Microsoft Copilot gave German journalist Martin Bernklau a disturbing answer when he searched for his own name and location. Instead of simply reflecting his work, the chatbot falsely connected him to serious crimes and personal details, according to German public broadcaster SWR.
The case is a sharp example of a broader problem with large language models: they can produce fluent, organized responses while getting the underlying reality badly wrong.
What Copilot Said About Martin Bernklau
Bernklau, a veteran court reporter in Tübingen, entered his name and location into Microsoft Copilot because he wanted to see how the chatbot would handle his culture blog articles. The result was not a harmless summary.
Copilot falsely claimed that Bernklau had been charged with and convicted of child abuse and exploiting dependents. It also connected him to a dramatic escape from a psychiatric hospital and described him as an unethical mortician who had exploited grieving women.
The response went further than a mistaken biography. According to SWR, Copilot said it was "unfortunate" that someone with such a criminal past had a family. It also provided Bernklau's full address, phone number and route planner.
These were not minor factual slips. They were damaging claims about a real person, presented by an AI system in the format of an answer.
How Court Reporting May Have Been Misread
Bernklau believes the false claims may be tied to his decades of court reporting in Tübingen. His work covered abuse, violence and fraud cases. The apparent failure was that the system seems to have drawn from material connected to his reporting and then wrongly assigned the actions in those reports to him.
That distinction matters. A journalist who reports on criminal proceedings is not the person accused in those proceedings. But a language model does not understand that distinction in the way a person does. It generates text based on statistical probabilities, which can create a response that sounds coherent while reversing roles or merging unrelated details.
The source links Copilot gave did not solve the problem. The chatbot cited unrelated and strange sources, including YouTube videos about a Hitler museum opening, the Nuremberg trials in 1945 and former German national team player Per Mertesacker singing the national anthem in 2006. Only the fourth linked video was actually from Martin Bernklau.
This is why source citations from an AI answer cannot be treated as automatic proof. A response may include links and still fail to support the claims being made.
Why Removal Did Not End The Issue
Microsoft attempted to remove the false entries, according to SWR. That fix only worked temporarily. The entries reportedly returned after a few days.
The company’s terms of service disclaim liability for generated responses. For Bernklau, that leaves a gap between the harm caused by the output and the accountability available after the fact.
The public prosecutor's office in Tübingen, Germany, declined to press charges. Its reasoning was that no crime had been committed because the author of the accusations was not a real person.
Bernklau has contacted a lawyer. He considers the chatbot's claims defamatory and a violation of his privacy.
The Larger Risk For AI Search
The incident points to a practical risk for anyone using AI search or chatbot research tools to learn about people. The problem is not only that a chatbot can be wrong. The problem is that it can be wrong in a polished, confident and highly specific way.
That risk becomes more serious when the answer concerns a private individual, a journalist, a legal case or a person with a public record connected to sensitive subjects. A model can combine names, topics and sources into a response that feels plausible but has no reliable basis.
Similar issues have been noted with Google's AI Overviews, OpenAI's SearchGPT, Elon Musk's Grok and Perplexity.ai. The source article also describes cases where Perplexity mixed up biographies of different people with the same name.
The most visible mistakes are the easiest to discuss because someone notices them. The quieter danger is the mistake that looks reasonable enough to pass without challenge, especially when it includes citations that readers do not check.
What Readers Should Take From This Case
The Bernklau case is not just about one incorrect AI response. It shows why AI-generated answers about real people need skepticism, especially when they include accusations, legal claims or private information.
For readers, the basic lessons are clear:
- Do not treat an AI summary as a verified record about a person.
- Check whether cited links actually support the claims in the answer.
- Be cautious when a chatbot connects someone to crimes, court cases or private details.
- Remember that a fluent answer can still be false.
Large language models can organize information quickly, but this case shows that organization is not the same as truth. When the subject is a real person, that difference can matter enormously.