2026-08-03 · Jack Stovell

ChatGPT says something wrong about you — how to fix it

When ChatGPT says something wrong about you, the error comes from one of two places: stale training data baked into the model's weights, or an outdated or ambiguous page it is reading live. You cannot edit the weights — that only changes with future training cycles. What you can do now is fix what it reads: publish one canonical, current, crawlable bio, and correct or retire the outdated pages it keeps citing.

Why does ChatGPT say something wrong about you?

Because it is either remembering or reading, and both can fail. Remembering means the fact came from training data — text scraped before a cutoff date, compressed into the model. If your job changed after that cutoff, the model's memory of you is frozen in the past. Reading means the assistant searched the web mid-answer and summarised the pages it could fetch; if the most fetchable page about you is a 2021 speaker bio, that is what you become.

Telling the two apart is useful because the fixes differ. If the answer cites sources or links to pages, it is reading — find those pages and fix them. If it states something confidently with browsing off and no sources, it is remembering, and the correction path is slower. What ChatGPT actually sees when it looks you up covers the retrieval side in detail.

Why is AI citing an outdated version of your bio?

Because the outdated version is the most readable, most stable page about you that retrieval can reach. Old bios tend to live on established domains as plain, static HTML: a former employer's team page, a conference site from three years ago, a directory entry you forgot exists. Your current bio, meanwhile, may sit on a JavaScript-heavy personal site most AI crawlers cannot render, or behind LinkedIn's login wall where they cannot go at all.

Retrieval does not weigh truth. It weighs what it can fetch, parse and trust. If the stale page is the legible one, the stale page wins — every time, until you change the inputs.

How do you remove outdated info from AI answers?

You cannot force deletion from a model's weights — no removal process reaches inside a trained model, and anyone claiming otherwise is selling something. What exists: vendor feedback channels (in-app report buttons, OpenAI's feedback forms, and formal privacy or data-correction requests in some jurisdictions). These are worth filing for serious errors, but for ordinary individuals they are slow, opaque and unreliable. Treat them as a secondary track, not the plan.

The reliable lever is owning the sources. Update or remove every outdated page you control. Ask the owners of pages you do not control to correct them — most sites will update a factual bio when the subject asks. Then give retrieval something better to prefer: a canonical, current page about you that any crawler can read without executing JavaScript.

The fix, in order

Find what it is reading. Ask ChatGPT (with browsing on) what it knows about you and note every page it cites or paraphrases. That list is your work queue.
Fix or retire the outdated pages you control. Update old bios you can edit; delete pages that should not exist. A page that genuinely returns 404 drops out of retrieval; a page that lingers keeps feeding the wrong answer.
Request corrections on pages you do not control. Former employers, event sites and directories rarely refuse a short, factual bio update from the person concerned.
Publish one canonical bio at a stable URL — current, plain HTML, crawlable, with your name, role, location and links to your real accounts. How to create an AI-readable profile of yourself walks through this; a hilyt profile (see the live example at i.hilyt.it/jack/llm.txt) is one free way to get it done.
Make the details agree everywhere. Same name form, same title, same employer, same location across your site, socials and directories. Consistency is how machines decide which version to believe.
Re-check in a few weeks. Retrieval-based answers update when the sources do — sometimes in days, usually within weeks.

What this won't fix

A wrong fact baked into training data will keep resurfacing when browsing is off, until a future model version trains on the corrected web. That timeline is not yours to control; the sources are. And being readable is necessary, not sufficient — clean sources remove the reasons an assistant gets you wrong, but retrieval, ranking and model behaviour still decide what any given answer says. You can test how readable you are right now with the free AI visibility checker.

Check your AI visibility — free