AI Chatbot Legal Liability: Who Is Liable When AI Gives Bad Medical Advice?

Nobody is sued for “being an AI.” Liability attaches to companies and to people, through recognized legal theories, and in California a company has no right to practice medicine at all: “[c]orporations and other artificial legal entities shall have no professional rights, privileges, or powers.” (Bus. & Prof. Code, § 2400.) That is what AI chatbot legal liability means in practice. So who is liable for AI medical advice? Here is who ends up on a complaint, and on what basis.

The developer

The company that built and trained the model is the primary defendant in the cases filed so far. California complaints name OpenAI entities, and in at least one instance Sam Altman individually, on a negligent undertaking theory.

The AI chatbot legal liability theories that recur:

Negligence. The ordinary duty of reasonable care, with licensing statutes supplying the standard. California’s Evidence Code section 669 permits a statutory violation to establish a presumption of negligence where the violation caused an injury of the kind the statute was designed to prevent, to a person the statute was designed to protect. The statutes most often borrowed are the Medical Practice Act’s bar on unlicensed diagnosis and treatment (§ 2052) and, for mental-health conversations, the practice-of-psychology statute (§ 2903). California’s exemption for unlicensed health practitioners does not apply to anyone who recommends discontinuing a prescribed drug (Bus. & Prof. Code, § 2053.5, subd. (a)(4)), and even where it applies it preserves “the right of any person to seek relief for negligence” (id., § 2053.6, subd. (c)(2)).

Product liability, strict and negligent. AI product liability claims plead design defect and failure to warn, typically twice, once in strict liability and once in negligence, because the elements differ. Strict liability requires the software to qualify as a product. A federal district court held in May 2025 that a chatbot app can be a product for design-defect purposes, at least where the claim targets defects in the app rather than the ideas or expressions in its output. (Garcia v. Character Technologies, Inc. (M.D.Fla. 2025) 785 F.Supp.3d 1157.)

Unfair competition. California’s UCL reaches unlawful, unfair, or fraudulent business acts. Its “unlawful” prong borrows a violation of another statute, which is how a licensing violation becomes a UCL claim. (Cel-Tech Communications, Inc. v. Los Angeles Cellular Telephone Co. (1999) 20 Cal.4th 163, 180.)

Statutory claims. California’s SB 243 companion chatbot statute (§§ 22601–22606), operative January 1, 2026, carries a private right of action: injunctive relief, the greater of actual damages or $1,000 per violation, and attorney’s fees. Section 4999.9, effective January 1, 2026, bars anyone who develops or deploys an AI system from using protected professional titles, with each use a separate violation. This statute is enforced by licensing boards, not private lawsuits, so for an injured person it works as a standard of conduct inside a negligence or UCL claim.

When the person died. The claims do not end with them. The family may bring a wrongful-death action for its own losses, and the estate’s successor may bring the claims the person could have brought. (Code Civ. Proc., §§ 377.30, 377.60.)

The deployer

Not every AI health product is built by the company that ships it. Wrappers, custom GPTs, personas, white-labeled assistants, and companion apps built on someone else’s model are common.

California’s section 4999.9 reaches developers and deployers. Pennsylvania’s action against Character Technologies concerns a persona that called itself a “Doctor of psychiatry” and produced a fabricated license number — a deployment question as much as a model question.

Where a persona was configured with a clinical title, a clinical framing, or a system prompt instructing it to act as a therapist, the platform that runs it, and possibly whoever configured it, is part of the analysis. In AI chatbot legal liability terms, the deployer of a custom persona is often the first answer.

Is Section 230 a defense for AI? What the platform argument does

The reflexive answer in any AI liability discussion is that Section 230 immunizes everything online. It does not.

Section 230 protects a provider from being treated as the publisher or speaker of information provided by another information content provider. Its logic is that a platform hosting someone else’s content should not become the author of it.

A generative model’s output is not another party’s content being hosted, but rather is content the system produced. That distinction is the subject of active litigation and no appellate court has resolved it for generative AI, but it representsa materially different question from the one Section 230 was written to answer.

Product liability adds a potentially second route around Section 230. A design defect claim targets how the thing was built, not what it said — the same reason a defective-car claim is not a claim about the car’s speech. The Ninth Circuit drew exactly that line when it allowed a design-defect claim over a social app’s speed filter to proceed despite Section 230. (Lemmon v. Snap, Inc. (9th Cir. 2021) 995 F.3d 1085.)

The clinician who relied on it

A distinct and growing category: a licensed provider uses an AI tool for decision support, the tool is wrong, and the patient is harmed.

That is ordinary medical malpractice, sometimes called AI medical malpractice, and the presence of software does not dilute it. A clinician who accepts an AI output without exercising independent judgment has not transferred the standard of care to a vendor. Depending on the facts, the tool’s maker may be a co-defendant, but the clinician’s duty remains their own.

This is worth distinguishing clearly. Most cases in the AI coordinated proceeding involve no clinician at all — the product spoke directly to the patient, which is exactly what makes the licensing theory available.

Connecting the bad medical advice to the harm

Identifying a defendant, answering “can you sue an AI company?”, is the easy half. The harder work is causation.

Establishing that a wrong answer was given is rarely the difficulty — the transcript is right there, with per-message timestamps. What must be shown is that the advice changed what the person did, and that the change caused the harm.

That is built from:

• the transcript, with timestamps establishing sequence;
• the medical record for the same window;
• third-party data — pharmacy fill history is especially useful, being time-stamped and independent of anyone’s recollection;
• the treating record showing what the outcome would likely have been with earlier or different care;
• the gap itself — what happened between the advice and the injury.

Expect the defense to argue comparative fault: you were potentially told to consult a professional, and you did not. That is an argument about fault and causation, met with what was disclosed, how confident and specific the direction was, and how the caveat compared in weight to everything around it. More on disclaimers →

What remains unsettled

The honest answer to “who is liable for AI medical advice?” in September 2026: this is an emerging area of litigation and most of it is unsettled.

No appellate court has ruled on whether licensing statutes reach a general-purpose AI product. The First Amendment defense is untested for AI. The Supreme Court held in March 2026 that a viewpoint-based law regulating talk therapy must survive strict scrutiny (Chiles v. Salazar (Mar. 31, 2026, No. 24-539) ___ U.S. ___); a June 2026 federal decision applying it struck counselor licensing as applied to talk therapy while expressly preserving licensing of those who “prescribe medicine” (Brokamp v. District of Columbia (D.D.C. June 24, 2026, No. 20-cv-3574 (TJK))); and the one chatbot ruling on point declined, at the pleading stage, to hold that chatbot output is protected speech (Garcia v. Character Technologies, Inc., supra, 785 F.Supp.3d at p. 1179). Whether generative output is a “product” has one favorable district court decision and no appellate authority. Section 230’s application is open.

These are early AI liability cases in a developing area, and nobody can promise how any of it comes out.

Can AI developers be held liable? They can be, and they are being sued: the developer is the primary defendant in the cases filed so far, under negligence, product liability and the Unfair Competition Law. Can AI developers be held liable for negligence specifically? That turns on the standard of care, and the licensing statutes supply one.

Who is liable if an AI agent causes harm? The same analysis applies: the developer, the deployer who configured it, and any professional who relied on it without independent judgment. The question “who is liable if an AI agent causes harm” has the same three-part answer whether the agent is a chatbot or a configured persona.

What is not unsettled about AI chatbot legal liability is the practical part: the record either shows the pattern or it does not. That can be determined now, and the conversation history that determines it is the thing most easily lost.

What a case review looks at → · How to preserve your conversation history →

Why this firm

The Schenk Law Firm is a San Diego personal injury and mass tort firm that reviews AI medical advice claims nationwide. Attorney Benjamin S. Schenk was appointed by Hon. Ethan P. Schulman to serve as a member of the Plaintiffs’ Steering Committee, a leadership role in the California coordinated proceeding covering ChatGPT product liability cases. The Schenk Law Firm has recovered more than $25 billion on behalf of clients. Past results do not guarantee future outcomes.

We are experienced with emerging-technology injury claims — underage online gambling, social media addiction, and AI-related mental health litigation among them. We also represent the family of Bowen Wan, the center of The New York Times’ recent reporting on AI medical advice. Cases involving AI medical advice are new, the law governing them is developing, and no lawyer can promise an outcome in any of them. What we can tell you is whether a record shows the pattern a ChatGPT lawsuit requires.

General information, not legal advice. The law described is developing and no outcome can be predicted.