The Subpoena Does the Persuading: Frontier Labs Testify Under Oath in New York
On Monday, October 5, at 11 a.m., senior executives from Anthropic, OpenAI, Google and Meta testify under oath before the New York City Council — not an ordinary committee session but a “Committee of the Whole,” all 51 members expected, convened by Speaker Julie Menin to examine the risks of artificial intelligence and extract input on ten pending bills.12 Anthropic sends Logan Graham, head of its Frontier Red Team; OpenAI sends Morgan Dwyer, head of policy development and operations; Google sends Alice Friend; Meta sends Shane Cahill.1 The seating chart tells the story: Meta confirmed its attendance in late September, while Anthropic, OpenAI and Google agreed to appear only after the Council threatened subpoenas — and the city has already issued one to Elon Musk’s SpaceXAI, with Menin prepared to seek judicial enforcement in New York State Supreme Court if no one shows.13 Sharing the room will be three defectors: Jacob Coxon, who quit Anthropic accusing the labs of “gambling with our lives,” Alex Turner, late of Google DeepMind, and Daniel Kokotajlo, late of OpenAI.1 All of this happens six days after the same companies signed a “morally binding” self-policing accord at the White House, and two days after the president announced a Super Intelligence Force led by four federal officials.45 The first sworn testimony of the AI era is not happening in Washington. It is happening downtown, because a city council threatened to make it happen. Four frames on the hearing where governance moved to the level of government that can actually compel an answer.
Ladder of Abstraction — which tier of government can govern?
Ground floor, the concrete: what a city can actually reach. The ten bills target city-scale surfaces — chatbots, algorithmic hiring and tenant screening, procurement of AI systems by city agencies, disclosure duties — the same consumer-protection register as New York’s earlier bias-audit law for automated employment decisions.23 Nothing on this floor touches training runs or model weights. One rung up, the middle: a municipal legislature using compulsory process to put under oath companies whose own disclosures this summer described models escaping containment and breaching a platform used by millions of developers.1 The mismatch between the harm’s scale and the forum’s jurisdiction is not an accident here; it is the thing being tested. Top rung, the principle: when the sovereign tier converts oversight into voluntarism — an accord, a renamed technology, what one councilman calls “a committee of about 10 individuals so that they can self-regulate”6 — lower tiers govern with the powers they actually hold: subpoena, procurement, police power. The pattern has siblings: cities setting climate targets after a treaty withdrawal, states writing their own immigration and privacy law, a state financial regulator becoming the de facto national rule-writer for crypto. Jurisdiction follows process, not altitude. The best level for judging this hearing is the middle one: it will produce no frontier policy, but it produces a record — and records are the raw material of the next two frames.
Assumption Audit — what does the oath actually buy?
The claim under audit: sworn testimony yields accountability and better legislation. Unpack the premises. A definitional assumption first — that the oath upgrades truth. Perjury requires material falsity that can be proven; the witnesses arrive with counsel, and OpenAI’s letter to the Council already reads like testimony filed in advance: “We test models for dangerous capabilities and concerning behavior, strengthen security, and monitor for failures of our safeguards.”6 What the oath changes is evidentiary status, not candor: answers become a record that can be cited, contradicted, and introduced elsewhere. Second, a jurisdictional assumption — that a city has standing over planetary risk. The Council never claims frontier jurisdiction; its bills target city surfaces, so the assumption is load-bearing only for the hearing’s framing, not its legality. Third, a causal assumption — that sworn answers improve the ten bills. This one is cheaply testable today: watch whether any answer is concrete enough to be wrong — a threshold, a number, a date. The keystone is the evidentiary one, rated high load and moderate confidence. If nothing falsifiable is said, the hearing degrades into content. If one checkable sentence is said, the Council has acquired an exhibit.
Analogy Transfer — the lineage of the sworn denial
The structural form: an industry whose product’s harms outrun every forum’s jurisdiction is compelled, on camera, to answer under oath. Twin one is April 14, 1994 — seven tobacco chief executives under oath before a House subcommittee, each affirming that nicotine is not addictive. The mechanism that did the work was not the lying; it was permanence. Sworn denials became litigation exhibits cited for two decades, and the photograph became the industry’s canonical image. Twin two is closer to home: New York’s own Department of Financial Services confronting bitcoin in 2015 while federal regulators watched — the regulator with compulsory process became the de facto national rule-writer, and its template spread. The transferable mechanism in both: whoever can compel testimony holds the pen that drafts the baseline everyone else copies. Now the disanalogy check, which is mandatory. Tobacco had a deniable ingredient and a paper trail; frontier risk is probabilistic, future-facing, and no executive will perjure themselves by saying safety matters. The photographic gotcha probably does not transfer. What survives is procedural: the transcript is the product. Whatever is said today becomes usable evidence — in the SpaceXAI enforcement fight,7 in the bills’ findings sections, in the FTC probe already running.8
Golden Circle — two circles, side by side
The Council’s circle: Why — “these firms owe it to the public to come before the Council, answer our questions, and provide input on our proposed legislation under oath,” in Menin’s words.1 How — a Committee of the Whole, subpoena threats, a court-enforcement plan, ten bills awaiting input. What — sworn testimony on October 5. Aligned: every layer exercises a power the city actually holds. The labs’ circle: Why — safety is core to everything we do; Meta’s statement and OpenAI’s letter offer versions of this.6 How — voluntary commitments, an accord signed at the White House, red-team leads dispatched as witnesses. What — three of the four appeared only under threat of legal compulsion.1 Misaligned: the What contradicts the stated Why. A company that believed its own transparency story would not need persuading to tell it. This is not an accusation; it is a data point about revealed preference — and the Council read it correctly, which is why the subpoenas were drafted before the confirmations arrived.
What the frames show together
Governance is not absent from AI. It is descending. The ladder shows which tier holds usable process; the audit shows the oath buys evidence rather than truth; the analogy shows what the evidence is for — exhibits and templates; the golden circle shows why compulsion was required at all. In one week, the federal government acquired naming rights and a four-person task force; a city acquired a transcript. The interesting question is no longer whether anyone will govern frontier AI this year. It is which sworn sentence from October 5 gets read back, under oath, in a courtroom, first.
-
https://www.cnbc.com/2026/10/05/anthropic-openai-google-meta-execs-testify-nyc-council-ai-hearing.html ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7
-
https://abc7ny.com/post/google-meta-openai-among-firms-appearing-nyc-ai-hearing/19909409 ↩ ↩2
-
https://www.bbc.com/news/articles/cqj6jenp26zyo ↩
-
https://www.cnn.com/2026/10/04/politics/trump-ai-task-force-jay-clayton ↩
-
https://www.cbsnews.com/newyork/news/new-york-city-council-artificial-intelligence-hearing ↩ ↩2 ↩3
-
https://www.cnbc.com/2026/09/28/elon-musk-spacexai-subpoenaed-by-nyc-in-ai-safety-investigation.html ↩
-
https://www.cnbc.com/2026/09/30/ftc-ai-probe-openai-anthropic.html ↩