible outcome as well as the expected benefit. I ask who bears the loss, what records will exist, and whether we can correct a recurring problem before it becomes catastrophic. Scientific AI deserves the same questions because its success could be enormously valuable. Good risk management is not the enemy of progress. It is how people become confident enough to adopt an innovation at scale. Give me transparent evidence, useful independent tests, and a chain of responsibility that survives the press conference. Then the promise of faster discovery begins to look less like a wager and more like an investment worth making.

There is one protection I would especially favor: an obligation to preserve the record of material errors discovered after publication. If a machine-assisted finding turns out to rely on a mistaken citation, flawed model assumption, or misinterpreted experiment, the correction should be easy for later researchers to find. A quiet revision buried in an updated webpage is not sufficient when others may already have incorporated the claim into their own work. Insurance depends on learning from claims and near misses; science should do the same. The story of an error can be as valuable as the story of a successful model, provided the institution is willing to disclose it. That is the sort of transparency that earns confidence over time rather than demanding it at the outset.

Billings, Montana

There is a difference between knowing what the forecast says and watching the weather arrive. I work as an insurance adjuster in country where hail can make a well-maintained roof look like it has been beaten with a sack of rocks. A storm model can be helpful. It can tell an insurer where to expect claims and help a family decide whether to bring the truck inside. But the model is not the damaged shingle, the broken window, or the field that has been flattened. When we inspect the aftermath, we often learn something the forecast did not capture. That is the attitude I bring to the question of artificial intelligence and scientific discovery. The models are becoming astonishingly capable. They still have to meet the world, and the world is not obliged to agree with them.

The evidence is strong enough that the argument can no longer be whether AI has learned to do anything scientifically interesting. It plainly has. The mathematical systems that reached the silver-medal level at the International Mathematical Olympiad in 2024 improved to a gold-medal-standard performance in 2025. The problems were externally graded, which is a useful safeguard against exaggerated claims. Protein structure prediction has given scientists an entirely different starting point for certain experiments. The newest research assistants described this May can propose biological hypotheses and help researchers decide what to test next. We should welcome all that. But we should be careful about confusing a map of possible answers with verified knowledge about how the world actually behaves.

Out here, the connection between prediction and observation is practical. A flood model may identify high-risk ground, but the actual path of water depends on local drainage, maintenance, conditions upstream, and events that do not fit neatly into a dataset. Land-use policy depends on measurements, maps, and assumptions, some of which are imperfect. Scientific AI faces a version of that challenge whenever the subject is a complicated natural system. A model can predict a protein’s structure, but how it behaves in a cell depends on an environment that the prediction may not fully represent. A system can suggest why a particular ecological pattern occurs, but field observations may uncover a variable the model never considered.

I would like the discussion to give more respect to the disciplines built around collecting difficult data. Scientists who measure river flows, soil conditions, animal populations, atmospheric behavior, and biological systems are often doing work that cannot be accelerated simply by writing better software. Instruments must be calibrated. Samples have to be collected at the right time and in the right place. Weather and access may limit fieldwork. A research project may depend on observations that take several seasons to become meaningful. Faster literature review is helpful, but nature does not increase its pace because a computer has a deadline. That limitation is neither anti-technology nor sentimental. It is the condition under which empirical science operates.

There are, of course, opportunities in those disciplines. AI might help researchers recognize signals in huge collections of satellite imagery, weather observations, acoustic recordings, and ecological measurements. It could help identify areas requiring closer investigation or combine data sources that people have struggled to reconcile. The benefit might be especially significant in places with enormous territory and limited field staff. A wildlife researcher may find more time for direct observation if software reliably sorts images. A hydrologist may detect patterns worth testing. An agricultural scientist may be able to prioritize samples more effectively. These are concrete ways technology could improve scientific work without pretending that every useful advance begins with an autonomous scientific genius.

The word “reliably” matters. An automated classification that mistakes cloud shadows for a physical change can generate false alarms. A model trained on common species may misidentify rare animals. A data system can combine records collected under incompatible conditions and create an artificial trend. The people using these tools should understand their error rates and know when to return to the original observation. In my work, a photograph can support a claim, but it often does not replace a careful inspection. Context matters. What looks like one type of damage may have a different cause when you examine the material and the surrounding conditions. Scientific models need the same humility about incomplete evidence.

The recent reports involving Co-Scientist and Robin are valuable because they move from plausible ideas toward experimental tests. The systems helped generate hypotheses and guide laboratory research, and the published work describes biological findings with some experimental support. That is a significant improvement over a model simply producing a persuasive essay. But a laboratory result is still one kind of observation under particular conditions. We need replication and explanations of failed attempts. We also need to know whether the work changed the next experiment for the better. A useful research assistant should be judged not only by the few promising ideas that attract publication but by its effect on the entire course of inquiry.

Mathematical reasoning occupies a somewhat different place. A logically valid proof is not subject to a thunderstorm changing its answer. That makes advances in automated proof generation easier to define, though not necessarily easy to achieve. The 2025 Olympiad milestone demonstrates that systems have developed impressive abilities in structured reasoning. Yet the most interesting scientific questions are not always well-posed exercises. Researchers often have to notice something surprising, determine whether the observation is real, invent a way to measure it, and only then formulate the question. Some of the hardest work happens before there is a problem statement anyone could feed into a machine.

The temptation to concentrate on subjects with good data worries me. A January 2026 Nature study suggested that AI-assisted researchers can achieve stronger individual publication outcomes while scientific investigation as a whole becomes more concentrated. That is a caution against equating an increase in measurable output with an increase in discovery. The problems that matter in Montana do not always come with enormous well-labeled datasets. Long-term changes in rangeland, water availability, wildlife movements, and rural health can be difficult to study precisely because the observations are sparse or expensive. A research system optimized for familiar material might direct money toward questions it already knows how to answer and away from questions that need patient collection of new evidence.

That would be a shame because one of the promises of technology is to make neglected fieldwork more productive. Imagine a small research team studying how water quality changes across a large watershed. AI could help organize historical reports, identify sampling gaps, compare records, and propose competing explanations. But the team would still need to collect samples and verify the claims on the ground. The useful innovation would be a better combination of computational assistance and field expertise, not a substitution of one for the other. If grants begin favoring flashy model demonstrations over the less visible labor of measurement, the technology may reduce the quality of the evidence it ultimately depends upon.

Scientific publishing faces a similar risk. The AI Scientist work published in March demonstrated that a computational system could produce research manuscripts and meet a workshop review threshold in one instance. That achievement raises a question about the supply of papers. If the number of manuscripts grows rapidly while the number of skilled people able to check them grows slowly, the literature could become harder rather than easier to trust. AI assistants might then be needed to filter an increasingly noisy collection of machine-produced articles. I would prefer that we invest in better standards for methods, source citations, negative results, and replication before celebrating unlimited production. Science ought to be an effort to reduce uncertainty, not a competition to create the greatest volume of confident prose.

This is not a plea for researchers to abandon computational methods and return to notebooks. I use sophisticated tools in my own work and appreciate what they accomplish. It is a plea to identify precisely where the technology adds value. Faster organization of evidence is valuable. Better detection of measurement anomalies is valuable. A biological candidate that survives laboratory testing is valuable. An independently checked mathematical argument is valuable. These successes are meaningful without being forced into a larger claim that machines have already transformed every branch of research. Some parts of science may accelerate dramatically; others may remain constrained by the time needed to observe the natural world.

There is an economic side that rural communities will recognize. Advanced AI systems require computing infrastructure, electricity, specialists, and often access to proprietary services. A small college or regional laboratory may be able to purchase a subscription but not afford the equipment necessary to test its suggestions. Research capacity is uneven, and the gap could grow if institutions spend heavily on computational output while reducing support for field instruments and technical staff. We should make investments that connect the pieces. Shared research equipment, well-maintained data repositories, open methods, and trained technicians may create more useful knowledge than another expensive demonstration that only a select group can reproduce.

I would also resist the idea that a lack of immediate practical application diminishes scientific worth. Some discoveries in mathematics or basic biology may take decades to influence daily life. We should not demand that every research grant justify itself like the purchase of a farm tractor. The ability to pursue questions without a guaranteed payoff is one of the strengths of scientific institutions. But when someone claims that AI makes discovery faster or cheaper, that is a measurable proposition. We can ask how much time was saved, which tasks were improved, whether results were independently confirmed, and what new questions became possible. Curiosity need not be treated as a business plan. A productivity claim should still be examined as one.

The ownership of data presents additional challenges for field science. Measurements may involve private land, local communities, tribal interests, sensitive environmental locations, or information gathered through public funding. Researchers should be thoughtful about placing such material into outside services and about the uses a vendor may make of it. If AI tools become central to field research, institutions need clear agreements covering confidentiality, access, version history, and the ability to reproduce findings. A system that loses the provenance of its inputs may turn good observations into questionable evidence. Data are not just raw material for a model; they often represent the work and trust of people who collected them.

We could use independent trials to measure the real benefits. Let researchers with comparable field questions use different combinations of AI tools and conventional methods. Record whether the software helps them detect important anomalies, design better sampling plans, and produce findings that other teams can verify. Measure false alarms and overlooked events, not just apparent successes. Include smaller institutions and difficult data, because that is where broad usefulness will be tested. Developers may discover that their systems perform exceptionally well in one type of environment and poorly in another. Such information would make the tools better and allow researchers to select them intelligently.

The strongest argument for optimism is that scientific progress has often followed improvements in instruments. Microscopes, telescopes, computers, and new measurement techniques opened parts of reality that people could not previously examine. AI may join that history by helping us connect observations, build useful models, and explore possibilities too numerous for manual study. Some of its achievements already warrant excitement. But a good instrument does not relieve the scientist of checking whether the result makes sense. The better the instrument becomes, the more important it is to understand what it measures and what it cannot. A powerful model without a disciplined connection to observation can be precisely and confidently wrong.

So my answer is qualified but positive. We are seeing meaningful acceleration in mathematical problem solving, information retrieval, and parts of biological design. The recent experimental work suggests AI can contribute to genuine findings, not just polished speculation. What remains difficult to prove is a broad acceleration in the rate at which science produces reliable knowledge across all fields. In some disciplines, better tools may already be changing everyday work. In others, the essential bottlenecks are physical observation, experimentation, and the slow accumulation of evidence. We should be clear about the difference rather than insisting on a single verdict for all of science.

When a hailstorm crosses eastern Montana, the computer’s forecast has already done its job. The claims adjuster still has to examine what happened. Scientific AI has reached the point where its forecasts of possible explanations and structures deserve serious attention. That is an accomplishment. The next step is to go outside, make the measurements, and see whether reality agrees. The future of discovery will belong to people who combine extraordinary computational tools with the old discipline of checking the world as it is. I see no conflict between those habits. The danger begins only when we decide the forecast is so impressive that the inspection can be skipped.