Executives from OpenAI, Google, Meta and Anthropic answered questions at City Hall about who pays when an AI system causes serious harm — and whether Albany or the Council writes the rule that decides it.
Wednesday's session ran nearly four hours. Council members pressed the four companies on disclosure, testing records and the gap between a model developer's warning label and a deployed product that acts on a customer's behalf. Most of the argument landed on one question: whether the city can write a liability rule at all, or whether that belongs to the state.
More policy coverageThe Policy committee called the session after a year of complaints from small businesses, schools and city agencies that bought AI tools without any clear statement of who carries the risk if the tool misfires.
Council members opened by describing what they said they see in their districts: a pharmacy that adopted an AI assistant for scheduling and lost a week of shifts to a bug nobody could explain; a nonprofit that signed a vendor contract with a liability clause it could not read past the third page. The companies were asked to say, plainly, who is responsible when a model produces a harmful outcome.
The executives drew a line between the model and the product built on it. Google and Meta argued that a general-purpose model is closer to a press than a product, and that liability belongs with whoever deploys it into a specific setting. Anthropic took the sharper position, saying developers should not be able to hand off an unpriced risk to the smallest customer in the room. OpenAI said it supports a national framework and warned that a patchwork of city rules would push the hardest decisions to whichever jurisdiction is cheapest to sue in.
"If the model ships without a price tag for the harm it can cause, the customer pays it later. That is not a market. That is a subsidy."
A question from the committee chair, paraphrased here, that set the tone for the rest of the afternoon.
Wants one national rule rather than 19,000 city codes. Testified that a city-level liability standard would slow deployment and put smaller vendors out of business before they could absorb the cost of compliance.
Compared a base model to a printing press: the maker builds the machine, the user decides what it prints. Said it supports disclosure requirements and a proposed incident-reporting registry, and would accept a shared liability split in high-risk settings.
Argued that the deployer — the hospital, the school, the retailer — holds the duty of care because it knows the context. Suggested the Council require vendors to hand customers a written risk summary before signing.
Said developers should keep a share of liability even when a downstream partner deploys the tool, and pushed for independent testing that the public can inspect. Its counsel called the current contract landscape "unreadable by design."
Council counsel told the room what everyone in the chamber already suspected: the city can regulate its own purchasing, its own agencies, and the contracts it signs. It cannot rewrite a general standard of care that the state sets.
Every city contract with an AI vendor could carry a plain-language risk clause and a named point of contact for incidents. This is the lever the Council can pull without waiting on Albany.
A bill in committee would require vendors selling to city agencies to state what the system can and cannot do, and what testing was run, before a signature — not in a footnote after.
Members floated a city registry for harmful outcomes tied to AI tools used by agencies, modeled loosely on the 311 complaint system. Whether a voluntary register is worth much was the most contested line of the day.
The hard part. Council counsel said a city cannot invent a new standard of care for private suits, which is why the chamber spent so long arguing about which level of government should. No bill text has been introduced yet.
Wednesday's hearing is the third city-level session on automated decision-making we have covered since spring. Each one narrowed the same gap a little further.
The hearing was informational, and the committee closed without moving any bill. Two members said they expect draft text on procurement disclosure before the end of the session; a third said liability language should wait for the state to act first, because a city rule that courts strike down wastes two years.
Several things are worth watching for readers who follow local AI policy. Whether the disclosure requirement picks up co-sponsors before the next stated meeting. Whether the four companies file written responses to the questions they deflected in the room — Anthropic and Google said they would; OpenAI and Meta did not commit. And whether the state's own liability bill, still in committee in Albany, moves before the Council meets again.
The Council also heard from two people outside the industry: a Brooklyn school administrator who said her district adopted an AI grading assistant without any clear contract language on accuracy, and a small-business owner from Queens who said he had no way to tell whether a vendor's tool would be usable next year. Both asked the same thing in different words — a plain answer about who is on the hook.
No. This was an informational hearing, which means the committee called witnesses and gathered testimony but did not advance a bill. Any draft text would need to be introduced, referred back to committee and calendared before a vote. Members said to expect a procurement disclosure draft in the coming weeks, but nothing was filed the day of the session.
Not a general standard of care. Council counsel explained that the city can control its own contracts, its own agencies and what vendors must disclose before selling to the city — but a broad liability rule for private lawsuits is generally set at the state level. That distinction is why much of the hearing turned into a debate about Albany rather than about a specific bill.
It does not really. Readers who arrived searching for torchys landed on a policy story, not a review. The hearing had nothing to do with restaurant chains or pricing tools built for kitchens — that thread lives in the antitrust case filed this week against McDonald's, which we covered separately and which touches far closer to food-service vendors than anything said at City Hall.
Two of the four said they would send written follow-ups to questions they could not answer on the dais. Written testimony submitted before the hearing is typically entered into the record and posted by the committee. We will update this story if the record is published with new material, and we will note any correction at the bottom of the page rather than quietly editing the text.
Bills, hearings, procurement rules and the contracts behind them — filed with dates, names and the parts still unresolved. If you have a document worth a second pair of eyes, send it over.
