The showdown had been building for weeks. As scrutiny of frontier AI labs intensified, City Council Speaker Julie Menin called a rare Committee of the Whole hearing and demanded that Anthropic, OpenAI, Google and Meta answer questions under oath about the dangers their systems could pose. Three of the companies agreed only after the Council threatened subpoenas, according to Menin’s office.¹

On Monday, the hearing exposed the distance between the Council’s demand for concrete safeguards and the industry’s guarded assurances. Menin asked executives how they quantified AI risk in a worst-case catastrophic scenario. OpenAI policy executive Morgan Dwyer replied: “I don’t know,” adding that it did not matter whether the chance was 1%, 10% or 20%, because no such level of risk was acceptable.²

Menin was unmoved. “To say you don’t know and it doesn't matter is flippant at best,” she said, after Dwyer said OpenAI should not train models that could not safely remain under human control.² The companies did not provide clear answers on whether they could halt a model that failed a safety test, insure against a catastrophe or quantify the probability of one.²

Former employees framed the problem more starkly. Jacob Coxon, who left Anthropic, told councillors that the sector’s “move fast, break things” ethos might suit a photo-sharing app, but not “the most powerful technology ever built.”³ He appeared with former Google and OpenAI workers pressing for stronger oversight.

The sharpest political rebuke came from Public Advocate Jumaane D. Williams, who called Dwyer’s response a “God-awful, fantastically terrible answer.”³ The hearing left both sides aligned on one abstract point — catastrophic risk is unacceptable — but divided over whether the firms can demonstrate that their accelerating systems are actually under control.