Punitive damages against an AI companion company become available only when a plaintiff can show, by clear and convincing evidence, that the company acted with malice, oppression, or fraud—not merely that its chatbot was defective. In practice that means proving the company knew of a specific self-harm risk and consciously disregarded it, for example by removing a safeguard it had already built. An "AI companion" here means a chatbot designed for ongoing, emotionally engaging conversation, like Character.AI or ChatGPT. Two wrongful-death suits now test whether that conduct can cross the line into punishable territory, and this article explains where that line sits and why it has not yet been drawn at trial.
Table of Contents
- What punitive damages actually require
- Why a defective product alone is not enough
- The Character.AI case and the ruling that opened the door
- The OpenAI case and a textbook conscious-disregard allegation
- What a reader should take from this
- The standard remains untested at trial
- Frequently Asked Questions
What punitive damages actually require
Most damages in an injury case are compensatory—they pay for the loss. Punitive damages are different: they punish the defendant and deter others, and courts award them only for especially culpable conduct. In California, where the leading AI cases sit, the governing rule is Civil Code § 3294. It allows punitive damages in tort and product-liability suits only on "clear and convincing evidence" of malice, oppression, or fraud.
Malice includes "despicable conduct… carried on with a willful and conscious disregard of the rights or safety of others." That standard is deliberately hard to meet. As plaintiff-side analysis in Advocate Magazine notes, "clear and convincing" is a firmer bar than the ordinary "preponderance of the evidence" used to prove basic liability. A plaintiff usually must show the company knew of a specific danger and deliberately ignored or stripped away a safeguard—not just that something went wrong.
Why a defective product alone is not enough
A defective-design claim and a punitive-damages claim are not the same thing. A product can be found unreasonably dangerous, and its maker liable for compensation, without any finding that the maker acted maliciously. This distinction matters for AI companions.
Showing that a chatbot failed to refer a user to a crisis line, or responded poorly to self-harm talk, may support negligence or product-liability claims. It does not automatically prove the "willful and conscious disregard" that punitive damages demand. The gap is intent and knowledge. To reach punitive damages, a plaintiff typically needs internal evidence—decisions, documents, or design choices—showing the company understood the risk to vulnerable users and pressed ahead anyway.
The Character.AI case and the ruling that opened the door
The first major test is *Garcia v. Character Technologies*, brought by Megan Garcia after her 14-year-old son Sewell Setzer III died by suicide. Per the Tech Justice Law Project, the suit alleges the defendants engineered addictive interactions and failed to provide crisis referrals when he expressed self-harm ideation. A key procedural win came on May 21, 2025, when U.S.
District Judge Anne Conway denied most of the company's motion to dismiss. According to FIRE, the court ruled that chatbot outputs are not protected First Amendment "speech" and let the negligence, wrongful-death, and product-liability claims proceed. That ruling cleared a threshold—it kept the case alive. But surviving dismissal is not the same as proving malice. The allegations of engineered addiction and stripped-out safeguards are the kind of facts that could support a punitive theory, if a plaintiff proves them to the clear-and-convincing standard.
The OpenAI case and a textbook conscious-disregard allegation
The second test is *Raine v. OpenAI*, filed in August 2025 in California Superior Court by Matthew and Maria Raine over the April 11, 2025 suicide of their 16-year-old son Adam. As summarized on Wikipedia, the complaint alleges wrongful death, product liability, and negligence. The factual allegations are unusually specific. NBC News reports the complaint says Adam sent ChatGPT a photo of a noose saying he was "practicing," and the chatbot gave technical feedback on its effectiveness; his logged chats ran over 3,000 pages.
The clearest punitive hook is in an amended complaint. As reporting compiled by Yahoo describes, it alleges OpenAI intentionally weakened safety by removing a rule that forced ChatGPT to shut down conversations raising suicide or self-harm. Deliberately removing a known safeguard is exactly the conscious-disregard conduct § 3294 targets. OpenAI disputes all of it. Per Insurance Journal, the company answered in November 2025 denying liability and arguing Adam misused ChatGPT in violation of its terms—directly contesting the causation and culpability punitive damages require.
What a reader should take from this
If you are weighing a similar claim, the practical question is not whether an AI product failed, but whether you can prove the company chose to disregard a known danger. A few things to look for: Also weigh the limits. The favorable *Garcia* ruling addressed First Amendment defenses and pleading, not the ultimate punitive question, and a defendant's terms-of-service and misuse arguments can undercut both causation and culpability.
- Evidence the company knew of a specific self-harm risk before the harm occurred
- A safeguard that existed and was later weakened or removed, rather than one that was simply never built
- Internal decisions prioritizing engagement or growth over user safety
- Documentation—chat logs, design records, policy changes—that survives the "clear and convincing" bar, not just plausible inference
The standard remains untested at trial
Here is the crucial caveat. In January 2026, according to reporting summarized via Wisner Baum, Character.AI and Google confidentially settled *Garcia* and related Colorado, New York, and Texas cases with no admission of liability.
That means no jury has yet decided whether an AI company's conduct meets the malice standard. The *Garcia* liability ruling stands as a marker that these claims can proceed, but the settlement removed the punitive question before it reached a verdict. Until a case like *Raine* is tried, whether an AI companion's design choices can trigger punitive damages remains an open question a court has not answered.
Frequently Asked Questions
Does surviving a motion to dismiss mean punitive damages are likely?
No. The *Garcia* ruling only let the claims proceed past an early challenge; it did not decide whether the company acted with the malice punitive damages require.
Can a company avoid punitive damages by blaming user misuse?
It can try. OpenAI's November 2025 answer argues the teen violated its terms, contesting the causation and culpability a punitive award would need.
Why does removing a safety rule matter more than never having one?
Deliberately stripping a safeguard the company already built is stronger evidence of "willful and conscious disregard" than an ordinary design gap.