New York's Bill to Slow AI Down
This post is part of my Medium blog.
There's a bill sitting in the New York Senate Rules Committee right now that would give one state office something no American regulator has ever had: the power to tell the companies training frontier AI models to use less computing power.
It's S.10731, introduced October 28 by State Senator Andrew Gounardes and Senator Patricia Fahy. It's the follow-up to New York's RAISE Act, the frontier AI law Governor Hochul signed in March that takes effect January 1. The first law required disclosure and incident reporting. This one goes further. This one treats a training run like a bank's balance sheet: something the regulator can mark down.
I want to walk through what it actually does, because the summary floating around underplays the strange part.
The headline provision is a new section called "Computing power limitations." The bill creates a permanent office inside the Department of Financial Services. It's called DIGIT, which is a genuinely funny acronym for the agency holding the most blunt regulatory instrument yet proposed in this country. That office may issue an order limiting the amount of computing power a large frontier developer can use to train new models, or to run automated AI research. The order applies to every large developer at once, though the office keeps discretion to tailor limits per company. There has to be a reason: catastrophic risk, a critical safety incident, or a covered emergency. And a licensed examiner verifies compliance afterward. But no court order comes first. The office decides, and the throttle turns.
The rest of the bill is the scaffolding that makes the throttle plausible. Independent frontier examiners, third parties who get access to unreleased models, internal documents, and employee interviews, are phased in: regulations licensing them by January 2028, and a hard requirement that every large developer use only licensed examiners starting January 2029. Examiners who believe an emergency is underway must report it to the developer, the office, and the attorney general within seventy-two hours.
The reporting obligations also tighten. Large developers already have to file quarterly assessments of catastrophic risk from the internal use of their own models. The bill adds a 72-hour deadline whenever their own assessment of risk jumps materially, and requires a specific number for something labs have mostly kept vague: what percentage of their AI research and development is currently automated. Annual audits of the frontier AI framework, the safety document each lab publishes, become explicit, with a certified senior officer signing under penalty of perjury.
And then there are the penalties, which is where the bill stops being polite: up to one million dollars per day for a first violation, up to three million dollars per day after that. Not per violation. Per day. That's the penalty structure you use when you expect the defendant to be a trillion-dollar company that treats eight-figure fines as a line item.
Read the bill yourself, it's short enough: Senate Bill S10731, with full text and the sponsor's memo on the New York Senate site. Gounardes sponsored the original RAISE Act too, so this is a deliberate second act, not a stray filing.
Here's where I land on it.
The odds this becomes law, as written, this session: low. It's in Rules. It's one chamber of one state. There's an active federal preemption fight: the Justice Department has a litigation task force aimed at exactly these state laws, and a White House that ran on challenging them. A state claiming authority to throttle frontier training runs will end up in court before it ends up on the books.
But the interesting question isn't whether it passes. It's that someone drafted it seriously. The compute-limit section isn't a press-release gesture. It has compliance mechanics, verification through examiners, an emergency trigger, and per-day penalties sized for the industry it targets. This is what it looks like when a regulator stops trying to read a lab's homework and starts reaching for the actual lever.
In Redundant, the confrontation between the people who build frontier systems and the people who are supposed to oversee them is mostly improvised on both sides. The oversight arrives late, frantic, and after the damage. The scenario that kept bothering me while I wrote it was the opposite one: oversight arriving early, with real authority, and what that world looks like for the people inside the labs. A bill like this is the paperwork version of that question.
The committee hasn't voted. The bill has two sponsors and a long way to go. I'll be watching whether it gets a hearing, because if the throttle provision survives even the committee stage, every frontier lab's Albany lobbying budget is about to get interesting.