OpenAI and Anthropic questioned under oath as NYC weighs AI shutdown rules

What happened at the NYC AI safety hearing?

New York City's Committee of the Whole held an AI oversight hearing at City Hall on October 5, examining risks and proposed protections. The Council's official record confirms the hearing took place and lists the measures considered. Council meeting record

The Next Web reported testimony from former Anthropic researcher Jacob Coxon, former OpenAI researcher Daniel Kokotajlo and former Google DeepMind researcher Alex Turner. Company representatives included OpenAI's Morgan Dwyer, Anthropic's Logan Graham, Google's Alice Friend and Meta's Shane Cahill. The Next Web's hearing report

This brought two competing accounts into the same proceeding: former employees warning that control mechanisms are inadequate, and current representatives defending their companies' approach to safety.

Why were OpenAI and Anthropic asked to testify under oath?

The Council wanted company answers on potential AI dangers and legislative safeguards. In its September 28 announcement, it said OpenAI, Google and Anthropic agreed to participate after subpoena warnings; Meta had agreed earlier. It also announced a subpoena for SpaceXAI. Council announcement

The Council cited its authority to investigate matters concerning city government, compel attendance and take testimony under oath. This was an oversight proceeding, not a criminal trial or a finding that any company had broken the law.

For readers following Anthropic's reported IPO risk disclosures, the hearing raises a related question: what evidence should accompany an AI company's account of the risks its systems could create?

What did former AI employees warn about?

They warned about the possibility that increasingly capable AI could escape effective human control. According to The Next Web, Coxon argued that the current development trajectory could end in loss of control and human extinction. Kokotajlo questioned the reliability of methods used to detect misalignment, while Turner gave his own substantial estimate of takeover risk. The Next Web

These are expert warnings and judgments. Their appearance in testimony does not establish a measured probability of catastrophe.

That distinction matters. A witness can identify a credible failure mechanism without knowing how often it will occur. The policy question is what evidence would justify deployment, what would trigger intervention and who would have the authority to act.

How did OpenAI respond to questions about catastrophic AI risk?

Speaker Julie Menin asked company representatives to quantify worst-case catastrophic risk. AP reported that OpenAI's Morgan Dwyer said she did not know, while arguing that no level of catastrophic risk was acceptable. Associated Press

The Next Web reported that Anthropic's Graham supported a role for state and local regulation, while Google's Friend favoured a comprehensive federal framework. The Next Web

The exchange exposed a practical gap: a commitment to safety does not, by itself, identify a testable threshold for deciding whether a system is safe enough to deploy.

Our reading is that regulation will increasingly turn on those thresholds. Who evaluates a model, what they can inspect and what happens after a failed assessment could matter as much as the stated commitment.

Would NYC require an AI kill switch and independent safety testing?

One proposed bill would require both independent validation and a technical capability for a human operator to stop an AI model functioning, temporarily or permanently.

The draft covers marketing, offering for sale, selling and deploying AI models in the city. Assessments would include task performance, bias, data handling and safety. Validators would have to disclose their financial or other interests.

Specified violations, including offering a model without validation or a shutdown capability, would carry a $25,000 penalty per instance. The draft would take effect 180 days after becoming law. Full validation and shutdown proposal

An AI kill switch is therefore part of the proposal's substance. Whether a shutdown mechanism would work reliably across real deployments is a separate technical question. The draft requires assessment of its functionality; it does not demonstrate that every system can already meet that requirement.

Could AI companies be sued when someone misuses their models?

A separate proposal would create a private right of action for harm occurring in the city through third-party misuse or malicious use of an AI model.

The draft requires three conditions: misuse caused the harm, the provider failed to implement reasonable safeguards, and the harm was foreseeable. It would allow claims for damages and other remedies, including injunctions. Proposed private cause of action

This would not automatically make a provider responsible for every harmful output. The proposed test connects liability to foreseeability and inadequate safeguards.

Economically, that could make records of risk assessment and mitigation more consequential. A safety process would need to support the company's decisions when examined after a harmful event. That is our analysis of the draft, not an estimate of future litigation costs.

Would AI safety incidents have to be disclosed within 24 hours?

A proposed city-contracting measure would create reporting obligations for covered contractors and city agencies. Contractors would notify Cyber Command within 24 hours of becoming aware of a reportable incident; agencies would face a similar duty.

Cyber Command would then have a 24-hour public disclosure obligation after becoming aware of an incident. The draft excludes information whose publication would compromise specified interests, including cybersecurity, trade secrets, public safety or national security. Incident reporting proposal

The scope matters: this is a proposal concerning city contracts and related incidents, not a blanket 24-hour reporting rule for every AI company worldwide.

Its potential value is reducing the gap between an organisation discovering a failure and the public learning that it happened.

What would change for chatbot privacy and AI safety advertising?

The chatbot proposal would establish privacy, security and transparency requirements. It includes access to users' own data, restrictions on data use, disclosure that the user is interacting with a chatbot, and a prohibition on suggesting equivalent advice to a licensed professional. Chatbot privacy proposal

Another proposal targets advertising. It would require AI advertisements in the city to disclose whether a model had received third-party validation and prohibit materially false or misleading statements about risks, safeguards or validation. Violations could carry penalties of up to $25,000. AI advertising proposal

Together, these measures connect frontier safety debates to the claims consumers encounter when choosing a product. The distinction between a vendor's assurance and an outside assessment would become visible at the point of promotion.

Would whistleblowers receive a share of recovered penalties?

A civilian-enforcement proposal would let eligible individuals submit evidence of certain AI violations to the city's Department of Consumer and Worker Protection.

If the city pursued a qualifying complaint and recovered proceeds, the complainant would receive 25%. The share would rise to 50% when an authorised complainant served a violation notice or a designated person brought the action under the bill's procedures. Civilian enforcement proposal

The Council has described the package as including whistleblower incentives. This particular draft is broader than employee whistleblowing: it creates a route for eligible members of the public to report covered violations. An allegation alone would not trigger a payment.

Has New York City passed these AI regulations?

The proposals examined here have not been shown as enacted in the legislative records checked for this article. The liability and civilian-enforcement records, for example, list committee consideration and leave enactment dates and law numbers blank. Holding a hearing is not the same as passing a law. Liability bill status, enforcement bill status

The package could still change. Its eventual effect would depend on the final text, implementing rules and enforcement.

For developers and businesses, the immediate signal is a set of questions worth preparing to answer: can the system be stopped, can an outsider evaluate it, can incidents be identified promptly, and can safety claims be supported?

Reporting note: We reviewed the Council's announcements, meeting record and published bill texts. Accounts of spoken testimony are attributed to AP and The Next Web; we did not independently review a complete hearing transcript or recording. Interpretation is identified as our analysis.