Yes. Courts in Germany, the US and China held that the organisation deploying a chatbot and the professional using AI output remain responsible for faulty answers, and that a model's probabilistic nature is no excuse. So set up demonstrable human verification, source limitation and incident correction before you release AI output.
On 12 May 2026 the Oberlandesgericht Hamm held that a company was responsible for incorrect statements made by its AI chatbot. That chatbot had presented non-existent specialist qualifications. The operator defended itself with the argument that the chatbot was an independent third party. The court rejected that: the answers counted as commercial acts of the operator, because it could influence the chatbot's operation and data basis. The case was brought by the Verbraucherzentrale Nordrhein-Westfalen.
That line does not stand alone. The Supreme People's Court of China published a nationwide opinion on AI disputes on 7 September 2026. The opinion addresses, among other things, the responsibility of providers and users and requires that AI-generated legal documents be checked before filing; technology does not constitute a general exemption from responsibility here. In the US, the District Court of Connecticut on 4 August 2026 imposed sanctions and a disciplinary referral on a lawyer who had filed unverified, AI-generated citations.
What exactly did the courts in Hamm, the US and China decide about AI hallucinations?
Our analysis of the German ruling, the American sanctions decision and the Chinese judicial opinion is that faulty AI output does not simply shift responsibility to the model. In Perez-Castillo v. Blanche, the Seventh Circuit addressed a lawyer's submission of fictitious AI-generated legal material, in that case generated by OpenAI's ChatGPT. In our analysis, the decision supports the importance of human verification before AI-generated legal content is filed. In Hamm, the operator remained responsible for misleading chatbot claims. In Connecticut, a lawyer was sanctioned because he had not verified AI output. The Supreme People's Court stated in its nationwide judicial opinion that providers and users may, in certain circumstances, be liable. Our reading is therefore that a model's probabilistic nature does not in itself constitute an automatic exemption from responsibility; the concrete consequences continue to depend on the applicable legal rule, the party's role and the circumstances.
Important for the interpretation: these cases do not create a universal liability standard. A court ruling, a sanctions decision and a policy guideline each have their own scope. The Chinese document is a guideline for judges, not a direct obligation for a Dutch organisation; for European parties such direct obligations follow rather from frameworks such as the EU AI Act (Regulation (EU) 2024/1689). In our analysis, the judicial sources together form a clear signal that courts may treat AI hallucinations as a governance and responsibility problem, not as a reason to place responsibility squarely on the model.
Why does a chatbot not count as an independent third party?
The Oberlandesgericht Hamm found that the operator had influence over how the chatbot worked and over which data it used. It is precisely that control which makes the answers acts of the operator. Whoever decides which model runs, which sources it consults and which limitations apply cannot hide behind the idea that the model speaks autonomously.
This touches on a reasoning that often crops up in practice: "the model did it, not us." In our analysis, the Hamm ruling makes it harder for operators of comparable public chatbots to attribute responsibility solely to the model; the precise consequences remain dependent on the facts and the applicable law. The connection with the other cases is that responsibility cannot simply be attributed to the model: with a customer service bot that may concern the operator, while with filed legal content the signing professional may be held to verification.
What do these rulings mean for directors, lawyers and CISOs?
Our analysis: because the Oberlandesgericht Hamm held the operator responsible for chatbot claims about qualifications, every organisation with a public chatbot runs a real risk of misleading commercial statements; the responsible director should therefore now have it recorded which factual claims the bot may and may not make, and enforce that limitation technically in the sources and system prompts. Because the District Court of Connecticut sanctioned a lawyer for unverified AI citations, this affects every professional who signs or goes public on the basis of AI output; therefore a firm or department should build in an explicit verification step before filing, in which the signatory demonstrably checks every source and citation themselves. In our analysis, partly in light of the Chinese judicial opinion’s attention to responsibility and evidence, it is advisable for the CISO to record for each AI workflow which model version ran, which source material was used and who assessed the output. This may create an evidentiary record useful for later review.
This line connects to the signing lawyer's duty to verify: in our analysis, as AI touches more sources, the need grows to test the output yourself rather than delegate it to the model.
Which controls must you demonstrably set up to limit liability?
In our analysis, the rulings operationalise AI governance in the form of source limitation, scope, human verification, ownership, warnings, escalation, correction and retention. Record the following per AI system, and keep the evidence of it:
- Source limitation: which information the chatbot was allowed to use and which factual claims were forbidden or sensitive, such as titles, qualifications or legal conclusions.
- Scope: for which tasks the system is intended and where it explicitly makes no statements.
- Human verification: how output was checked before publication or filing and who signed off on it.
- Ownership: who within the organisation is responsible for the operation, data basis and behaviour of the system.
- Warnings and escalation: which disclaimers the user saw and when a query was passed to a human.
- Correction and retention: how an identified error was withdrawn and rectified, and which evidence of those controls was preserved.
What counts is not a log file in itself, but evidence that control actually took place. An activity log that merely records that something happened is something other than evidence that a human assessed the output; on this see when audit trail software produces real evidence. Whoever records these decisions in advance builds defensible AI governance and responsibility instead of having to explain afterwards why the model said something. The same logic applies to automated summaries, where AI summaries of files require tracing back to the source before someone decides on them.
Sources and references
Sources: The article relies on rulings by the Oberlandesgericht Hamm, the District Court of Connecticut and the Seventh Circuit, and on a guideline from the Supreme People's Court of China.