How do AI models decide what to say about a person?

AI models don't look you up the same way Google does. They assemble an answer about your name from three layers, in this order: what they absorbed during training, what they retrieve from the open web at the moment of search, and how confidently the public record connects your name to a single, consistent identity. Where those layers agree, the answer is consistent. Where there is variability, the model fills the gap with the most statistically likely version of someone with your name, your industry, and your city. That variability is where most people get an incomplete or incorrect description of themselves.

Layer one: training data

Every large model is trained on a fixed snapshot of data. That snapshot includes news, court records, company pages, directories, forum threads, and archived copies of webpages that may no longer exist in search. Once trained, the model holds an impression of digital artifacts that relate to a string of letters which, when typed out, equate to your name. This is not a file of you as an individual. The impression is weighted by repetition and by source authority. So if a name appears on one profile page and fifty aggregator listings, the description from the aggregator listings takes precedence.

Two consequences follow from that. Nothing published after the snapshot exists to the base model. And anything published before it persists in the model's weights even if the page comes down later. Deletion removes the page from the web, but it doesn't remove the pattern from the model.

Layer two: retrieval at the moment of search

ChatGPT, Claude, Perplexity, Gemini, and Google's AI Mode add live search to the base model. When someone asks about you, the system runs a query on your name, pulls the top results, and writes an answer from what it fetched in real time. This is why the answer changes between sessions and between tools. It's also why the first page of search results still governs the AI answer, since retrieval draws on the same kind of ranked index.

Machine retrieval also favors efficiency, preferring pages that are easy to parse. A clear heading, a direct first sentence, a date, a stated role, and a named organization outperform a beautifully designed page that says nothing specific. Court records and news archives are easy for AI assistants to parse, which is why they surface in AI answers with disproportionate frequency. They don't necessarily hold a better account of you. They're simply the most machine-readable documents about you.

Layer three: entity resolution

Before a model can say anything about you, it has to decide which "you" the question refers to. It does this by matching the name against corroborating signals: a consistent role, a consistent organization, a consistent location, links between your official pages and third-party pages, structured data such as Schema.org Person markup, and knowledge graph entries where they exist.

When the signals agree, the model treats you as one entity and answers with confidence. When they conflict, or when there is a more prominent person with your name, the model either merges the two of you or defaults to the one with the stronger record. Name collision is the most common cause of a confidently wrong AI answer, and it's invisible to the person it happens to unless you look yourself up or someone tells you. Retrieval fetches the sources, but reconciliation decides which of them wins.

What the model does with gaps

A model rarely says "I don't know" about a person. It hedges its response by generalizing. My own search results kept me at an employer I had left, in a city I no longer lived in, doing work that had stopped being current. Nothing in that answer was false. It was the version with the most weight behind it, and it stayed the version every system repeated until the record was corrected. The averaging out of your public record is the AI model's protection against being too biased on one source. But don't mistake absence for neutrality. Even if your record is not well-documented, the model will offer its opinion about you, fair or not. The result is plausible, generic, and statistically average. An average is always smaller than the things it averages.

What can and cannot be changed

I can shape layers two and three. The pages that exist about you, how clearly they're written, how consistently they describe you, how they link to one another, and whether the structured signals resolve to a single identity. These are all within reach, and can meaningfully change retrieval within weeks and entity resolution within months.

I cannot edit layer one. No one can. A model's training data is fixed until the next model is trained, and the next model is trained on the web as it stands at that time. The only lever on training data is to make the record accurate and consistent now, so the next snapshot learns the corrected version.

I also cannot remove accurate public documents from the record. There is always a breadcrumb no matter what anyone says. A lawsuit that exists will continue to exist. What changes is whether it's the first and only thing the model knows about you, or just one line inside a fuller and better-sourced account.

What this means if your name is about to be researched

Ask ChatGPT, Perplexity, and Google AI Mode the same question about yourself: "Who is [your name]?" Compare the three answers to one another and to the truth. Note where they agree, where they disagree, and what they leave out. The disagreements show weak entity resolution. The omissions show where the record is thin. The errors show what the models filled in for you.

That comparison is the starting point for every review I complete.