OpenAI Faces Multistate Attorney General Investigation After IPO Filing

A state attorney general subpoena is not usually a data governance story, but this one reads like one. The document list New York Attorney General Letitia James' office sent OpenAI, on behalf of a coalition The Wall Street Journal reported at 42 states, doubles as an informal audit checklist for any team shipping a consumer AI product: advertising and engagement telemetry, health-data handling, age-assurance practices for minors and seniors, model-behavior records tied to "sycophancy," and internal safety-policy documentation. The scope is the real story here, not the subpoena itself. It previews what state regulators now expect a frontier AI lab to produce on demand, and it lands days after OpenAI confidentially filed for a U.S. IPO.

This line of inquiry did not start with the subpoena. In a December 2025 letter, James and a bipartisan coalition of 41 other attorneys general urged 13 AI chatbot makers, including OpenAI, Meta, and Microsoft, to warn users about "sycophantic and delusional AI responses" and to publicly report where their models produce biased or sycophantic outputs, according to the New York attorney general's office. Sycophancy, in ML terms, is a model's tendency to validate a user's stated belief or emotional state rather than challenge it, a pattern that can be reinforced by engagement-optimized fine-tuning. A subpoena request for sycophancy-related records reads as an attempt to determine whether that tuning tradeoff was tested and documented before shipping, not just whether harmful outputs later occurred.

The subpoena's interest in "internal company policies" has a specific, recent example attached. OpenAI staff had flagged and banned a user's account for disturbing content months before that user was allegedly involved in a mass shooting in Tumbler Ridge, British Columbia, and the company did not alert law enforcement at the time, a lapse CEO Sam Altman later apologized for, according to TechCrunch. That sequence, flag, ban, no escalation, is close to the exact kind of internal-policy record regulators appear to be probing: not whether a safety system exists, but whether it connects to any real-world escalation path.

Timing sharpens the exposure. OpenAI filed confidentially for a U.S. IPO on June 8, 2026, working with Goldman Sachs, Morgan Stanley, and JPMorgan as lead underwriters on a listing that could value the company at up to $1 trillion, up from an $852 billion private valuation set in March, according to Reuters and The Next Web. A live, multistate investigation into a company's core product is the kind of legal proceeding securities lawyers typically treat as material and expect addressed in an S-1 registration statement. How OpenAI frames ChatGPT-related litigation risk once that filing becomes public is likely to be one of the more closely read sections in it.

James' office served OpenAI with the subpoena on June 12, 2026, seeking the categories above. An OpenAI spokesperson said the company takes the attorneys general's concerns "seriously" and intends to "engage constructively" with their offices, a statement reported by Reuters, Bloomberg, and the Journal. The often-cited figure of 42 participating states traces to the Journal's sourcing, people who reviewed the subpoena, and neither OpenAI nor the coalition has published an official list, so treat the count as reported rather than confirmed. The probe adds to a widening set of legal actions against OpenAI, including Florida's June 1 civil suit against the company and Altman personally, and a separate Florida criminal investigation into ChatGPT's alleged role in an April 2025 shooting at Florida State University, according to The Next Web and the BBC.

• •Whether other coalition members file their own civil investigative demands, which would multiply the jurisdictions OpenAI has to satisfy with consistent answers.
• •How OpenAI's S-1 characterizes the investigation once the filing becomes public, since that language sets a template other frontier labs may follow for disclosing AI-specific regulatory risk.
• •Whether the sycophancy line of inquiry produces findings that tie specific fine-tuning or product decisions to documented harms, handing other state regulators a blueprint for similar subpoenas.

regulators are starting to treat engagement logs, safety-eval artifacts, and escalation records as discoverable evidence, and teams that cannot reconstruct that paper trail on request are the ones most exposed.