New York City doesn't regulate AI safety today. A bill package unveiled by City Council Speaker Julie Menin ahead of an October 5 hearing is designed to change that — and on the day of the hearing itself, it produced something unusual: on-the-record executives from OpenAI, Anthropic, Google and Meta, under questioning, repeatedly declining to put a number on how likely their own technology is to cause catastrophic harm.
Quick summary: Ahead of an October 5, 2026 hearing convening all 51 City Council members, Speaker Julie Menin unveiled a bill package requiring third-party validation and a human-controlled "kill switch" before AI systems can be marketed, sold, or deployed in New York City (with a $25,000 penalty per violation for both the business and the validator), plus a first-in-the-nation program paying whistleblowers a share of any fines recovered from AI companies. At the hearing, policy executives from OpenAI, Anthropic, Meta and Google each declined to quantify catastrophic-risk probabilities when asked directly, and former Anthropic researcher Jacob Coxon testified that the industry is being "extremely reckless" and that it's "more likely than not that humanity loses control to these AIs."
The bill package: kill switches, validators, and a $25,000 penalty
The centerpiece proposal would require independent third-party validation before an AI system can be marketed, sold, or deployed within city limits, and would mandate that any such system include a human-controlled kill switch. Both the business deploying the system and the third-party validator would be liable for a $25,000 penalty for each instance an AI system is marketed, sold, or deployed without valid third-party validation, or where that validation was falsified. A second measure, billed by the Council as a first-in-the-nation approach, would give whistleblowers a cut of any fines or penalties the city recovers from AI companies after they report violations — a direct financial incentive for insiders to come forward rather than resign quietly (Sources: amNY, aiweekly.co).
What the executives said under oath — and what they wouldn't say
Menin had asked the chief executives of Anthropic, OpenAI, Google and Meta to appear personally, and said the Council would consider using its subpoena power if they didn't; in the end, each company sent a senior policy or technical representative rather than its CEO. The sharpest moment came when Council members asked each company to put a probability on their own technology causing catastrophic harm. OpenAI's Morgan Dwyer, head of policy development and operations, declined to name a figure, arguing, "I don't think it matters whether it's 1% or 10% or a 20% chance that something catastrophic will go wrong" — a notable non-answer given that OpenAI's own researchers have previously discussed risk estimates internally. Anthropic's Logan Graham and Meta's Shane Cahill both declined to give a percentage. Google's Alice Friend gave the most legally precise answer in the room, arguing no rigorous scientific method exists yet to produce such a number, while also noting on liability that "if something is illegal without AI, it is still illegal with AI" (Source: amNY).
| Company | Who testified | Title | Stance on risk quantification |
|---|---|---|---|
| OpenAI | Morgan Dwyer | Head of Policy Development and Operations | Declined a number; said the specific percentage "doesn't matter" |
| Anthropic | Logan Graham | Head of Frontier Red Team | Declined to give a percentage |
| Meta | Shane Cahill | Director for Privacy and AI Legislation | Declined to give a percentage |
| Alice Friend | Policy representative | Said no rigorous scientific method exists yet to produce one |
The whistleblower in the room
Menin had specifically invited AI whistleblowers to testify, and the most pointed testimony came from Jacob Coxon, the former Anthropic researcher whose September 9 resignation — and warning that the industry was "gambling with our lives" — went viral earlier this fall (a story LumiChats covered in depth at the time). A month later, in front of the Council, Coxon doubled down rather than walked it back: "On the current path, I think it is more likely than not that humanity loses control to these AIs, and it could end in human extinction," he told the packed hearing room. He described the industry's pace as "extremely reckless given the stakes," arguing that "the companies run on a startup mindset: move fast, break things, fix them later... It does not work for building the most powerful technology ever built," and added bluntly that "we do not know how to control any AI system yet" (Source: Yahoo/Politics coverage of the hearing).
Will any of this actually become law?
An important thing to keep in mind: a hearing is not a passed law, and a bill package unveiled by the Council Speaker is not the same as something the full Council has voted on. New York City would be an unusual jurisdiction to lead on AI safety regulation specifically — most binding AI rules in the US so far have come from states (California's employment-focused SB 947, for instance) or from federal agencies, not city councils. Whether Menin's kill-switch and whistleblower-reward proposals survive committee, industry lobbying, and a full Council vote is genuinely an open question; this hearing is best read as the opening move in what will likely be a long process, not a settled outcome.
01What would NYC's proposed AI 'kill switch' law actually require?
It would require independent third-party validation before an AI system can be marketed, sold, or deployed in New York City, and mandate that the system include a human-controlled kill switch. Both the deploying business and the validator face a $25,000 penalty per violation if validation is missing or falsified.
02What is the whistleblower reward proposal?
A separate measure in the same bill package that would give whistleblowers a share of any fines or penalties the city recovers from AI companies after they report violations — billed by the Council as a first-in-the-nation incentive for insiders to come forward.
03Did the AI company executives admit their technology is dangerous?
Not directly. Representatives from OpenAI, Anthropic and Meta each declined to quantify the probability of their technology causing catastrophic harm when asked, and Google's representative said no rigorous method exists yet to produce such a number. None made a blanket commitment that failing a safety test would automatically block a release.
04Who is Jacob Coxon and why did he testify?
Coxon is a former Anthropic researcher who resigned in September 2026 warning that unchecked AI development carries real extinction risk. The Council specifically invited AI whistleblowers to testify at its October 5, 2026 hearing, and Coxon reiterated and sharpened his warning in person, saying it's "more likely than not that humanity loses control to these AIs."
05Has this bill package actually passed?
No — as of this hearing, it's a Council Speaker's proposal that has been unveiled and discussed, not a law the full Council has voted on. Whether and when it passes remains an open question.
Regulatory debates like this one move slowly, but your choice of which AI model to actually use doesn't have to wait on them. LumiChats gives you Claude, GPT, Gemini and 40+ other models in one subscription, so you can form your own view of how each company's product actually behaves rather than relying only on what its policy team says in a hearing room.
