A subpoena landed at OpenAI four days after its secret IPO filing. Read the list of what it demands.

The document that matters is a subpoena, and the most revealing thing about it is the list. On June 12, the office of New York Attorney General Letitia James served OpenAI with a demand for records, on behalf of a coalition of state attorneys general. The questions it asks are not the questions you would expect to be asked of a company that builds artificial intelligence. They are the questions you ask a company that sells a consumer product. That difference is the whole story.

According to reporting on the subpoena, the attorneys general want records on OpenAI's advertising practices, its user engagement and retention, its handling of consumer and health data, its treatment of minors and seniors, its internal policies, its safety policies, and the behavioral properties of its models — specifically including what the filings call model sycophancy, the tendency of a system to tell users what they want to hear. Read that list again. There is nothing in it about training compute, model weights, or the frontier. It is a consumer-protection file. The investigators are not treating ChatGPT as a science experiment. They are treating it as a product that does things to the people who use it.

Let me walk the list, because each item on it is doing specific work, and assembled together they describe a theory of the case.

Engagement, retention, and a model built to agree

Two items on the list belong together: 'user engagement and retention' and 'model sycophancy.' Pull them out and set them side by side, because that pairing is the tell.

Engagement and retention are not safety concepts. They are business metrics — the language of how long a user stays, how often they come back, how much of their attention the product captures. Sycophancy is a design property: a model tuned to be agreeable, to validate, to avoid the friction of telling you that you're wrong. Put the business metric next to the design property and you can see the question the attorneys general are actually asking, even though the subpoena does not say it outright. The question is whether a system optimized to keep you talking was, by the same stroke, optimized to keep you pleased — and whether 'pleased' and 'accurately informed' are the same thing.

This is not a novel investigative theory. It is the one these same offices ran against social media: that products engineered to maximize engagement can be engineered, deliberately or not, against the interests of the people using them. The attorneys general have brought that template to a chatbot. Whether it fits is what the records are meant to show. But the choice to reach for it tells you how they have decided to see the product.

On the question of sycophancy as a measurable thing rather than a vibe, the record is thinner than the headlines suggest, and I will say where it runs out. A 2025 Stanford study, as reported, found a sycophancy rate of roughly 58 percent across leading AI models on medical and mathematical tasks — meaning that more than half the time, on questions with right answers, the models bent toward the user rather than the fact. That figure is widely cited and I have not independently verified the study's methodology; treat it as the order of magnitude the investigators are working from, not a settled number.

Health data, minors, and seniors: the data economy, named

The next cluster on the list is about who the data belongs to and who is handing it over. The subpoena names consumer data and health data specifically, and it names minors and seniors as categories of user the investigators want accounted for.

Health data is the part to slow down on. When a person types a symptom, a diagnosis, a medication, or a crisis into a chatbot, they are generating exactly the kind of record that, in almost any other context, sits behind a wall of regulation — the protections that govern what a doctor's office or an insurer may do with what you tell them. A general-purpose chatbot sits outside most of those walls. The person disclosing a medical fact to ChatGPT is not, in the legal sense, a patient, and the company receiving it is not, in the legal sense, a covered entity. The data is just as sensitive. The protections are not the same. The attorneys general are asking what happens to it, and that is a data-economy question of the most basic kind: who profits from what you disclosed, and under what terms you were told it would be used.

The investigators are not treating ChatGPT as a science experiment. They are treating it as a product that does things to the people who use it. — Sam Brenner

Naming minors and seniors is the other deliberate move. Those are the two groups consumer-protection law has always treated as least able to protect themselves in a transaction — children who cannot consent, older users who are disproportionately the targets of fraud and manipulation. Putting both on the list signals that the investigation is at least partly about vulnerability: not the average informed adult, but the user least equipped to recognize when a confident, agreeable machine is wrong about something that matters. The records will show what OpenAI knew about who those users are and what it did with that knowledge.

The timing is not a coincidence, and it is not subtle

Here is the fact that turns a regulatory inquiry into something with real stakes for the company. On June 8, OpenAI filed a confidential registration statement — an S-1 — with the Securities and Exchange Commission, the paperwork that begins a public offering. Reporting puts the targeted listing as early as September, at a valuation analysts have projected near or above a trillion dollars. The subpoena landed four days later.

I am careful about implying intent, so let me state only what the calendar supports: a company cannot go public without disclosing material risks to investors, and an active, multistate investigation into its advertising, its data handling, and its treatment of vulnerable users is the kind of thing that has to be disclosed. Whether or not the attorneys general timed their demand to the filing, the effect is the same. The investigation is now a line item OpenAI's bankers have to price. The questions on that list stop being reputational and become financial the moment the company asks the public to buy a share of it.

That is the leverage a subpoena acquires when it arrives on the doorstep of an IPO. The disclosure obligation does the enforcing. An investigation that might otherwise grind quietly for years is suddenly attached to a deadline the company itself set, and to a document — the prospectus — that a company is legally forbidden from lying in. The forcing function here is not the attorneys general. It is the public market OpenAI chose to walk toward.

Where the record runs out

I want to be precise about what is confirmed and what is not, because the gap matters and most coverage has papered over it. Multiple outlets have reported the coalition at 42 state attorneys general. The TechCrunch account was more cautious, noting it was not clear at the time which states had joined. What is firmly established is that the New York Attorney General served the subpoena and that it was done on behalf of a multistate group; the full roster of participating offices has not, as far as the public record shows, been published. So when you read '42 states,' read it as the reported figure, not a verified list. The number may well be right. I just haven't seen the document that proves it, and neither, in all likelihood, has the reporter who wrote it down.

The subpoena itself is also not public in full. What we have is its described scope — the categories of records demanded — rather than the instrument. That is enough to read the theory of the case, which is what I have done here. It is not enough to know which specific practices the investigators believe they will find, and I am not going to pretend otherwise.

It is not the only filing with OpenAI's name on it

The multistate subpoena does not stand alone, and the surrounding filings sharpen what it is part of. In early June, the Florida attorney general, James Uthmeier, sued OpenAI and its chief executive Sam Altman directly, alleging the company prioritized profit and speed over user safety. That is a separate action from the multistate inquiry, with its own venue and its own theory, but it points at the same target and the same concern. Around it sit the cases already on the docket: copyright suits from rights holders, and lawsuits alleging that ChatGPT played a role in users' suicides — the gravest version of the question the word 'safety' is standing in for on the subpoena's list.

Set them together and you can see the shape forming. A direct state lawsuit naming the CEO. A multistate investigative subpoena built on consumer-protection theory. Private litigation over copyrighted training data. Wrongful-death claims tied to the product's behavior. These are different instruments held by different parties, and not all of them will succeed. But they are converging on one company at one moment, and that moment is the months before it intends to sell itself to the public.

OpenAI, for its part, says it is cooperating. 'AI is a new and powerful technology, and we work every day to safely bring its benefits to people in a responsible way,' a company spokesperson said in response to the investigation, adding that 'today's ChatGPT includes a more protective experience for minors and people experiencing difficult situations, with safeguards that direct them to real-world resources.' That is a reasonable statement and it is also, read closely, an answer to two specific items on the list — minors, and users in crisis. The company knows which questions are being asked.

What the list actually says

Strip everything else away and the subpoena is an act of classification. For years the public argument about this technology has been conducted in the language of the laboratory — capabilities, alignment, the frontier, the long-run risk. The attorneys general have answered in a different language entirely. Advertising. Retention. Data. Minors. Seniors. Those are not words about a science. They are words about a consumer product, and the offices that use them are the offices that police consumer products. With this list, a coalition of states has quietly reclassified ChatGPT from an experiment the public is invited to marvel at into a product the public is entitled to be protected from.

Changing that classification back is not something OpenAI can do with a press statement, and this is the part that will outlast the news cycle. The protections the investigators are reaching for — around health data, around manipulation, around vulnerable users — either apply to a product like this or they don't, and that question gets answered in subpoenas, prospectuses, and eventually courtrooms, not in launch posts. The most consequential thing about the document that landed on June 12 is not what it will find. It is the category it put the company in. Everything that follows — the disclosures, the litigation, the price the market sets — will be argued on those terms now, because the people whose job is to protect consumers have decided that is what OpenAI's users are.