What AI companies said under oath at the NYC Council hearing on AI risk

What did AI companies say under oath at the NYC Council hearing on AI risk? On October 5, 2026, representatives of Anthropic, OpenAI, Google and Meta answered questions from the full 51-member Council in a Committee of the Whole. The sworn answers put a handful of facts on the record. They did not establish who pays when an AI system causes harm, and they did not produce a company estimate of catastrophe odds. The odds came from the three former employees who testified first.

What happened in the room

The hearing was the first time these firms testified under oath before a legislature on the risks of frontier AI, according to the Council's own announcement. Speaker Julie Menin framed it as a duty: the firms, she said, owe it to the public to come before the Council, answer questions and give input on proposed legislation under oath.

Attendance was not voluntary in spirit. Per the Council's account, Meta confirmed first, while Google and Anthropic missed the September 25 response deadline, and OpenAI and Google agreed on the Sunday before. Anthropic confirmed late that night, hours before a subpoena was due to be served.

Elon Musk's SpaceXAI is the exception. It ignored the Council's inquiry, was subpoenaed, and still did not appear. CBS New York reported that the company cited a wish to work with the city separately. Menin's response, as reported by The City Reporter: "The only response to the subpoena should be that they should be here today to testify. They are opting not to do that, so we are pursuing legal action."

The witnesses who gave numbers

The hearing opened with former insiders. Jacob Coxon, a former Anthropic researcher, said that on the current path it is more likely than not that humanity loses control to these AIs, and that this could end in human extinction. Alex Turner, formerly of Google DeepMind, put the chance of a complete AI takeover at roughly one in three, according to wire coverage carried by the Hanford Sentinel. Daniel Kokotajlo, formerly of OpenAI, called autonomous AI self-improvement a recipe for disaster.

These are personal probability estimates, not measurements. They are useful as testimony because they were given under oath by people who worked inside the labs, but nobody at the hearing showed a method that produces a figure like "more likely than not." Our explainer on P(doom) numbers covers why such figures swing so widely and why they should be read as judgments, not forecasts.

What the companies did and did not say

The company witnesses were Logan Graham, who leads Anthropic's Frontier Red Team, Morgan Dwyer of OpenAI, Alice Friend of Google and Shane Cahill of Meta. Three of the exchanges matter.

• Probability. According to the wire account, Dwyer said the likelihood of the worst case does not matter, only that no chance of it is remotely acceptable. That is a refusal to give a number dressed as a safety stance. CBS reported that Cahill declined to quantify risk, saying he lacked the frameworks to do so.
• Liability. Menin asked directly whether companies would bear legal responsibility if their models went rogue and caused harm. CBS reported the answers were vague: Dwyer said OpenAI takes its safety responsibilities very seriously without giving a yes or no. No source we opened reports a company accepting liability.
• Precaution. Graham pointed to Anthropic holding back its most powerful model this year for safety evaluation, per the wire account. That is a company claim about its own conduct. Nothing in the coverage we reviewed shows it was independently checked.

Councilmember Shekar Krishnan, per The City Reporter, criticized the answers for offering general statements rather than specific risks. Meta's written line was that earning public confidence requires rigorous safety work, transparency and a willingness to engage with policymakers, as ABC11 relayed.

So the record is thin where it matters most. Three former employees supplied alarming probabilities. Four current employees supplied process language. The gap between those two is the finding.

The bills and the jurisdiction problem

The Council is weighing 10 bills. Reporting describes a human shutdown requirement verified by a third party, with $25,000 fines, a whistleblower bounty program, a private right of action for New Yorkers harmed by AI agents, 24-hour incident reporting for city contractors, third-party validation of AI systems, and rules on chatbot data consent and unauthorized AI depictions of public officials.

Can a city do this? The City Reporter says plainly that the Council lacks power to regulate the industry as a whole. The realistic reach is narrower: procurement rules for city contractors, consumer protections for people in New York, and a private right of action that lets courts do the enforcing. Each of those invites a legal challenge on jurisdiction, and the companies reportedly questioned both technical feasibility and the Council's authority, according to one analysis post. Mayor Zohran Mamdani told the hearing that national action is needed.

Federal context cuts both ways. The Trump administration has favored speed over regulation, which is part of why Menin says the Council is stepping in. It is also why any city rule could be contested. Congress is moving on liability from other directions, such as the Hawley and Murphy bill on criminal liability for agents, and a federal regime would likely override whatever New York writes.

What to watch

Three things will tell us whether this hearing was theater or a start. First, whether the Council moves any of the 10 bills to a vote, and which survive as drafted. Second, whether Menin's legal action against SpaceXAI is filed and whether a court enforces a municipal subpoena against a company of that size. Third, whether the four firms submit written follow-ups with actual numbers on liability and risk, since the sworn hearing did not produce them.

Our take

The hearing proved a city can make the largest AI labs show up, and that is not nothing. It did not prove those labs will accept legal responsibility for harm, and the only catastrophe probabilities on the record came from ex-employees, not the companies. The private right of action is the bill with teeth, because it moves the liability question from a hearing room to a courtroom. Expect the fight to be about whether New York has the authority to pass it at all.