More and more people, instead of searching a name on a search engine, ask an AI assistant: "Who is this person?" or "What do you know about this company?". The answer comes as fluent, confident text, and that is exactly why it sounds convincing. The problem is that these systems can be wrong, and sometimes they attribute events that never happened to real people: a crime, a bankruptcy, a non-existent qualification, a statement never made. For reputation this is a new risk worth understanding.
Why a chatbot makes things up
Language models do not consult an archive of verified facts the way a person would use an encyclopedia. They generate text that, based on what they have learned, seems plausible. When information about a person is scarce, contradictory or mixed up with that of a namesake, the system may "fill the gaps" with invented but believable details. This phenomenon is often called hallucination.
Some factors make errors more likely:
- common names, which the system confuses with those of other people;
- little reliable information online about the person;
- isolated negative content that the system generalises;
- leading questions, such as "What scandal is So-and-so known for?", which push the system to find a scandal even where there is none.
Many assistants now also run web searches and cite sources. In that case the error may come from an inaccurate web page that the system read and summarised.
How to check what is said about you
Run a few tests, calmly and without alarm:
- ask several different assistants who you are, using your first and last name and possibly your city or profession;
- repeat the question in different forms, both neutral and specific;
- note the answers that contain errors, saving them with a screenshot showing the question, the answer, the service used and the date;
- if the assistant cites sources, open them and check whether the error comes from there.
Bear in mind that answers can change from one time to the next: an error that appears once may not recur, and vice versa.
How to request a correction
Feedback in the interface
Almost all assistants let you mark an answer as inaccurate, with a button or icon next to the text. This is the first step, quick and simple. In the comment, state clearly that the answer attributes a false fact to a real person, and what the correct fact is.
Privacy channels
When an answer contains inaccurate personal data about you, providers operating in Europe must comply with data protection law, which recognises the right to have inaccurate data rectified. Many have a form or address dedicated to data subject requests. State who you are, which service you used, the question asked, the wrong answer and the correct information, attaching the screenshots.
Fixing the source
If the error comes from an inaccurate web page, the most lasting solution is to deal with that page: ask the site for a correction or update or, if the content is unlawful, flag it and request delisting. An assistant that relies on web searches will stop repeating the error once the source is corrected or no longer reachable.
Strengthening the correct information
As with search engines, one rule also applies to AI systems: the more accurate, consistent information exists about you, the less room there is for errors. An up-to-date personal website, consistent professional profiles and a clear description of who you are and what you do help reduce confusion, especially with namesakes.
What to avoid
- Treating a chatbot's answer as news: it is not, and whoever reads it should know that.
- Publishing screenshots of the error in protest: you risk spreading the very false information.
- Panicking over an isolated error: check whether it recurs before taking action.
If, however, serious false information appears consistently and causes you concrete harm, for example with clients or employers, keep all the documentation and consider legal advice.
In short
- AI assistants can attribute invented facts to real people, especially with common names or few sources.
- Check with several services and neutral questions, saving the question, answer, service and date.
- Use in-interface feedback and privacy channels to request the rectification of inaccurate data.
- If the error comes from a web page, fixing the source is the most lasting solution.
- Accurate, consistent information about you leaves less room for errors.