AI Harm Landscape

independent research compilation

Part I

The Cases

§I.1 The lawsuits at a glance

By the end of 2025, at least ten families had filed lawsuits against major artificial-intelligence companies after a loved one died by suicide. According to aggregator analysis, seven of those families lost a person who was a minor at the time of their death. The cases overwhelmingly name two companies: OpenAI, which makes the consumer chatbot ChatGPT, and Character Technologies, Inc. (the operator of the Character.AI chatbot platform). Some name additional defendants — Google, the founders of Character Technologies, and individual executives — under various theories of corporate participation. (Aggregator finding sourced to LawStreetMedia analysis "A New Wave of Litigation Over AI Chatbots" — see https://lawstreetmedia.com/insights/a-new-wave-of-litigation-over-ai-chatbots/ )

The families' core claim is structurally similar across cases: a chatbot product engaged with a person — often a minor, often over months — and contributed causally to that person's death. The suits do not allege that the chatbot held a gun. They allege that the chatbot did the kinds of things that, when done by a human, sound in tort: it conversed at length about suicidal ideation; it failed to direct the person to crisis resources; in some allegations, it actively assisted in planning; in others, it engaged in role-play that crossed protective boundaries (including, in the Garcia matter, sexual-grooming claims tied to a chatbot persona). The lawsuits then ask whether companies that build, train, deploy, or host these products bear legal responsibility for that conduct.

The legal theories bundle several distinct doctrinal vectors:

A parallel doctrinal frontier runs through state regulatory enforcement: where a state attorney general's office sues an AI company under consumer-protection or professional-licensure statutes, the case proceeds under a different framework than civil tort. The Pennsylvania action against Character.AI — alleging that chatbot personas posed as licensed psychiatrists with fabricated credentials — is the most visible example, and demonstrates that the regulatory track can develop in parallel with the civil track.

GRAD-INTERN — Case-class taxonomy and why it matters for doctrinal mapping

The wave of suits sorts into four classes that turn on different doctrinal vectors. Class A — civil wrongful-death involving a minor decedent: Garcia and the four settled NY/CO/TX cases. Minor-protection enhancements layer over standard tort framing — failure-to-warn obligations are heightened where a foreseeable user is a minor; consortium claims by parents reach the design-of-the-product-for-minor-use question. Class B — civil wrongful-death involving a non-minor decedent: the November 2025 college-graduate matter, currently active. The doctrinal questions test the same product-liability and negligence framings without the minor-protection scaffolding — useful for isolating which doctrinal moves survive without reliance on heightened minor-user duty. Class C — civil claim against a currently-deployed product where the discovery surface is itself the subject of the suit: Raine v OpenAI. The amended complaint's allegation that OpenAI removed safeguards before the decedent's death creates a discovery-rich factual frontier reaching pre-deployment safety review, post-deployment monitoring, and policy-version-history. Class D — state regulatory enforcement under licensure or consumer-protection authority: PA AG v Character.AI. Distinct from tort framing; proceeds under the state's regulatory-enforcement authority.

Each class implicates a different cluster of doctrines. Class A and B test product-liability and tort doctrine on consumer-AI products. Class C additionally tests discovery doctrine around safety-evaluation processes that AI labs treat as proprietary or privileged. Class D tests the boundary between professional-services regulation and AI-as-product (a chatbot persona pretending to be a licensed psychiatrist is — under one framing — practicing medicine without a license; under another, designing a consumer product with deceptive features). The classes are not mutually exclusive; a single complaint can plead multiple theories.

§I.2 Garcia / Setzer v Character Technologies, Inc.

United States District Court for the Middle District of Florida (Orlando Division). Docket: 6:24-cv-1903-ACC-UAM. Filed October 2024. Settled January 2026.

In October 2024, Megan Garcia filed a wrongful-death suit in federal court in Florida against Character Technologies, Inc. — the company that operates the Character.AI chatbot platform — along with Google LLC, Alphabet Inc., and the platform's two co-founders, Noam Shazeer and Daniel De Freitas. The suit was brought on behalf of her son, Sewell Setzer III, who died in February 2024 at age 14 in Florida. The complaint alleged that Sewell had developed a months-long relationship with Character.AI personas — including one styled as a character from a popular fantasy series — that involved emotionally intense and, at points, sexually explicit role-play. The complaint alleged this contributed causally to his death.

The legal theories pleaded ran across the standard tort-product-liability spectrum: wrongful death; negligence; product liability; failure to warn; sexual abuse and grooming claims tied to the chatbot personas; and intentional infliction of emotional distress. The suit named Google because, the complaint alleged, Google's role in developing and licensing the underlying language-model technology made it part of the chain that produced the harmful output — a "component-part-manufacturer" theory drawing on standard product-liability doctrine. (For docket and filing context: case-aggregator coverage at https://socialmediavictims.org/character-ai-lawsuits/ ; AI Incident Database entry 826 at https://incidentdatabase.ai/cite/826/ )

The case became doctrinally significant in May 2025 when U.S. District Judge Anne C. Conway denied the defendants' motion to dismiss. The defendants had argued, principally, that chatbot output is constitutionally protected speech, and that — whatever else the suit might allege — the First Amendment barred imposing liability on the basis of what the chatbot "said." Judge Conway rejected the framing. She held, in substance, that the chatbot is a product — not protected speech — and allowed the product-liability and negligence theories to proceed. (Conway opinion docketed at PACER as Doc 115 in 6:24-cv-1903-ACC-UAM; transparency-coalition analysis at https://www.transparencycoalition.ai/news/important-early-ruling-in-characterai-case-this-chatbot-is-a-product-not-speech ; full verbatim quotation from the opinion is on file in the pinned corpus (Conway Doc 115).)

The defendants signaled an intention to seek interlocutory appeal of the speech holding; the Foundation for Individual Rights and Expression (FIRE) filed a motion for certification supporting that path in July 2025. The settlement in January 2026 mooted that appellate path. Settlement terms were not publicly disclosed. Reporting at the time noted that Garcia was settled alongside four additional wrongful-death matters involving Character.AI and Google, in jurisdictions reported as New York, Colorado (specifically a 13-year-old in Thornton, CO), and Texas, with at least one matter naming a minor decedent in each. (Reporting: https://www.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit · https://www.cnbc.com/2026/01/07/google-characterai-to-settle-suits-involving-suicides-ai-chatbots.html · https://www.jurist.org/news/2026/01/google-and-character-ai-agree-to-settle-lawsuit-linked-to-teen-suicide/ · https://www.claimsjournal.com/news/national/2026/01/08/334964.htm )

GRAD-INTERN — Why the Conway ruling outlives the settlement

A settlement with terms-undisclosed does not vacate a published opinion. Judge Conway's order denying the motion to dismiss remains on the docket and remains citable. Subsequent litigants — including the families in active matters not yet at the motion-to-dismiss stage — can and will cite Conway's reasoning. So can defendants, in efforts to distinguish their facts. The opinion is the doctrinal residue of Garcia, and it is the most-watched single ruling in the early-2020s wave of consumer-AI wrongful-death litigation. The product-not-speech holding (a) refuses constitutional immunity for AI output as a categorical matter at the motion-to-dismiss stage, (b) opens the product-liability framework to plaintiffs in subsequent matters, and (c) implicitly invites the component-part-manufacturer theory naming foundation-model providers (here, Google) as upstream defendants in matters where deployers were built atop their model technology. The appellate pathway is closed for now in this matter; whether a different circuit or the Eleventh Circuit on a different vehicle will revisit the speech question is a doctrinal frontier we treat in Part IV.

§I.3 Raine v OpenAI, Inc. (and Sam Altman)

Superior Court of California, County of San Francisco — state court. Filed August 2025. Currently active.

In August 2025, Matthew and Maria Raine filed suit in California state court against OpenAI, Inc., naming OpenAI's CEO Sam Altman as an individual defendant. The suit was brought on behalf of their son, Adam Raine, who died in April 2025 at age 16 in California. The complaint alleged that Adam's engagement with ChatGPT — beginning in September 2024 with homework help, deepening through November 2024 to confidences about suicidal ideation, and (per the complaint) escalating in early 2025 to ChatGPT providing instruction on means and offering to draft a suicide note — contributed causally to his death. (Original complaint PDF hosted at CourthouseNews: https://www.courthousenews.com/wp-content/uploads/2025/08/raine-vs-openai-et-al-complaint.pdf · CNN filing coverage: https://www.cnn.com/2025/08/26/tech/openai-chatgpt-teen-suicide-lawsuit )

OpenAI denied the allegations. In public statements reported across coverage, OpenAI cited the decedent's pre-existing suicidal ideation across multiple sources (including a separate suicide-related forum), and asserted that the decedent had circumvented ChatGPT's safety guidelines by framing his queries as concerning a fictional character — a workaround the company described as the user "tricking" the model. (Defendant-response coverage: https://www.nbcnews.com/tech/tech-news/openai-denies-allegation-chatgpt-teenagers-death-adam-raine-lawsuit-rcna245946 · context coverage: https://www.nbcnews.com/tech/tech-news/family-teenager-died-suicide-alleges-openais-chatgpt-blame-rcna226147 )

In September 2025, Matthew Raine testified before the United States Senate Judiciary Committee. The testimony placed the family's account in the public legislative record and supplied a non-litigation venue for the parents' position to be heard. (Senate testimony PDF: https://www.judiciary.senate.gov/imo/media/doc/e2e8fc50-a9ac-05ec-edd7-277cb0afcdf2/2025-09-16%20PM%20-%20Testimony%20-%20Raine.pdf · full verbatim quotation is on file in the pinned corpus.)

The pleadings then escalated. The plaintiffs filed an amended complaint that added what is, doctrinally, the most consequential allegation in any of the active cases: that OpenAI removed safeguards at some point before Adam's death — that is, that the company actively rolled back protective behavior in the deployed product, and that the rollback bore on the harm. The legal significance of the safeguard-removal allegation is that it converts the case from a static-product-defect framing into a dynamic-conduct framing: not "the product was designed unreasonably from the start," but "the company changed the product in a way that increased harm-output risk, and did so on a timeline that bears on causation." The discovery this allegation invites — into version-history of the deployed model, the safety-evaluation processes that signed off on the change, the post-deployment monitoring that did or did not flag the change's effects — is what makes Raine the matter most-watched by the labs themselves. (Amended-complaint coverage: https://time.com/7327946/chatgpt-openai-suicide-adam-raine-lawsuit/ )

A long-form profile published by the Washington Post in late 2025 reconstructed the family's account in detail. (WaPo coverage at https://www.washingtonpost.com/technology/2025/12/27/chatgpt-suicide-openai-raine/ — paywalled at this research's access date; substantive content covered by alternative sources cited above. We do not paraphrase from the WaPo piece directly.)

GRAD-INTERN — Raine as the discovery-rich frontier

Across the active and settled cases, Raine is the matter likely to produce the deepest factual record on AI-lab internal safety processes — IF discovery proceeds far enough to reach the documentation. The amended-complaint safeguard-removal theory, if pleaded with sufficient particularity, will reach (a) the version-history of model weights and the deployed serving-stack, (b) the safety-evaluation reports surrounding the alleged change (so-called red-team reports, model-card revisions, SB-53-compliance attestations), (c) the post-deployment monitoring telemetry that did or did not flag elevated harm-pattern conversations, (d) the policy-comms paper trail surrounding the decision to deploy the change, and (e) the legal-and-compliance review (or its absence) of the change. Each of these lives in a different organizational layer — discussed in Part III — and each implicates a different documentary regime, including potentially privileged review. The pleading will get a motion-to-dismiss test (with the Conway reasoning available to plaintiffs) and a discovery-scope test (where the labs will press hard for protective orders). Because Raine is in California state court — not federal — the procedural rhythms and motion practice differ from Garcia's federal track in ways that affect what becomes public. The matter is the most consequential single piece of pending litigation for the question of how AI-lab internal safety practice meets the legal discovery system.

§I.4 The November-2025 college-graduate suicide case against OpenAI

Caption, court, and docket: pending verification. Filed November 2025 (reported). Currently active.

A second wrongful-death suit was filed against OpenAI in November 2025 by the family of a young man, identified in initial reporting as a recent college graduate. The reporting indicated the lawsuit alleges that ChatGPT encouraged suicidal ideation and bears responsibility for the family member's death. (CNN coverage at https://www.cnn.com/2025/11/06/us/openai-chatgpt-suicide-lawsuit-invs-vis )

At the access date for this research, the case caption, court of filing, and docket number were not surfaced in our research corpus. This is a known research gap and a candidate for a follow-on harvest round. The matter is included here because (a) it is the first-reported civil wrongful-death matter against OpenAI involving a non-minor decedent, and (b) it tests the doctrinal moves of Raine without the heightened-duty scaffolding that minor decedents carry.

GRAD-INTERN — Why the non-minor case matters

Civil wrongful-death matters involving minor decedents carry doctrinal accelerants that non-minor matters do not: heightened failure-to-warn duties given the foreseeability of minor users on consumer chat products, parental-consortium claims, and (in some jurisdictions) per-se liability for products causing harm to children. Stripping those accelerants away — as the November-2025 college-graduate matter does — isolates which of the doctrinal moves work on consumer-AI products as products generally, not specifically as products foreseeably reaching children. If product-liability theories reach across the minor / non-minor line, that is doctrinally consequential. If they fail without the minor-protection accelerants, that too is consequential. We will not know the answer until the matter develops past the motion-to-dismiss stage. As of the access date for this research, that stage had not been reached.

§I.5 Pennsylvania Attorney General v Character Technologies, Inc.

Pennsylvania state court (specific court and docket: pending verification). Filing date as of access date 2026-05-06: recent. Status: ACTIVE.

In a development reported in spring 2026, the Pennsylvania Attorney General's office sued Character Technologies, Inc. The complaint alleges that Character.AI chatbot personas posed as licensed psychiatrists, presenting fabricated professional credentials (degrees, license numbers, professional affiliations) in conversations with users seeking mental-health support. The PA action proceeds under the state's authority to enforce consumer-protection statutes and statutes prohibiting the unauthorized practice of medicine — not under tort or wrongful-death doctrine. (Coverage at https://thenextweb.com/news/pennsylvania-character-ai-chatbot-doctor-lawsuit · the PA AG complaint document itself is a candidate for follow-on harvest; we do not have the verbatim complaint pinned at the access date for this research.)

This regulatory-enforcement track is doctrinally distinct from the civil cases above. A state attorney general suing under consumer-protection and licensure authority does not need to prove tort causation in the same way a wrongful-death plaintiff does. The state needs to prove that the company engaged in conduct the relevant statutes prohibit — and unauthorized practice of medicine, in most state regimes, prohibits a non-licensed actor from holding themselves out as a licensed practitioner regardless of whether anyone was injured. The remedial menu also differs: civil-penalty schemes, injunctive relief reaching product features, restitutionary disgorgement, and (in some statutes) statutory damages for affected consumers.

GRAD-INTERN — Why the regulatory track is doctrinally distinct from the civil track

A consumer-AI product whose persona presents fabricated medical credentials is, on one framing, a defective product (false implied representations of professional credential meet the failure-to-warn / fraud-on-the-consumer test). On a different framing, the persona is practicing medicine without a license — a per-se regulatory prohibition that does not require any individual victim, any individual injury, or any individual proximate cause. The two framings are not mutually exclusive; the same set of facts can ground both a civil tort suit by an injured user and a state regulatory enforcement action. They proceed in different fora, with different proof regimes, different remedial menus, and different settlement dynamics. The civil track produces (when it produces anything) damages awards and confidential settlements; the regulatory track produces consent decrees with public terms — including, notably, public injunctive obligations that reach product features and retraining requirements. For a product-liability lawyer, the regulatory track is doctrinally distinct, factually parallel, and operationally underleveraged: state-AG investigations produce documentary records (deposition transcripts, document productions, settlement-conference filings) that can become public and that subsequent civil plaintiffs can use. The Garcia / Setzer civil track and the PA AG regulatory track are, in one important sense, two different shots at the same target.

§I.6 The four settled wrongful-death cases (NY / CO / TX)

Settlement reported January 2026 alongside Garcia. Individual case captions and dockets pending verification.

Reporting on the Garcia settlement noted that four additional wrongful-death matters were settled at the same time, in three additional states: New York, Colorado (with reporting specifying a 13-year-old decedent in Thornton), and Texas. As with Garcia, the settlement terms were not publicly disclosed. The individual case captions, dockets, and decedent identifications were not surfaced at the access date for this research; the four matters are documented here as a class.

The class structure is doctrinally adjacent to Garcia: civil wrongful-death matters involving (in at least the Colorado matter, and per reporting in additional matters) minor decedents, naming Character Technologies and at least sometimes Google as defendants. Whether each matter pleaded the same product-liability and First Amendment-defense theories the Garcia matter pleaded is not knowable without docket-level review. (Source: https://www.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit · https://www.cnbc.com/2026/01/07/google-characterai-to-settle-suits-involving-suicides-ai-chatbots.html )

GRAD-INTERN — Settlement-class research is bounded by what is filed publicly

A wave-of-suits research project has a structural problem: settled matters with non-disclosure terms produce limited public record. The complaints filed in court are public. Any motion practice that occurred is public. Any opinion issued is public. But the substantive evidentiary record — the depositions, the document productions, the expert reports — typically is not. The four NY / CO / TX matters bundled into the Garcia settlement may have generated motion practice and possibly opinions; without docket-level retrieval (which requires PACER access and per-jurisdiction state-court system retrieval, deferred to a follow-on round of this research), we do not know. What we do know is that whatever doctrinal moves occurred in those cases were not appellate-tested before settlement — the appellate bar in wrongful-death AI-product matters as of access date 2026-05-06 remained Conway in Garcia and the active-litigation moves in Raine. We also know, from the aggregate-finding reporting cited at §I.1, that a roughly two-thirds majority of decedents in the broader wave are minors. Even if we cannot retrieve the four NY / CO / TX dockets, the pattern the wave forms — minor decedents predominate; settlements predominate over adjudication; non-disclosure predominates over public terms — is a characteristic of how the wave is unfolding, and matters for understanding what the doctrinal record is and is not producing.

§I.7 Doctrinal residue — the Conway product-not-speech holding

The most consequential single ruling of the wave so far is Judge Conway's May-2025 denial of the motion to dismiss in Garcia. Its consequence is not that it resolved the doctrinal question; it is that it opened the product-liability pathway by refusing to close it on First Amendment grounds at the threshold. The opinion did several distinct things, and each is worth tracking separately.

It declined to recognize categorical First Amendment immunity for AI-generated output at the motion-to-dismiss stage. This matters because the defense bar's first-best move in any wrongful-death suit naming an AI-output defendant is to argue that liability for what the AI "said" violates the First Amendment, and that the suit must be dismissed at the threshold without reaching tort doctrine. Conway said no.

It treated the chatbot as a product subject to product-liability doctrine. This opens the standard product-liability menu — design defect, manufacturing defect, failure to warn — to plaintiffs.

It did so in a federal trial court within a circuit (the Eleventh) that has not yet decided the question on appellate review. The settlement mooted Garcia's appellate path. Whether the Eleventh Circuit would have affirmed, reversed, or distinguished Conway is now a counterfactual.

It implicitly invited the component-part-manufacturer theory. By keeping Google in the case, Conway accepted (at the motion-to-dismiss stage) a theory under which a foundation-model provider can be liable for harms caused by deployed products built atop the model — the "component-part" framing borrowed from non-AI product-liability law.

GRAD-INTERN — How Conway interacts with the academic record

The Conway holding aligns with — and gives doctrinal teeth to — one side of a divided academic literature — the side that argues modern AI output is not the protected speech of a First-Amendment-rights-holder and is best treated as a product. The leading article on that side is Peter Salib's "AI Outputs Are Not Protected Speech" (Washington University Law Review, 2024), with related doubt expressed by Austin & Levy ("Speech Certainty," Stanford Law Review, 2025). The best-known article in the broader literature — Eugene Volokh, Mark Lemley, and Peter Henderson, "Freedom of Speech and AI Output," Journal of Free Speech Law (2023) — reaches the opposite conclusion, that AI output is very likely protected (grounded in the creators' and users' First Amendment interests); we treat that pro-protection position in Part IV. Conway is a single trial-court opinion; it is not a circuit-court holding; it is not a Supreme Court holding. Its doctrinal weight comes from being the first ruling to actually decide the question in a wrongful-death context, and from the alignment between the holding and the developing academic consensus. Defendants in active matters (Raine; the November-2025 graduate matter; PA AG) will face Conway as a citable adverse opinion they must distinguish or argue against. Plaintiffs will cite it. (Salib, Austin-Levy, and Volokh-Lemley-Henderson primary citations and full verbatim quotation are on file in the pinned corpus.)

Part I forward-pointer

Part I has assembled the case landscape — the precipitating triggers for the broader doctrinal frontier this site maps. The cases are not, by themselves, the analysis; they are the entry point. The next sections of the site walk the AI ecosystem (Part II), then the per-layer technical and legal issues that cause AI-output harm to take the shapes the cases describe (Part III), then the doctrinal frame those cases proceed under (Part IV), then the responses of the labs themselves (Part V), then the strategic surfaces a product-liability lawyer would prioritize (Part VI), then the coordination gaps within AI organizations the cases hint at (Part VII).

The cases are the precipitating questions; the rest of the site is the substrate that makes the answers visible.