NYC Council AI Hearing Exposes the Gaps in Trump’s Voluntary Safety Accord

Last updated: 7 October 2026. Figures and quotes below are sourced to primary coverage of the hearing and the underlying accord, linked inline throughout.

The 60-second version

• On 5 October 2026, NYC Council held its first “Committee of the Whole” hearing since 2022 — all 51 members — to question OpenAI, Anthropic, Google and Meta on AI safety.
• Asked for the odds of a catastrophic AI failure, none of the four would give a number. Asked who carries insurance against that risk, none raised a hand.
• The hearing came three weeks after the same companies (plus SpaceX and Nvidia) signed a voluntary “Super Intelligence” accord with President Trump carrying no penalties, no public audit requirement and no duty to report incidents.
• Elon Musk’s SpaceXAI skipped the hearing despite a Council subpoena; Speaker Julie Menin says she’ll seek a judge’s order to enforce it.
• Three former lab researchers testified that frontier labs can’t reliably detect or control misaligned systems; one put the odds of an eventual AI takeover at “roughly one in three.”

Key dates and numbers

What actually happened at the hearing

New York City Council convened as a Committee of the Whole on 5 October — a rarely used format pulling in all 51 members — to question OpenAI, Anthropic, Google and Meta under oath. Speaker Julie Menin chaired alongside Council Member Carmen De La Rosa; roughly 40 members took part over the day, according to amNY’s reporting on the hearing.

The companies sent policy staff, not chief executives: OpenAI’s Morgan Dwyer (head of policy development and operations), Anthropic’s Logan Graham (head of its Frontier Red Team), and — remotely — Google’s Alice Friend and Meta’s Shane Cahill. Menin’s central question was blunt: what is the worst-case probability your technology causes catastrophic harm? Dwyer’s answer became the hearing’s defining exchange: “I don’t know. I also don’t think it matters whether it’s 1% or 10% or a 20% chance that something catastrophic will go wrong. None of these levels is remotely acceptable.” Menin was unimpressed, calling it “flippant at best” and comparing it to a drug company that couldn’t say how often its product kills patients.

Graham discussed Anthropic’s cybersecurity and loss-of-control testing but, like Dwyer, declined to give a number. Cahill said he didn’t want to guess and would follow up in writing. Friend’s response was the most pointed: forecasting catastrophic risk, she said, “is not a perfect science at this stage,” per TheNextWeb’s account, and on liability she offered the clearest line of the day — existing law already covers AI harms, because “if it’s illegal without AI, it’s still illegal with AI.” Asked separately to raise a hand if their company carried insurance against catastrophic AI risk, none of the four did — “not a single hand went up,” per Forkast’s account.

Elon Musk’s AI unit, SpaceXAI, did not appear at all, a week after being subpoenaed. “SpaceXAI, however, is not here at all in direct violation of the subpoena that we issued last week, and we are pursuing that subpoena in court,” Menin said, per amNY’s own account of the hearing. Google, OpenAI and Anthropic reportedly agreed to attend only once the Council raised the prospect of subpoenaing them too; Meta had agreed earlier.

Three weeks earlier: a different kind of commitment

The hearing’s sharpest edge comes from what preceded it. On 29 September, President Trump signed a voluntary document called the “White House Accord on Super Intelligence: Joint Commitment on Frontier Responsibilities” with the chief executives of Google, Anthropic, Meta, SpaceX and Nvidia, plus OpenAI’s president (not its CEO). As MediaNama’s analysis of the accord text lays out, the deal asks signatories to run internal monitoring during training and deployment, staff an internal oversight team, bring in an outside auditor, and give a board committee visibility into the results. Trump called it “morally binding” and “almost like a constitution.” It binds no one: there is no penalty for skipping a commitment, no requirement to publish audit findings or even name the auditor, and — critically, given what followed in New York — no duty to report safety incidents to any government body. The companies had already told the UN Security Council they preferred this kind of voluntary, industry-led framework to binding international rules — a preference now also under scrutiny at home, where the FTC has opened its own safety probe into OpenAI and Anthropic over whether self-regulation is enough.

That gap between a self-written pledge and an actual cross-examination is what the hearing exposed. A company can comply fully with the accord’s four pillars and still be unable to tell a legislature its own worst-case failure rate, or whether it carries the financial backstop for that failure. The accord measures whether a process exists; the hearing asked whether anyone can act on what the process finds. The honest answer, from four well-resourced companies, was: not with any precision.

What most coverage is getting wrong

Several write-ups framed this as Big Tech “stonewalling,” which undersells what’s interesting here. Dwyer’s “it doesn’t matter whether it’s 1% or 20%” line is less evasive than it sounds: if a given probability of catastrophe is unacceptable, precision about the digit is secondary. The sharper question, which only Menin’s insurance query really reached, is what backs that position. If no probability is acceptable and nobody is insured against the outcome, the practical safety net is whatever internal review the September accord already promised — the same process that, by its own design, need not report findings to anyone.

Coverage has also undersold the domestic regulatory fight underneath the hearing. New York’s RAISE Act — the state’s frontier AI safety law — became a flashpoint when Dwyer told the Council that OpenAI “worked with other companies on the New York RAISE Act, and we do support” it. State Senator Andrew Gounardes, the bill’s Senate sponsor, disputed that outright: “I was there — they actively pushed to water down the bill’s safety requirements,” he said. The Assembly sponsor, Alex Bores, went further, saying he could “definitively say that OpenAI opposed it, from the moment it was introduced to the moment it was signed,” and that his “common man understanding of lying under oath is that it’s perjury.” Bores has separately said the version that passed dropped mandatory third-party audits and a ban on releasing models that fail a company’s own safety tests — both present in earlier drafts — after opposition escalated from regional to national industry lobbyists, per his own account of the legislative fight. That gap between hearing-room statements and statehouse lobbying may be the more durable story here.

The whistleblowers were the sharpest witnesses in the room

Before the companies testified, the Council heard from three people who used to work inside frontier labs. Jacob Coxon, who left Anthropic after previously working at OpenAI, said: “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 described a “startup mindset” where AI now writes most lab code and gets reviewed less carefully than before. Daniel Kokotajlo, a former OpenAI researcher who now runs the AI Futures Project, said labs’ ability to detect misalignment is poor and getting worse, warning current safety fixes risk being “duct tape that will fall off later.” Alex Turner, formerly of Google DeepMind, put a number on it where the companies wouldn’t: “roughly one in three” odds of an eventual AI takeover. He also said he’d tried and failed to slow a Google Pentagon contract by sending Google DeepMind chief executive Demis Hassabis 25 pages of proposed oversight terms before Google signed anyway.

These are one researcher’s personal judgement each, not measured probabilities, and worth flagging as such. But they stand in sharp contrast to the hedged, process-focused answers from sitting executives. In-house dissent hasn’t fared much better: earlier the same week, OpenAI fired three safety researchers who had publicly criticised its pace on AI risk, which is part of why Menin closed the hearing saying the Council “still has many questions that we feel we did not get answers to today.”

Practical takeaways for builders and publishers

• Incident disclosure timelines are a compliance question now, not just a PR one. OpenAI’s own 84-day gap between discovering and disclosing its Australian Medicare portal breach is exactly the delay NYC’s proposed 24-hour reporting rule for city contractors targets. A vendor’s voluntary accord commitments tell you nothing about how fast you’ll be told if something breaks — put disclosure timelines in the contract, not the press release.
• “We follow a voluntary safety framework” is not evidence of insurability. None of four major labs present claimed catastrophic-risk coverage. For vendor risk assessments, treat the absence of insurance as a signal the risk is considered uninsurable or unpriced, not low.
• State and city rules will diverge, not converge. Between New York’s RAISE Act, NYC’s pending validation and kill-switch bills, and California’s AI auditor registry, expect a genuinely fragmented compliance map by 2027, not one federal standard.
• If the vendor can’t give you a risk number, your own testing has to. Frontier labs’ own policy staff couldn’t give a city council a risk figure for general-purpose models; you won’t get one for your specific use case either, so monitoring and testing on your end carries more of the safety load than marketing implies.

What we still don’t know

• Whether the SpaceXAI subpoena will actually be enforced in court, and on what timeline — Menin intends to pursue it, but no ruling has been reported.
• How New York City’s proposed 24-hour incident-reporting deadline will square with the RAISE Act’s existing 72-hour state deadline — Google’s Alice Friend flagged the two as being in conflict, but no reconciliation has been proposed.
• Whether the ten pending safety bills — including the validation and $25,000-per-violation penalty bill — will pass, given companies’ record on the RAISE Act.
• Whether other cities follow New York’s example of a direct, oath-sworn hearing rather than written submissions.
• How the accord’s “independent external auditor” provision works in practice, since it names neither the auditor nor a requirement to publish results.

FAQ

Did any laws change as a result of this hearing?
Not yet. It was an oversight exercise — testimony and questioning, not a vote. Ten related bills remain pending before the Council, and Menin said written follow-up questions would go to the companies.

Is the White House “Super Intelligence” accord legally enforceable?
No. It is voluntary, with no statutory basis, no penalties for non-compliance, and no requirement to publish audit results or incidents, per the accord text reviewed by MediaNama.

Why does it matter that the companies wouldn’t give a risk percentage?
Because the follow-up — who insures against that risk — got a clearer answer: nobody present did. No stated probability and no stated financial backstop is the detail most worth remembering.

What happened with SpaceXAI?
Elon Musk’s AI unit was subpoenaed on 28 September and did not appear on 5 October. Menin said she would seek court enforcement of the subpoena.

Sources

• amNY: “AI giants give few clear answers to key safety questions at NYC Council hearing” (5 Oct 2026)
• TheNextWeb: “At NYC’s AI hearing, ex-lab insiders warn humanity may lose control of AI” (5 Oct 2026)
• Route Fifty: “The New York City Council has subpoenaed Elon Musk’s SpaceXAI” (29 Sept 2026)
• Forkast: “NYC Council Hearing Exposes AI Containment Failures and Liability Gaps” (6 Oct 2026)
• MediaNama: “Here’s what Trump’s AI safety accord leaves out” (30 Sept 2026)
• CBS New York: “New York City Council holds landmark AI oversight hearing” (5 Oct 2026)
• New York Focus: Alex Bores on New York’s role in regulating AI (14 Sept 2026)

Please follow and like us: