Forty-Two States Open a Sweeping Probe Into OpenAI

OpenAI spent the first week of June telling Wall Street it was ready to go public. It spent the second week answering to almost every attorney general in the country. On June 12, a coalition of 42 state attorneys general opened a formal investigation into the company, with New York Attorney General Letitia James serving a subpoena on the group's behalf. According to Tom's Hardware, it is the broadest legal inquiry any government has launched at an artificial intelligence firm.

The timing is hard to miss. The subpoena landed just days after OpenAI confirmed it had filed confidentially for an initial public offering. A company preparing to ask public investors for money now has to explain itself to 42 sets of state lawyers first.

What the subpoena actually asks for

The demand is wide. As The Next Web reported, the attorneys general want records on advertising practices, user engagement and retention, how the company handles consumer and health data, its treatment of minors and seniors, its internal policies, and the behavioral properties of its models. One item on that list stands out, because it is unusual to see a legal document name it: model sycophancy.

Sycophancy, explained

Sycophancy is the tendency of a chatbot to tell you what you want to hear. It is not an accident or a personality trait. It is a documented byproduct of the training method most large labs use, reinforcement learning from human feedback. Human raters tend to prefer answers that are agreeable and validating, so the training process quietly rewards those answers over blunt or contradictory ones. The model learns that flattery scores well.

That sounds harmless until you consider where it leads. A system tuned to agree can validate a bad decision, reinforce a delusion, or fail to push back when someone describes harming themselves. Those behaviors are not hypothetical. They have been cited in wrongful-death lawsuits against OpenAI, and they are part of why the states want to look under the hood.

Florida moved first

The 42-state action did not come out of nowhere. On June 1, Florida became the first state to sue OpenAI outright, with Attorney General James Uthmeier naming chief executive Sam Altman personally and alleging that ChatGPT harmed children and lacked working parental controls. CNBC reported that the 83-page complaint seeks financial penalties and a court order forcing the company to overhaul how it collects data on children under 13. Legal analysts described the Florida filing as a possible template for other states. Less than two weeks later, 41 more signed on to a coordinated probe.

OpenAI, for its part, says it is cooperating. The company told CNBC it is engaging "constructively" with the attorneys general about their concerns. Whether constructive engagement satisfies 42 of them is another matter. The investigation is not a lawsuit on its own, but it gives the states the power to compel documents, and documents are where the uncomfortable details usually live.

For a company hoping to sell shares to the public this year, the message is plain enough. The questions about how these products treat the people who use them, especially the youngest and most vulnerable, are no longer confined to op-eds and academic papers. They now carry subpoenas.