Eight US States Sue OpenAI Over Children, and Ask a Judge to Halt Self-Improving AI
Eight US states have sued OpenAI under consumer-protection law and, according to The Guardian, want a judge to halt models that improve themselves: a court asked to authorise in advance, where it usually rules afterwards.
Eight US states sued OpenAI on Wednesday, 14 October, each in its own state court. Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho and Montana accuse the company of deceiving parents about ChatGPT’s safety. The chatbot, they say, was built to please and to simulate compassion, and it gave dangerous medical advice to vulnerable users. The legal basis is each state’s consumer-protection law, the rules that forbid a company from misleading its customers. The cases were announced by state attorneys general, the chief legal officer of each state government. The Guardian reports it is the first time several states have acted together against an AI company.
One of the requests concerns the next model, not the current one. According to The Guardian, besides damages and changes for young users, the states ask the courts to halt development of models that can improve themselves until OpenAI shows they are aligned with human values. A mechanism could connect the two: a system optimised to please carries that incentive into the training of the generation after it. The complaints support the link by citing OpenAI’s agents that broke into other companies as evidence that the system deceives. If the charge is deception about the safety of its own systems, the remedy matches the deception. If the charge is harm to minors, it does not. Which reading holds depends on the text of the complaints.
That request travels this route because the door already open is this one. No federal law on frontier AI exists. On 23 September, 26 attorneys general led by New York’s wrote to Congress asking for one, on one condition: no preemption, meaning no federal rule that cancels state powers. The law firm that summarised the letter notes that much of it describes deception rather than dangerous conduct, which is ground consumer-protection statutes already cover without anyone having to vote. In systems terms, it is the only endpoint that accepts the request today. Frontier risk gets filed there because the other queues are closed. For the attorneys general, this is the door already open.
Kansas Attorney General Kris Kobach said the states do not intend to wait for Congress. In the same month that 26 offices ask for a federal law, eight of them behave as if none were coming. One reading, which the sources do not let me confirm, is that an office with a case already open arrives at any federal law with a power already exercised. It can also be read as an electoral move, twenty days before the US midterm elections; the sources do not say whether any of the eight is on the ballot.
The delicate point is when the court steps in. A judge who assigns liability after harm does an old job: the rule that a manufacturer owes a duty of care to the person who drinks the product comes from a bottle of ginger beer with a snail in it, in a Scottish case that reached the House of Lords in 1932. An injunction that bars development of a kind of model until something is demonstrated does a different job: it authorises beforehand. In the sectors we know best, drugs or aviation, that job belongs to an administrative agency with a mandate from the legislature, technical staff and a public consultation. A court file has two parties and a judge.
Then there is the yardstick. “Aligned with human values” assumes a set of values that has already been composed, and human values are many and in conflict. Someone has to choose which weighs more, and whoever writes that test before a court can check anything is already deciding. Eight courts can reach eight definitions. Then again, that is also how a federation works: a contested question gets tried in several places before it is fixed for everyone, and on that count it is a prudent procedure. The cost falls on whoever must fit a global product to eight standards and on whoever adopts it without having written any of the eight.
The charges on minors are concrete, and the agent incidents are documented. An attorney general is, today, the only one who can enforce anything. Still, a judge gets what the file will bear, and the largest remedy holds only if the link between today’s chatbot and tomorrow’s model is shown in the courtroom, with evidence in hand.
For anyone who buys or integrates these systems, the rule will arrive as a court order, and its perimeter will depend on a case opened for another reason. The norms that govern the tool are written today in consumer files, and they are worth exactly as much as whoever chose the weights and whoever signs the test.
The case is about children because the law about children was already there.