When AI Invents Your Credentials: Hallucinated Bios and What to Do

Somewhere out there, a machine is confidently describing a version of you who studied at a university you never attended and wrote a book you have never heard of. The unsettling part is not the error. It is the fluency, delivered with the same calm confidence as every true fact sitting right beside it.

Why does AI make things up about people?

Because a language model is not a filing cabinet, it is a prediction engine. It does not look you up; it continues a sentence about you in the most statistically plausible way. When the public record about you is deep and consistent, plausibility and truth coincide, and the model gets you right. When the record is thin, the model still owes an answer, so it reaches for what usually fits a person of your title, industry and era. Consultants of your sort tend to have MBAs, so you acquire one. Authors in your field cluster at certain publishers, so your imaginary book finds a home there. The fabrication is a portrait painted from the average of people like you, signed with your name.

Note what this implies: hallucination about people is inversely proportional to documentation. The model rarely invents credentials for someone whose every role, degree and byline is publicly and repeatedly recorded. It invents most freely in the silence between sparse facts. That inversion is the entire strategy for fixing it, and we will get there.

What does a hallucinated bio actually look like?

Having run this audit for many names, we can report that invented bios are not random noise. They come in recognisable species, and diagnosing which one you have determines the repair.

If you have already caught an engine mangling your story, the diagnostic detail matters, and we walked through a real case in what to do when AI gets facts wrong about you. And do not assume a flattering invention is harmless. The Promotion feels like free marketing until a client repeats your imaginary doctorate in an introduction, a journalist prints it, or a due-diligence check quietly flags the mismatch between what the machine says and what your record supports. Invented credentials cost you credibility in both directions: the false ones cannot be verified, and their presence casts doubt on the true ones sitting beside them.

Why do thin records invite invention?

Think of the model as a court sketch artist working from a witness description. Given rich testimony, the sketch resembles you. Given three vague sentences, the artist still produces a complete face, borrowing the average nose and the average jaw from every face they have drawn before. Every blank you leave in the public record is an invitation for an averaged feature. The professionals who suffer worst from hallucinated bios are almost never the heavily documented ones. They are the quietly excellent ones who let their websites lapse, kept their wins private, and assumed the absence of information would read as neutrality. To a prediction engine, absence reads as a canvas.

Can you get an AI to correct your bio?

Directly, sometimes. The major providers accept feedback on wrong answers, and their privacy channels handle formal requests about personal data, a route with real but bounded power that we examine in whether you can make AI forget you. But direct appeals share a weakness: they treat the symptom in one model, in one version, this month. The bio that ChatGPT hallucinated has cousins waiting in Gemini, Claude, Perplexity and whatever launches next quarter, all drawing on the same thin public record that caused the problem.

The reliable correction is upstream, because modern assistants increasingly retrieve before they answer. When an engine browses the live web and finds a clear, current, corroborated account of you, retrieval overrides the model's fuzzy memory. Fix the sources and you fix every engine that reads them, including the ones that do not exist yet.

The correction playbook, in seven steps

1. Audit before you assume. Ask ChatGPT, Gemini, Claude and Perplexity who you are, in fresh sessions, and save every answer verbatim. You are mapping the errors, not debating them. Our companion piece on Googling yourself versus asking ChatGPT gives the full routine.
2. Classify each error against the table above: Blend, Promotion, Fossil or Phantom. Each species points at its own cure.
3. Publish a canonical bio page. One page on a domain you control, stating name, role, real credentials, employers with dates, and genuine publications. Plain prose a machine can quote without interpretation.
4. Mark it up. Add Person structured data: name, jobTitle, alumniOf, worksFor and sameAs links to your real profiles. Google's structured data documentation covers implementation. This is how you state facts in the machine's own grammar.
5. Align every profile. Same name form, same title, same history across LinkedIn, directories, speaker pages and press mentions. Contradictions between your own profiles are the raw material of Blends and Fossils.
6. Report the specific falsehood through the provider's feedback channel, especially for defamatory inventions. It is the slowest lever, but for serious errors it belongs in the file.
7. Re-audit on a schedule. Monthly for a quarter, then quarterly. Corrections propagate as retrieval indexes refresh and models update, so expect improvement in waves rather than overnight; the timelines behave like the ones in how long until AI knows who you are.

One caution for step three and beyond: resist the urge to inflate while you are correcting. The same consistency machinery that spreads your fixes will happily spread an exaggeration, and engines cross-check against third-party sources. The record you want is boring, precise and verifiable everywhere.

Why arguing with the chatbot does not work

A tempting shortcut deserves a specific debunking, because nearly everyone tries it. You spot the error, you tell the model it is wrong, the model apologises gracefully and restates your correction, and you leave feeling the record has been set straight. It has not. You corrected one conversation. The model's weights are untouched, the retrieval sources are unchanged, and the next user who asks about you starts from the same flawed baseline, minus your correction. Unless memory features are involved, and those are scoped to your own account, the apology evaporates when the session ends.

Worse, the in-chat argument can mislead you about severity. Models are agreeable under challenge; they will concede errors they did not make and restate falsehoods with new confidence a session later. The only reliable test of whether a hallucination persists is the one you run cold: fresh session, no history, the same neutral question a stranger would ask. That is why the playbook above begins and ends with audits rather than arguments. You are not trying to win a debate with a language model, which is roughly as durable as winning an argument with the sea. You are trying to change the evidence the sea keeps drawing from.

How long until the corrections show up?

Faster than most people fear, for the retrieval half of the problem. Engines that browse can reflect a corrected canonical page within weeks of it being crawled and trusted. The parametric half, what the base model memorised in training, only shifts when models are retrained, so a Fossil may linger in offline answers for months after the live web has been put right. Treat the two on separate clocks: measure retrieval-based engines in weeks, parametric memory in model versions, and do not let the slow clock convince you the fast one is broken. There is also churn working in your favour: the sources cited in AI answers rotate constantly, with 40 to 60 percent changing month over month according to Semrush's AI Visibility Index, as reported by Similarweb. The answer about you is rebuilt regularly. Your job is to make sure the rebuild keeps finding better bricks.

A worked example: the invented doctorate

Here's a simple, hypothetical case to make the mechanics concrete. Someone works as an independent management consultant, has a genuine bachelor's degree, and has never claimed anything further. Their public footprint is thin, a sparse LinkedIn profile, no personal website, a couple of old conference mentions. Asked to describe this person's background, an AI assistant produces a confident, fluent bio that includes a doctorate from a well-regarded university. Nobody fed the model this specific lie, no malicious actor invented it. The model simply noticed that consultants with a similar profile, title, industry, years of experience, very often hold advanced degrees, and it filled the gap with the statistically likely version rather than an honest "unknown." The fix isn't a furious correction email, it's publishing one clear, accurate page stating the real degree plainly, so the next retrieval has something true to find instead of a gap to fill. This is a made-up example, but the shape, thin record, plausible invention, honest fix, repeats constantly across real names.

The real insurance is being too documented to misquote

Here is the reframe worth keeping. A hallucinated bio is not primarily an AI failure; it is a documentation failure that AI made visible. The professionals who never worry about invented credentials are the ones whose real credentials are stated identically in twenty places the machines trust. Building that record, canonical pages, structured identity, third-party corroboration, is the everyday work of people engine optimization, and it doubles as reputation insurance. If you would rather not build it alone, that is the work we do. Either way, do the audit this week. The machine is already answering questions about you. The only question is whether it is working from your evidence or its imagination.