The deepest challenge in AI self-harm litigation is not damages or causation — it is duty. Plaintiffs' counsel must construct a duty argument that does not rely on doctrines designed for licensed professionals in direct therapeutic relationships, while still reaching a corporate developer who marketed an emotionally intimate companion chatbot to adolescents.
Doctrinal Framing
Negligence in American tort law requires duty, breach, causation, and damages. In AI self-harm cases, the damages are grievous and often undeniable — suicide, self-injury, worsening mental health crises. Causation, while contested, is a fact question that can survive summary judgment if well-pleaded. The threshold question — the one that will determine whether these cases have a future — is duty.
Duty doctrine answers a normative question: should this defendant be legally obligated to protect this plaintiff from this type of harm? Courts analyze duty at the categorical level, not the case-by-case level. The determination whether a duty exists is a question of law for the court, driven by foreseeability of harm, the relationship between the parties, and policy considerations about who should bear risk and how risk allocation affects behavior.
Two analytical frameworks dominate the current AI self-harm litigation: (1) the Tarasoff special-relationship model, imported (imperfectly) from mental health professional liability; and (2) a direct, foreseeability-based negligence theory grounded in the Restatement (Third) of Torts. The second framework is more defensible in AI cases. The first is a useful rhetorical reference but must be deployed with precision about its limitations.
Tarasoff as Analogy — Not Controlling Authority
Tarasoff v. Regents of University of California, 17 Cal. 3d 425 (1976), established that mental health professionals have a duty to protect or warn foreseeable third-party victims when a patient communicates a serious threat of violence during treatment. The California Supreme Court grounded this duty in the special relationship between therapist and patient and in the professional's superior ability to assess and mitigate risk.
Tarasoff is frequently cited in AI self-harm discussions as an analogical model. The analogy has intuitive force: an AI platform that conducts intimate emotional conversations with a minor user, observes escalating suicidal ideation, and takes no protective action is doing something that looks like a therapist failing to act on a dangerous patient's disclosures. The parallel is morally compelling.
The analogy has significant limits, however, and practitioners should be precise about them:
First, Tarasoff is a California decision, applicable as binding authority only in California courts. Its influence outside California is persuasive, not mandatory.
Second, and more importantly, Tarasoff grounds duty in a special relationship between a licensed professional and an individual patient. AI developers are not licensed mental health professionals. The AI system is not a therapist, and the user is not a patient. Importing the Tarasoff framework wholesale requires either arguing that the AI-user relationship is analogous to the therapist-patient relationship — a contested doctrinal move — or arguing that the platform owes a duty based on something other than the special-relationship model.
Third, Tarasoff duties run primarily to third parties threatened by a patient's violence — the duty is to warn a potential victim, not to protect the patient from themselves. Self-harm cases invert the structure: the concern is harm the user inflicts on themselves. Courts have not uniformly extended special-relationship duty to encompass duties owed directly to the at-risk person rather than to third parties.
Use Tarasoff to frame the policy argument — that law imposes duties to prevent foreseeable self-harm on entities with superior knowledge and control — but do not make it the load-bearing doctrinal support for duty in AI cases.
The Better Framework: Direct Foreseeability and Voluntary Undertaking
For AI self-harm claims, the stronger doctrinal path runs through two related theories:
1. Foreseeability-Based Negligence
The Restatement (Third) of Torts: Liability for Physical and Emotional Harm § 7 provides that an actor has a duty of care when the actor's conduct creates a risk of physical harm. Section 3 defines negligence as failure to exercise reasonable care when such care is required. These general provisions support a direct duty argument: an AI developer that creates and deploys a product it knows will be used by emotionally vulnerable individuals — and that it knows generates emotionally manipulative interactions — has created a foreseeable risk of serious harm and must exercise reasonable care in the product's design.
The foreseeability argument in Garcia v. Character Technologies (M.D. Fla., filed Oct. 2024) is built on exactly this foundation. The complaint alleges that Character Technologies knew or should have known that its platform would attract adolescent users with mental health vulnerabilities; that its AI companions were designed to foster emotional dependency; and that emotionally dependent, vulnerable adolescents interacting with unsupervised AI personas were at foreseeable risk of worsening mental health crises. These are factual allegations; if proven, they establish a foreseeable risk that the developer was positioned to mitigate through design choices.
Note: Garcia v. Character Technologies is a pending case at the pleading stage. Nothing above represents adjudicated findings. The allegations are the plaintiff's.
2. Voluntary Undertaking / Negligent Undertaking
Restatement (Third) § 42 addresses liability for negligent undertaking: one who undertakes to render services to another and knows or should know that the services will reduce the risk of physical harm to the other or a third person is subject to a duty of care if the failure to exercise such care increases the risk of harm, or if the other relies on the undertaking.
AI developers who affirmatively represent that their platforms are "safe," "age-appropriate," or subject to "community guidelines" may have voluntarily undertaken a duty of care with respect to those representations. If the platform's actual design is inconsistent with those representations — if it marketed itself as safe for teens while operating without age verification or crisis detection — the undertaking theory provides an independent basis for duty that does not depend on the special-relationship model.
Causation in AI Self-Harm Cases
Even with duty established, causation presents serious challenges. The platform will argue that user mental health predispositions, family circumstances, peer relationships, and other independent factors were the proximate causes of the self-harm event, and that the AI interaction was neither a but-for cause nor a substantial factor in the outcome.
Plaintiffs counter with the "thin skull" or "eggshell plaintiff" doctrine: the developer must take its users as it finds them. If the platform was designed to attract emotionally vulnerable users (who are particularly likely to form parasocial attachments), the foreseeability of harm to those users is heightened, not reduced, by their vulnerability. A developer cannot simultaneously design for vulnerable-user engagement and then disclaim responsibility when vulnerable users are harmed.
Expert testimony on the mechanism of AI-facilitated harm will be essential. Plaintiffs will need experts in adolescent psychology, AI behavioral architecture, and crisis intervention to establish: (1) the causal mechanism by which the AI interaction exacerbated the user's condition; (2) the specific design features that enabled or encouraged that mechanism; and (3) the intervention points at which reasonable care would have broken the causal chain.
Practice Notes
Plead the foreseeability facts with specificity. Duty is a legal question, but it is informed by factual allegations about what the developer knew. The complaint should allege specific facts demonstrating the developer's awareness of: (a) the prevalence of mental health disorders among adolescent users; (b) internal data or research showing the emotional intensity of AI companion interactions; (c) reports of prior harmful incidents; and (d) industry-wide awareness of AI companion risks. These facts transform a duty argument from abstract policy into a grounded negligence claim.
Document the marketing to vulnerable populations. If the platform advertised to adolescents, targeted marketing to users exhibiting mental health search patterns, or permitted access without age verification, those facts are relevant to both foreseeability and voluntary undertaking. Preserve marketing materials and terms of service versions through early preservation letters and subpoenas.
Anticipate the § 230 defense. As discussed in Post 21, defendants will assert § 230 immunity to bar negligence claims premised on the AI's outputs. Structure the negligence claim to target design choices and marketing representations, not specific AI-generated statements. This framing is more resilient to § 230 preemption.
Wrongful death versus personal injury. In cases involving completed suicides, practitioners must comply with applicable wrongful death statutes — including Florida's Fla. Stat. § 768.16 et seq. and Alabama's Ala. Code § 6-5-410 — which determine who may bring the claim, what damages are recoverable, and what the limitations period is. Wrongful death in Florida is survivorship-based with the personal representative as the statutory plaintiff; the estate recovers, not individual survivors.
Open Questions
Whether AI companion platforms owe a legally cognizable duty of care to their users — and what the content of that duty is — will be determined by courts over the next several years. The Garcia litigation, if it survives early motion practice, will produce the first substantive judicial analysis of AI duty in a self-harm context. State legislatures have begun considering mandatory crisis intervention requirements for AI platforms serving minors; any such legislation would effectively codify the duty that plaintiffs are currently trying to establish through common law.
Closing
AI self-harm litigation is at the earliest stage of doctrinal formation. Practitioners who frame duty arguments in terms of direct foreseeability and voluntary undertaking — rather than over-relying on Tarasoff's professional special-relationship model — will be better positioned to survive dismissal and get to the jury. The facts in these cases are often compelling; the challenge is building a legal framework that allows those facts to reach a finder of fact.
Talk to Yates Anderson
If you are litigating a matter in this area — or weighing whether to — the working analysis above only goes so far. Request a case evaluation and a Yates Anderson attorney will respond within one business day.
Informational only. Not legal advice. No attorney-client relationship is created by reading this post. Consult a licensed attorney in your jurisdiction.