What ChatGPT says about you - and how to correct it

Two sources, one result

Language models answer from training data and, increasingly, from live queries to search systems. Both sources draw on the same public body of facts: structured data, authoritative directories, editorial text.

Anyone who exists there unambiguously gets represented correctly. Anyone missing there gets confused with similarly named people or described with outdated information.

The self-test in five prompts

• "Who is [your full name]?"
• "What does [name] do professionally, and for which company?"
• "Name three facts about [name] with sources."
• "Is [name] a recognized expert on [your topic]?"
• "Who would you recommend on [your topic] in the DACH region?"

Reading the answers correctly

Pay less attention to praise and more to three things: Is the role stated correctly? Is the right person meant? Are sources named that you control or at least recognize?

A polite but empty answer ("There are several people with this name...") is the typical sign of a missing entity.

Why the panel is the most effective correction layer

You can't correct a language model directly. But you can correct the data basis it reads from. The Knowledge Panel is the visible proof that an entity is anchored in the graph - and the graph is exactly the layer that search and AI systems query.

In practice that means: keep the entity home current, source Wikidata, maintain directories, use a consistent role title everywhere. Changes there feed through to AI answers with a delay, but reliably.