Prescott Valley, Arizona Correspondent — The present approach is reactive — rules written after harm appears or after deployment is already widespread.

For these two the speed and the scale make pure reaction too slow. An anticipatory framework would identify high-risk applications early, require transparency about capability and limitation, and set baseline safety standards before mass deployment.

Anticipation isn’t a freeze. It means drawing lines for the highest-stakes uses before the facts on the ground make regulation politically impossible, which is what happened with the platforms.

Europe proposed a risk-tiered framework in April. It’ll be imperfect and it’s the only serious legislative attempt anywhere, and American firms will comply with it regardless of what Congress does.

Dayton, Ohio Correspondent — Waiting for damage and then regulating has worked tolerably for slower technologies.

The manufacturing use I’d point at is quality inspection. We now have systems making pass-fail decisions on parts, and the system can’t explain why it failed one. Under the old regime a human inspector could be asked and could be wrong in a way you could trace.

That’s the practical form of the problem. Not a machine deciding something dramatic — a machine deciding something ordinary at volume, with no audit trail, in a process where a defect reaches a customer.

Testing requirements and accountability for severe outcomes in high-risk uses. Not a permit for a spreadsheet.

Sydney, Australia Correspondent — Reactive regulation lags too far behind in fast-moving domains and we’ve demonstrated it locally.

Our government deployed an automated debt-recovery system against welfare recipients that issued hundreds of thousands of incorrect notices, and it took years and a class action to stop. The technology was unremarkable. The failure was that nobody was required to test it before it was pointed at people who couldn’t contest it.

That’s the anticipatory case and it doesn’t require imagining anything exotic. Government uses of automated decision-making against individuals are the highest-risk category available right now and they’re the least regulated.

Sheffield, Jamaica Correspondent — The biotechnology half of this question is being discussed very differently depending on where you sit.

The gain-of-function argument is currently a domestic American controversy about a laboratory and an origin. From here it looks like a governance question about work that is conducted in perhaps a dozen countries under standards nobody harmonizes and no one inspects across borders.

There’s no international framework worth the name. There’s a convention with no verification mechanism, which is a treaty that relies on everyone’s good intentions.

Whatever the origin of this pandemic turns out to be, the absence of that framework is the finding, and it will still be absent when the argument about the origin is exhausted.

Long Island, New York Correspondent — The current tools are analogue and some of them still work. Liability, fraud statutes, medical product law, export control.

What’s missing is a duty to test in the high-risk applications and a reporting trigger when a system crosses a threshold you can define.

I’d write it narrowly. Credit, employment, medical decisions, and anything a government points at a citizen. Those four already have anti-discrimination law attached, and the honest observation is that existing law probably already prohibits a lot of what these systems do — nobody has litigated it because nobody can see inside the system to prove it.

So a disclosure requirement might do more work than a new regulatory regime, and it’s achievable.

Knoxville, Tennessee Correspondent — I want what I want on a bridge. If it can kill somebody, inspect it. If it’s a toy, don’t.

A system that dispatches power or triages a patient is a bridge. A recommendation engine for footwear is a toy.

The public-works version is already here. We’re being sold predictive maintenance systems that tell us which culvert fails next, and I have no way to evaluate the claim. The vendor won’t show me the model and my county has nobody who could read it.

That’s a procurement problem before it’s a regulatory one, and every small government in America is about to buy something it can’t assess.

Jacksonville, Florida Correspondent — Export-control the genuinely dual-use, test the medical claim, and don’t regulate a spreadsheet like a pathogen.

The port has automation coming and the argument here will be about labor rather than safety. That’s a different fight and it’ll be louder than this one.

On the biotech side, I’d say the shipping industry moves an enormous quantity of laboratory material and nobody outside the industry knows what the controls are. There’s a real security question there that isn’t being asked because it isn’t interesting.

Las Vegas, Nevada Correspondent — We already use models to score a player and we still keep a human who can explain a decision.

That’s the regulated version and it works. The Gaming Control Board requires that a system be auditable, that its logic be filed, and that a patron be able to contest an outcome to a person.

None of that stopped anybody from building the systems. It made them explainable, and explainability turns out to be an engineering choice rather than an impossibility.

If an industry as commercially aggressive as this one can operate under that, the argument that it would strangle innovation elsewhere is weaker than it sounds.

Orange County, California Correspondent — The investment view is that regulatory uncertainty is worse than regulation.

Capital can price a rule. It cannot price the possibility of a rule arriving in three years and being retroactive, which is the current condition, and the result is that money goes to applications where the regulatory question is settled rather than to the ones with the most social value.

An anticipatory framework, even an imperfect one, would unlock investment rather than suppress it. That argument is available to the industry and it hasn’t been making it, because the industry’s near-term interest is in no rule at all.

Novi, Michigan Correspondent — Automotive has the most advanced version of this argument and it’s instructive.

We’re deploying driver-assistance systems into a regulatory framework built for mechanical failure. The federal safety standards test whether a component breaks. They don’t test whether software makes a bad decision in an unusual situation, and there’s no protocol for that because nobody knows how to write one.

Meanwhile a fatality involving these systems generates an investigation, a headline, and no rule. Twelve investigations open and no standard.

That’s the inadequacy in a concrete form. It isn’t that regulation is too tight. It’s that nobody knows what to measure and nobody is being funded to find out.

Owatonna, Minnesota Correspondent — I’d offer the historical caution, since somebody should.

Anticipatory regulation has a poor record because the anticipation is usually wrong. We wrote rules for the internet in the nineties that assumed a set of harms that didn’t materialize while missing the ones that did.

That’s not an argument for doing nothing. It’s an argument for regulating the process rather than the technology — requiring testing, disclosure, and accountability, which are stable requirements, rather than specifying what the technology may do, which will be obsolete.

Tyler, Texas Correspondent — Insurance is where this becomes real and nobody is ready.

I’m being asked to write coverage for firms deploying these systems and there’s no actuarial basis. No loss history, no standard of care, and no clear answer to who’s liable when a model makes a decision — the vendor, the deployer, or the person who trusted the output.

Until a court answers that, the market will price it badly or decline it entirely, and declining it is what’s mostly happening.

A framework that established a standard of care would do more for adoption than any subsidy, because it would let me write the policy.