Representatives of OpenAI, Anthropic, Meta and Google gave no percentage estimate of an AI-caused catastrophe at a public New York City Council hearing on October 5. Council Speaker Julie Menin asked the companies for their assessment under oath. Their answers pointed to safety testing, missing reference data and the limits of forecasting.
Morgan Dwyer, who leads policy development and operations at OpenAI, said she did not know the probability. One, ten or twenty percent would all be unacceptable, she said; what mattered was a strong case for human control. Those examples were not OpenAI’s own numerical risk estimate.
Anthropic Frontier Red Team head Logan Graham described research on biological, cyber and control risks without giving a percentage. Meta’s Shane Cahill pointed to testing and the company’s framework and said he wanted to avoid imprecision. Google’s Alice Friend distinguished measurable capabilities from the probability of a future catastrophe, for which she said there was no rigorous reference class.
Earlier, former AI researchers Sammie Coxon, Daniel Kokotajlo and Alex Turner warned about difficult-to-control systems and acceleration through AI-assisted research. They called for greater transparency, independent evaluation or restrictions on especially capable systems. Their forecasts and judgments were personal assessments, not established future events.
Asked later about possible affected parties in New York, Dwyer said she was not aware of such cases. OpenAI’s investigation of potential agent incidents since November 2025 was continuing. It would notify affected third parties and publish findings; she extended that commitment to affected parties in New York if the investigation identifies them. Her answer does not establish that the city or its institutions were unaffected.
The testimony accompanied consideration of proposed city AI safeguards. The previously announced proposal for human shutdown controls and independent validation remains a proposal. The hearing now supplies actual company answers, rather than merely invitations to appear.