Who Pays When AI Makes It Up? What Courts in China, Germany and Georgia Have Decided So Far

Lawyers keep asking the same question about generative AI errors: when a chatbot invents a crime, a campus address or a company scandal, who answers for it? Courts in China, Germany, the United States and Canada have now ruled on versions of that question, and the answers do not line up. A provider walked away in Georgia, lost a case over a search answer in Nanjing, was held to be the speaker of an AI summary in Munich, and was told by a Hangzhou court that a chatbot’s promise to pay 100,000 yuan meant nothing. The pattern only becomes visible when you stop reading each ruling as a verdict on “AI” and start reading it as a verdict on a specific decision someone made about how the output was built, labelled and published.

This guide walks through the rulings that matter most, explains where they point in different directions, and ends with the questions a company should be able to answer about its own AI features before a plaintiff asks them. Everything below describes court decisions and regulator positions as reported. Several of the German and Chinese rulings are interim or not yet final in every respect, and none of them is a binding rule outside its own jurisdiction.

1. Four questions that decided almost every case

Different legal systems, different causes of action, same underlying inquiry. Reading the decisions side by side, four questions do most of the work:

1. Whose statement is it? Did the provider or operator present the output as its own, or was it a machine response a user prompted and then chose to rely on or republish?
2. What would a reasonable reader take it to mean? A confident answer in a search box reads differently from a chat reply that arrives with repeated warnings.
3. What did the provider do to reduce the risk? Warnings, retrieval safeguards, complaint handling and speed of correction all show up in the judgments.
4. What harm actually followed? Several claims failed because nobody proved a loss, not because the output was acceptable.

Keep those four in mind. They explain why OpenAI won in Georgia and Google lost in Munich even though both involved a fabricated claim about a named person or company.

2. China: fault, not perfection, and users carry their share

The Hangzhou campus case

In June 2025 a user asked a generative AI app about university admissions. The system gave an incorrect campus address, then, when challenged, insisted it was right and said it would pay 100,000 yuan in compensation if it was wrong, inviting the user to sue in the Hangzhou Internet Court. The user did. According to a law firm analysis of the judgment, the court dismissed the claims in January 2026 and neither side appealed, so the decision is final.

Three holdings stand out. The AI’s promise did not bind the company, because an AI is not a legal subject and cannot make a declaration of intent for its operator. Generative AI was treated as a service rather than a product, which keeps strict product liability out of the picture and sends the dispute to ordinary fault-based tort rules. And the court sorted the provider’s duties into tiers: a heavy duty to prevent clearly illegal content, a duty to give clear and prominent warnings about functional limits (stronger in health, legal or safety settings), and a lighter duty of reasonable effort on accuracy, which is not a guarantee of correct answers. Public reporting on the case also says the plaintiff did not prove actual loss, and the presiding judge described inaccuracy as an unavoidable feature of how current models work while stressing that this does not excuse providers who skip their duties of care.

The Nanjing search-answer case

The Baidu case went the other way. In September 2024, a search for a Nanjing lawyer produced an “AI Smart Answer” stating that he had been sentenced to three years in prison for an explosion-related offense, next to a photograph of him in a lawyer’s robe. As reported in a case summary, the court distinguished between Baidu’s ordinary search suggestions and related-search terms, which it did not find infringing because they were generated from user behavior and Baidu had not edited them, and the AI-generated answer, which Baidu had processed and synthesized itself. Baidu argued that hallucination is technically unpredictable and that it had taken remedial steps afterward. The court acknowledged some unpredictability but still held that the answer clearly pointed to the lawyer and lowered his reputation. It ordered a written apology and refused his claims for financial loss and emotional distress for lack of evidence. The Nanjing Intermediate Court dismissed Baidu’s appeal in March 2026.

Read next to Hangzhou, the contrast is useful. A chatbot that gave a wrong but harmless answer, with warnings, and caused no proven loss, produced no liability. A search platform that displayed a false criminal conviction about an identifiable person did.

The Guangzhou user case

The third Chinese ruling concerns the person typing. In a case reported in Chinese court-affairs coverage, a finance commentator gave an AI tool a 1,700-word document of unverified allegations about a group company, told it to expand them into an article of at least 10,000 characters, and published the result as his “original” work without labelling it as AI-generated. The Guangzhou Internet Court found he had failed in a chain of duties: verifying the input, writing the prompt carefully, checking the output before publishing, labelling AI-generated content as the rules require, and exercising the extra care expected from someone presenting himself as a finance professional. It ordered an apology and 10,000 yuan in damages. The practical message is that “the AI wrote it” did not shield the publisher, and a prompt that asks for expansion of unverified material was treated as evidence of fault.

3. Georgia: the disclaimers worked, and so did the reporter’s skepticism

The best-known US ruling is Walters v. OpenAI. A reporter asked ChatGPT to summarize a real lawsuit, and after the system first said it could not access the link, it produced a false account accusing radio host Mark Walters of embezzlement. On May 19, 2025, Judge Tracie Cason of the Gwinnett County Superior Court granted summary judgment to OpenAI. Reading the court’s order, the reasoning has three parts. A reasonable reader in the reporter’s position could not have understood the output as stating actual facts, given ChatGPT’s own warnings, its refusal to open the link, its stated knowledge cutoff, repeated disclaimers in the terms of use, and the fact that the reporter quickly recognized the summary was wrong. There was no evidence that OpenAI had been negligent or acted with actual malice. And Walters conceded that he had suffered no actual damages.

That is a narrow ruling, decided by a state trial court, and it is highly fact-specific. The user who received the false statement never repeated it, and he knew it was false. A different plaintiff, with a different reader who believed the output and passed it on, would present a different case.

4. Germany: when the AI is the speaker

Hamburg and the Grok account

On September 23, 2025, the Hamburg Regional Court issued a preliminary injunction against xAI over a German-language reply from the public Grok account on X. The reply said the civil-society group Campact depended heavily on state funding, which the court treated as a false statement of fact. According to published summaries of the decision, the court held that the AI origin of the sentence changed nothing, that readers may even give an AI answer extra credibility, and that the operator had adopted the statement as its own by publishing it from the account. Because xAI made no showing that the statement was true, it was treated as untrue for the proceedings.

Munich and Google’s AI Overviews

The Munich I Regional Court went further on May 28, 2026 in a case brought by two publishers who found a Google AI Overview linking them to scams, subscription traps and other dubious practices. In a translated copy of the judgment, the court treats the overview as Google’s own statement, not a neutral display of third-party results, because the AI summarized and restructured the material in its own words, added a recommendation section and made claims that the linked sources did not contain. It rejected the argument that users can check the links, reasoning that a self-contained, confident statement with no sign of unreliability gives readers no reason to double-check. The court also held that the Digital Services Act’s liability privileges for hosting did not apply, and that the GDPR was irrelevant because the claimants were companies rather than natural persons. It granted an injunction on most of the challenged statements and refused a few others, including a claim about links to debt collection that the publishers had not shown to be false. Google has said it is reviewing the decision, which is not yet final.

5. A chatbot’s mistake is the company’s mistake

One older decision belongs in the same conversation. In 2024 a British Columbia tribunal ruled that Air Canada was responsible for incorrect bereavement-fare information given by its website chatbot, finding negligent misrepresentation and rejecting the idea that the chatbot was a separate entity or that the right answer elsewhere on the site excused the mistake. It is a consumer-contract case rather than a defamation case, but it points the same direction as Munich: if the company deploys the system to speak to customers, the company speaks.

6. What regulators have said

Data protection authorities approach the problem from accuracy rather than reputation. In its May 2024 ChatGPT taskforce report, the European Data Protection Board’s members expressed preliminary views that the GDPR’s data accuracy principle applies to a language model’s output, even though the model’s purpose is generating text rather than reporting facts. They noted that probabilistic training can produce made-up output that users are likely to treat as accurate, that providers must explain the limits of reliability, and that this explanation is not enough on its own to satisfy the accuracy principle. The report also says technical impossibility cannot be invoked to justify non-compliance. These are taskforce views while investigations were ongoing, not a final ruling, but they matter for any company processing personal data of people in the EU.

7. Side by side

8. What a business should be able to show

None of these courts demanded error-free output. What they looked for was evidence of decisions. If your company deploys a chatbot, publishes AI-assisted content or feeds model output into customer-facing products, the facts below are the ones a court or regulator is likely to ask about.

• The data flow. What the model receives, including personal data and uploaded documents, which vendor and model version process it, where output is displayed, who can see it, and how long prompts and responses are kept.
• Presentation. Whether the output appears as the company’s own answer or as clearly labelled model output, and whether warnings sit next to the answer or buried in terms of use. Munich and Georgia diverged on exactly this point.
• Named people and companies. Whether the system can generate statements about identifiable individuals or businesses, and if so what guardrails or retrieval checks apply.
• Complaint handling. A route for someone to report a false statement, a documented review, and a record of how quickly it was corrected. In Munich, the court noted that the publishers had notified Google before suing and that Google had not reviewed the statements even after that notice.
• Human review. Who checks AI-generated copy before publication, what they check, and where that review is logged. The Guangzhou decision made this a central question.
• Testing. Evidence that someone tried to break the feature with the kind of questions that produce invented claims, not just a completed risk assessment.

Captain Compliance’s artificial intelligence advisory work covers AI governance structures, transparency documentation and assessment of third-party AI tools, which is where these records usually start. Our template for building an internal AI governance committee is one place to begin. Where a false AI statement concerns an identifiable person, a request to correct or erase it can arrive through the same channel as any other privacy request, and a DSR portal can receive, route and document that request. Neither makes a model stop producing errors. They create the paper trail showing that someone received the complaint, made a decision and acted on it, and a closed ticket is not proof that the model has stopped repeating the claim. That still has to be tested on the live system.

The insurance side is moving too. We have looked at how insurers are repricing AI liability and at the argument that liability may be the most practical brake on runaway AI.

Frequently asked questions

Does a disclaimer protect an AI provider from defamation claims?

Not automatically. In Georgia, warnings were one reason the court found no reasonable reader would take the output as fact, but the reporter’s own knowledge and the lack of proof of negligence or damages also mattered. In Munich, the court emphasized that the AI Overview was a self-contained answer with no sign of unreliability. A disclaimer helps most when it appears where the reader sees the answer and when the surrounding facts support it.

Can a company be bound by something its chatbot promises?

The Hangzhou court said no for a hallucinated compensation promise, reasoning that an AI cannot make a declaration of intent for its operator. The Air Canada tribunal ruling, by contrast, held the airline responsible for misleading policy information its chatbot gave a customer. The difference is between a fanciful promise made after repeated challenges and a plausible answer on a core business matter that the customer relied on.

Are these rulings binding in the United States or the European Union?

No. The Chinese decisions bind no court outside China, and the German rulings are first-instance, one of them interim and not final. The Georgia order is a state trial court’s decision. They are best read as signals of reasoning, not as precedent in your jurisdiction.

Does the GDPR apply when AI invents facts about a person?

It can, if the output is personal data about an identifiable natural person and your organization is a controller for that processing. The European regulators’ taskforce views treat accuracy as applying to model output. It does not apply to claims about companies, which is why the Munich court decided under German civil law instead.

What should we do the first time someone says our AI said something false about them?

Preserve the prompt, the output, the model version and the display context, since a screenshot alone may not be enough. Assess whether the statement identifies a person or company, check it against reliable sources, correct or suppress it where you can, record who made the decision and when, and tell the complainant what you did. Then test whether the same prompt still produces the claim.

The catch in this body of law is that the fact separating the winners from the losers is rarely a technical one. In Georgia, the AI told the user it could not open the document and carried warnings the user plainly noticed. In Munich, the AI answered a question about two named publishers with confidence and no sign of doubt. The same kind of error produced opposite outcomes, and the difference came down to what the reader was told at the moment of reading.