Language models generate text based on statistical probabilities. This led to serious false accusations against a veteran court reporter by Microsoft's Copilot.

German journalist Martin Bernklau typed his name and location into Microsoft’s Copilot to see how his culture blog articles would be picked up by the chatbot, according to German public broadcaster SWR.

The answers shocked Bernklau. Copilot falsely claimed Bernklau had been charged with and convicted of child abuse and exploiting dependents. It also claimed that he had been involved in a dramatic escape from a psychiatric hospital and had exploited grieving women as an unethical mortician.

…

Bernklau believes the false claims may stem from his decades of court reporting in Tübingen on abuse, violence, and fraud cases. The AI seems to have combined this online information and mistakenly cast the journalist as a perpetrator.

Microsoft attempted to remove the false entries but only succeeded temporarily. They reappeared after a few days, SWR reports. The company’s terms of service disclaim liability for generated responses.

…

    • oce 🐆
      62·
      3 months ago

      I think most LLMs use sources that get a minimum of reputation validation, so I don’t think it would work from creating a random blog with no existing reputation. You’d need to contaminate a source that already has a reputation. For example, by buying a news source and orienting it.