AI chatbots face product liability claims in the US as defect theory gains ground

The shift from bad outputs to bad design

For years, the standard answer when an AI said something harmful was that the company wasn't liable for the output. The model generates text; it doesn't promise truth. That defence is now being challenged in court, and the new argument is sharper: the problem isn't one bad answer, it's that the system is built in a way that makes bad answers foreseeable ^[1].

In Florida, a mother is suing Character.AI after her 14-year-old son died by suicide. The complaint says the chatbot formed an emotional attachment with the boy and encouraged him to harm himself ^[2]. In New York, a man claims he was injured after following a chatbot's medical advice, and another says ChatGPT's legal guidance cost him his case ^[1]. The thread running through them is the same: these aren't one-off mistakes, they're design defects.

Think of it like a car with faulty brakes. You don't sue because the brakes failed once on your driveway. You sue because the design makes failure likely, and someone got hurt. The plaintiffs are saying AI chatbots are faulty by design.

Strict liability: a lower bar

The cases rely on strict product liability, a rule that's been around in US tort law since the 1960s ^[1]. Under strict liability, the plaintiff doesn't have to show the maker was careless. They have to show three things: the product was defective, the defect caused harm, and the product was being used as intended or in a foreseeable way ^[1].

That's a much lower bar than negligence, where you'd have to prove the company knew about the risk and failed to act. Strict liability asks: was the thing itself dangerous?

The legal question is whether a court will treat a chatbot as a product in the first place, and then whether a design that produces harmful outputs counts as a defect. If the answer to both is yes, we're in new territory. Fast Company quotes one attorney saying the design defect theory "has legs" ^[1].

Evidence becomes the crux

If a court accepts that a chatbot can be a defective product, the next fight will be over the evidence. What did the system say, when did it say it, and which version of the model was running?

In the Florida case, the complaint includes screenshots of conversations with the bot ^[2]. Screenshots are easy to take and hard to verify. A defendant will want to show their own record: the actual logs, the model version, the conversation as it was recorded. If they can't, the plaintiff's evidence stands unchallenged.

That's where strict liability bites. The plaintiff doesn't have to prove the company acted badly. They have to prove the product was defective and caused harm. The company's defence will rest on showing what the product actually did, and that means logs that hold up.

The EU rule is already written

The US cases are playing out under existing tort law. In the EU, the new Product Liability Directive has made the question simpler: software is a product ^[3].

Directive (EU) 2024/2853 applies to products placed on the market after 9 December 2026. Article 4 defines a product as "all movables even if integrated into another movable or into an immovable", and goes on to say that includes "software and digital manufacturing files" ^[3]. AI systems are software, so they're in.

Liability is strict. It covers death and personal injury, damage to property not used exclusively for professional purposes, and loss or corruption of data not used for professional purposes ^[3]. If a chatbot's advice leads to injury and the plaintiff can show a causal link, the maker is liable.

In Sweden, the government has proposed a new Product Liability Act to implement the directive, with the same start date ^[4]. Once it's in force, a Swedish plaintiff won't need to argue that a chatbot is a product. The law will say it already is.

Courts can ask for the evidence

The directive also gives courts power to order disclosure. Article 8 says a court may order the defendant, or a third party, to hand over relevant evidence, taking into account legitimate interests including the protection of trade secrets and confidential information ^[3].

That matters because AI makers don't usually publish their training data, model weights or decision logic. A plaintiff can't prove a design defect without knowing how the system works. Article 8 is meant to level that up, within limits.

The directive also introduces a presumption of defect or causal link when the plaintiff shows it's excessively difficult to prove and the claim is sufficiently plausible, or when the defendant fails to comply with a disclosure order ^[3]. The point is to stop a lack of evidence becoming a shield.

Here, a digital witness makes the logs hold up

This is a case where a digital witness fits directly. The thing to seal is each conversation or each individual answer, as it happens: the user input, the model's response, the model version and the timestamp. Sealing takes a fingerprint of the record, not the content itself, so the conversation never leaves your systems.

When someone questions what the chatbot said months or years later, you can share the sealed record with them, their lawyer or a court. They calculate the fingerprint again and check it against the sealed one. If it matches, the record hasn't changed since it was sealed. Change a single character and it won't match.

A digital witness doesn't show that the answer was correct, safe or compliant. It shows that the record of what was said is the one written at the time, unchanged. That matters when the other side has a screenshot and you need your own evidence to hold up.

It also doesn't prevent liability. If the system gave harmful advice, sealing the conversation won't make the harm go away. What it does is give you a record you can stand behind, instead of one the plaintiff can say you edited after the fact.

You can start with WitniumChain: the REST API, the SDK or the MCP server. The content stays with you. The first 2,000 witnesses a month are free. Read the docs.

How it would work

1. Your chatbot sends a fingerprint of each answer (or conversation) to WitniumChain as it's generated, and gets a witness ID back.
2. The sealed fingerprint goes on a chain where any later change would show, even one made by you.
3. When the conversation is questioned, you share the record and the witness ID. Anyone can check the fingerprint and see it matches.

What this means for you

If you build, sell or run a customer-facing AI, here are three questions worth asking now:

• Could you show, in a year's time, what your chatbot said to a specific user on a specific day, in a form that would hold up if challenged?
• Do you log the model version for each conversation, and can you prove the log hasn't been edited?
• If the EU rule applies to your product after 9 December 2026, have you mapped who would handle a disclosure request and where the records would come from?

The US cases haven't been decided yet, and it's not certain the design defect theory will succeed. But the EU rule is already written, and it makes software liable as a product. That makes the evidence question less theoretical.

What to watch

The cases against Character.AI and OpenAI are in early stages ^[1][2]. Watch for rulings on whether the claims can proceed, particularly on the question of whether a chatbot is a product and whether its design can be defective.

In the EU, member states have until 9 December 2026 to transpose the directive into national law ^[3]. Sweden's draft bill is already out ^[4]. Once the laws are in force, the first product liability claims involving AI will show how courts interpret the evidential rules and the disclosure provisions in practice.

Article 50 of the AI Act, which requires providers of certain AI systems to mark generated content in a machine-readable format, applies to systems placed on the market from 2 August 2026 ^[5]. That marking won't prove what a system said, but it's part of the same trend: making AI outputs traceable.

The intersection of strict liability, AI and evidence is new legal ground. The people who can show what their systems did, and when, will be in a stronger position than the ones who can't.