Oberlandesgericht Hamm held a company responsible

You now carry the duty to verify AI output before it reaches a customer, client or filing, and to document which model, sources and human checks applied to each decision. Courts across three jurisdictions have held that deploying faulty AI statements does not shift responsibility to the model itself.

An analysis of 21 October 2026 of operator liability when AI chatbots make false claims argues that courts treat AI hallucinations as a governance problem, not as grounds for the operator to escape responsibility. The Oberlandesgericht Hamm ruled in May 2026 that a company remained liable for false specialist qualifications its chatbot had presented, rejecting the defence that the model acted independently. In our assessment, this signals that courts will hold organisations accountable for the factual claims their AI systems make in public, and that a model's probabilistic nature offers no exemption from that duty.

Which organisations face direct exposure?

Any organisation operating a public chatbot or AI system that makes factual claims about products, services, qualifications or legal matters now faces real liability for those claims. The Hamm court found the operator responsible precisely because it controlled which model ran, which data sources the chatbot consulted and which limitations applied. That control—the ability to shape the system's behaviour—is what makes the output an act of the operator. You cannot defend a false statement by saying the model generated it independently if you chose the model, the sources and the constraints it operates under.

The same principle extends to professionals who file or publish AI-generated content. A lawyer in Connecticut was sanctioned in August 2026 for submitting unverified AI citations to court. The Supreme People's Court of China issued a nationwide judicial opinion in September 2026 stating that AI-generated legal documents must be checked before filing, and that technology does not excuse responsibility. Across these cases, courts treat the human decision-maker—the operator, the professional, the signatory—as the party who remains accountable.

What are the failure modes you must guard against?

• Unverified factual claims — chatbot or AI system output presented to customers without human confirmation that the claims are accurate.
• Untraced sources — AI-generated content filed or published without the decision-maker having checked the original sources themselves.
• Undocumented control — no recorded evidence of which model version ran, which data it used, or who assessed the output before release.
• Scope creep — AI systems allowed to make claims beyond the factual domain they were designed for, without explicit limitation.
• No escalation path — no mechanism to flag, correct or withdraw false statements once they have been made public.

Which concrete controls must you be able to demonstrate?

1. Record the model and its purpose — document which model version each workflow uses, which source material it consults and the lawful basis for the data it touches.
2. Define and enforce factual boundaries — specify in advance which categories of claim the AI system may and may not make, and enforce those limits technically in the sources and system prompts.
3. Require human verification before release — establish an explicit step in which the responsible person checks every factual claim and source themselves, with evidence of that check recorded.
4. Maintain an audit trail of decisions — record who assessed the output, when, which version of the model was used and what sources were consulted, kept as evidence of control rather than as a mere activity log.
5. Plan for correction and withdrawal — document the process for identifying false statements after release and the mechanism for correcting or withdrawing them.

How does this change your incident response?

When an AI system makes a false statement, you now need to treat it as a governance failure, not a model failure. The Hamm ruling makes clear that the operator's control over the system is precisely what creates the operator's responsibility. This means you cannot simply log the error and move on. You must investigate which controls failed—whether the source was wrong, whether the verification step was skipped, whether the system was used outside its intended scope—and you must be able to show that investigation to a court if necessary.

The Chinese judicial opinion's emphasis on responsibility and evidence suggests that regulators and courts will expect you to have traced the decision chain. An activity log that merely records that something happened is not the same as evidence that a human assessed the output. What counts is demonstrable control: a record made in advance of which claims the system may make, evidence that a person checked the output before release, and a clear chain showing who decided what and when.

What can tooling do, and what remains your own judgment?

Automated logging, source tracking and workflow enforcement can all help you build the record. But no tool can replace the human judgment that the output is actually correct. A lawyer who relies on an AI system to generate citations must still read each citation themselves; a customer service team that deploys a chatbot must still decide which factual claims it may make. The courts have been clear: responsibility does not transfer to the model, and it does not transfer to the tool. It stays with you.