Prescott Valley, AZ Correspondent — Regulate high-risk applications rather than attempting to license every model. Systems that touch safety, credit, employment, or an election.
Transparency about training data and limitations, plus clear liability for severe harms, creates accountability without freezing development.
The hardest problem is preventing the rules from becoming a barrier only the largest firms can clear, and the largest firms know it. Several of them spent this year asking to be regulated, which is not a sentence you hear from an industry that expects the regulation to hurt.
Clermont, FL Correspondent — The technology moves faster than a regulatory cycle, so identify the uses carrying the highest potential harm and put guardrails there.
Broad preemptive licensing becomes a moat.
What I’d note from the small-business side is that my clients are already using these tools daily and nobody is regulating that at all. The argument in Washington is about frontier models. The actual deployment is a twelve-person firm running client documents through a service, with no policy, no audit, and no idea where the data goes.
Novi, MI Correspondent — Target high-risk uses with transparency and liability. Avoid licensing that only the largest can navigate.
The executive order in October runs to more than a hundred pages and rests on a wartime production statute, which is a creative hook for a technology policy. Whatever you think of the substance, it’s the sort of authority that gets litigated and doesn’t survive a change of administration.
Durable rules come from a legislature. We don’t have one that will write this.
Sheffield, Jamaica Correspondent — Regulate the highest-risk applications and keep barriers low enough that new firms can enter, because heavy rules favour incumbents.
I’d raise the international dimension, since three jurisdictions are now writing rules that everyone else will inherit. Europe agreed the outline of its act this month, you have an executive order, Britain held a summit.
None of those processes included any country in my region, and all of them will apply to us in practice, because the models are built elsewhere and licensed here. That’s ordinary in technology and it’s worth someone saying aloud that the governance of this is being settled by about four governments.
Jacksonville, FL Correspondent — Regulate the use that can wreck a person or a grid. Credit, medicine, safety systems. Not the notebook and not the jingle.
Stifling happens when a bureau pre-clears every model. Concentration happens when only four firms can afford the bureau.
Write duties a second-tier lab can meet, and watch the incumbents try to wrap the duty around themselves.
Las Vegas, NV Correspondent — We use models to price a room and still keep a human who can explain a refusal. Same split.
Power concentrates if the statute is a novel three legal departments can read. Write short and enforce on deployment rather than on curiosity.
Curiosity is how a smaller shop eats somebody’s lunch, and that’s the anti-concentration policy that actually works.
Cheyenne, WY Correspondent — If it can kill or deny a loan, inspect it. If it writes a poem, don’t.
Short list. Review the list. Don’t build a church, because churches pick favorites and favorites are the concentration.
Knoxville, TN Correspondent — Bridges get inspected and toys don’t.
Write the bridge rule so a university laboratory can obey it without hiring counsel. That avoids both a free-for-all in the hospital and a castle on the coast, and the castle is the risk nobody in the hearing wants to name because the castle is in the room.
Long Island, NY Correspondent — The liability question is where this gets decided and almost nobody is working on it.
When a model produces a harmful output, who is responsible — the developer, the company that fine-tuned it, or the firm that deployed it without testing? Existing product liability doctrine was built for physical goods with identifiable defects, and it maps badly onto a system whose behavior is statistical.
Until a court answers that, the market prices the uncertainty by declining to insure it, which is what my colleagues in that industry report.
Settle the liability rule and you’ve done more for both safety and adoption than any licensing regime.
Orange County, CA Correspondent — The concentration is already substantial and it isn’t in the models. It’s in the compute.
Training a frontier system requires an amount of specialized hardware that perhaps five organizations can assemble, and the supplier of that hardware has effective control over who competes. That’s a bottleneck with a name and it’s more consequential than anything in the current rulemaking.
Any policy serious about concentration has to address access to compute. Nothing proposed does.
Dayton, OH Correspondent — The manufacturing deployment is quieter than the debate and it’s already here. Vision systems making pass-fail decisions, scheduling systems allocating work.
What I want is a testing requirement and a record. Not because the technology is sinister, but because we require that of every other process that touches a part, and nobody can currently tell me why this one is exempt.
Sydney, Australia Correspondent — We’re a rule-taker here and I’d say so plainly.
Our government is consulting on an approach and will land somewhere between Brussels and Washington, because those are the options available to a market of twenty-six million.
What I’d urge on the Americans is interoperability. If your rules and Europe’s differ in structure rather than in stringency, every firm outside both markets builds to the stricter one and complies with neither well. That’s what happened with privacy and it wasn’t anybody’s intention.
