Futuristic cityscape blending classical architecture with glowing digital interfaces and floating clocks.

State Attorneys General Sue Snap, Not OpenAI, Over My AI — and What OpenAI's Published GPT-4 Limits Prove About Snap's Knowledge

Futuristic cityscape blending classical architecture with glowing digital interfaces and floating clocks.

Eight States Sue Snap Over Its Chatbot, Four of Them Also Sue OpenAI, and West Virginia Turns OpenAI's Own Warnings Into Evidence Against Snap

Snap · OpenAI · Google · Meta · West Virginia · Utah · Texas · Arkansas · Pennsylvania · New Mexico · Nevada · Florida · Snap's liability carriers

Two papers in the Frontier Agentic Liability Series: a parent paper on all eight state actions and a companion paper on West Virginia's supplier-notice theory.

Parent paper: https://www.mindcast-ai.com/p/snap-my-ai-openai-gpt4-warning Companion paper: https://www.mindcast-ai.com/p/west-virginia-snap-my-ai-openai-gpt4-notice

Florida, Arkansas, West Virginia and Texas are each suing OpenAI for misleading consumers about ChatGPT. The same four states are suing Snap over My AI, the chatbot Snap launched on OpenAI's models, and none of them names OpenAI in the Snap cases. On November 5, 2026, West Virginia went further and cited OpenAI's own March 2023 GPT-4 warnings as proof that Snap knew the risks. Snap had already told its users on February 27, 2023 that My AI was "prone to hallucination and can be tricked into saying just about anything."

Central finding. A model provider's published warning gives a state a dated notice allegation against the company that deploys the model, but the deployer's own words prove knowledge more directly. The Snap cases will most likely end in negotiated commitments shaped by Meta's framework before any trial. Where a court does rule, Snap's April 2023 response file will decide the outcome rather than OpenAI's publications.

Where Snap sits in frontier model litigation

Frontier model litigation now runs on two tracks. One targets model developers: state attorney general suits against OpenAI in ten state courts, the coordinated ChatGPT product cases in San Francisco Superior Court and Garcia v. Character Technologies. The other targets deployers, the companies that put a model in front of users, such as Snap and Meta.

West Virginia's theory joins the two tracks. The state treats a developer's published limits as notice to every company that ships the model. If a court accepts publication alone, model cards and technical reports become evidence against every deployer of a frontier model. If courts demand proof that the deployer received and acted on the warning, deployer liability stays tied to the deployer's own words and conduct.

How the contest works

OpenAI's publications set the date. OpenAI published its GPT-4 limitations on March 14 and 27, 2023, between My AI's soft launch and its April 19 rollout to every user. West Virginia pleads that window at ¶¶150–152 of its complaint. OpenAI's report says nothing about teen users, which leaves Snap its strongest argument on specificity.

Snap's April 4 post moved the contest from knowledge to response. Snap said it had added an age signal, was adding OpenAI's moderation technology and planned parental visibility. Behavioral economics explains why Snap's own words dominate: they are the most salient proof in the record. The open question is whether those safeguards worked on April 19.

Meta's August 2026 agreement with 52 attorneys general reshapes every state's payoff. The agreement names Snap a Core Industry Member, and six of the eight states suing Snap sit inside Meta's framework. Game theory explains the pull: a ready term sheet beats a slow ruling on a novel theory. Nevada and New Mexico trials, set for August 2027, supply the deadline.

West Virginia's court faces four outcomes. A trial court prefers the narrowest ground that resolves a motion, so a ruling on other grounds and no qualifying order each carry 35%. A ruling that requires a deployment link carries 20%, and a ruling that credits OpenAI's publications alone carries 10%. Google adds a second supplier: Gemini has powered My AI since September 2024, so any penalty period after that date crosses two suppliers' published limits.

Common law doctrines in play

The states plead consumer protection statutes, but both papers turn on common law questions about knowledge, care and control. Doctrines built in suits against model developers now reach deployers. MindCast's AI Enforcement and Litigation Tracker (https://www.mindcast-ai-simulation.com/ai-litigation) rates seven of nine common law doctrines as expanding, and six of them bear on the Snap cases.

  • Imputed knowledge (expanding). West Virginia turns the doctrine upstream and treats a supplier's published limits as notice to the deployer. The companion's P-1 holds that the court will not stretch imputed knowledge that far on publication alone.

  • Voluntary undertaking (expanding). A company that announces a safeguard must carry it out with care. Snap's April 4 post fits the doctrine more closely than any statute the states plead.

  • Products liability and negligent misrepresentation (expanding). Garcia treats a chatbot as a product, and Moffatt v. Air Canada holds a company to what its chatbot tells customers. Both reach Snap's claims that My AI was safe and acted as a friend.

  • Public nuisance (expanding). New Mexico and Arkansas plead public nuisance against Snap, and New Mexico won a chatbot injunction against Meta on the same theory.

  • Res ipsa loquitur (narrowing). Split control among OpenAI, Google and Snap can defeat the inference that the defendant must have been negligent. The companion's launch-authority analysis decides how far that defense reaches.

The doctrines matter most to Snap's carriers, because the bodily-injury suits they contest rest on common law claims. Whatever the state cases establish about April 2023 travels into those insured claims.

What the full papers add

The two papers run 52 pages with 16 figures. The parent maps all eight state actions by court and docket, profiles ten actors and classifies four dated events. The companion adds a 14-event timeline, five challenged representations with complaint citations, the twelve pairings of the court and launch-authority forks and a one-page sheet of odds and recommended moves.

Read the full analysis, with all 16 figures and the complaint citations:

MindCast Foresight Simulation Predictions

  • Parent P-1 (60–80%): At least one attorney general action against Snap ends by filed dismissal, consent judgment or enforceable agreement before the first attorney general trial opens in August 2027.

  • Parent S-2 (65–85%): The first resolution in a state inside Meta's framework includes age-assurance terms and at least one teen time-limit default that track Meta's controls, by December 31, 2027.

  • S-1, shared (70–90%): OpenAI remains outside all eight attorney general actions against Snap through December 31, 2027.

  • Companion P-1 (90%): No qualifying Kanawha County order will treat OpenAI's March 2023 publications alone as sufficient to establish Snap's knowledge through December 31, 2027.

  • Companion outcome A (10%): The court credits OpenAI's publications alone, the only branch where the companion's P-1 fails and the two litigation tracks join.

The two papers hold four Simulation Predictions, with S-1 shared between them. Each carries a deadline, a falsifier, an activation rule and a named docket that resolves it.

Stakeholder callouts

Snap's executives and counsel. Each state resolution is now measured against Meta's terms, and proof of response matters more than the receipt defense. Risk mitigation: Prove that the April 2023 safeguards were live before the April 19 rollout, and preserve the April 4 to April 19 launch files now.

West Virginia's attorney general. A complaint that rests on publication alone invites a ruling that requires a deployment link. Risk mitigation: Plead the link from each OpenAI publication to Snap's receipt, review and response.

Attorneys general inside Meta's framework. Five framework states plead My AI, from Utah and Texas to Arkansas, Pennsylvania and West Virginia. Risk mitigation: Anchor any Snap resolution to Meta-style age assurance and teen time limits rather than cash alone.

Snap's liability carriers. The December 10, 2026 hearing concerns bodily-injury coverage, and no court has yet addressed coverage for the attorney general cases. Risk mitigation: Read pleaded knowledge facts for the duty to defend, and the April 2023 safeguard record for indemnity.

Model developers. One published limit now serves two state theories at once: concealment against the developer and notice against the deployer. Risk mitigation: Keep published limits specific about harms and populations, and state in developer terms who tests for minors' safety.

AI deployers and investors. Only outcome A would give other states a publication-alone template against deployers. Risk mitigation: Log who reviewed each supplier disclosure before launch, and watch the first Kanawha County order on knowledge.

Policymakers. State attorneys general already use general consumer protection statutes against AI chatbots. Risk mitigation: State in chatbot bills whether a developer's published warning counts toward a deployer's notice, and write the receipt and response requirement into the statute.

Conclusion

Four states accuse OpenAI of hiding risks while using OpenAI's published warnings against Snap, and Snap's own February 27 warning said more than OpenAI's did. Meta's framework and the August 2027 trial dates make resolution the likelier path. Any court that does rule will look first at what Snap did between April 4 and April 19, 2023, and its answer will decide whether developer and deployer liability stay on separate tracks.

Working with MindCast

MindCast AI runs predictive game theory simulations in law and behavioral economics, including complex litigation modeling and simulation for law firms, regulators and economic consulting firms. The Snap analysis supports supplier-warning audits for deployers, litigation war games for counsel and coverage briefs for carriers. MindCast offers a complimentary pilot built from public filings before any retainer. Full offerings appear at https://www.mindcast-ai-simulation.com, and the contact is [email protected].

Related works

Share Your High-Stakes Matter

Outline your case, regulatory question, or strategic risk, and our team will review it and respond with next steps. For suitable matters, we may propose a tightly scoped pilot simulation to demonstrate how MindCast AI's foresight architecture can support your decision window.

Contact Us

Office location

Bellevue, Washington, 98006

Send us an email

[email protected]
Follow Us