Whether you can correct what a machine says about you
You asked an assistant about yourself, or you read the summary sitting above the results for your name, and it said something wrong. Take the boundary first: no operator publishes a way for you to correct a statement an AI makes about you.
That is a strange sentence to read now, and it is the documented position. Two assistant companies publish a personal data request route a private person can use, and I name and quote both below, because they are real and free. Neither publishes a decision time, an appeal, or what happens when the answer is no. For the AI summary above a search for your name — the one most people are looking at — the published action is a thumbs-down and a comment box.
This is the thinnest-documented corner of the subject and the one where selling is loudest, so the rule here is simple. Where an operator has published something, it is quoted. Where nobody has published anything, I say so rather than fill the space with something that sounds like knowledge.
Where the sentence about you comes from
Start with the documented part, because it is more than most people expect. The AI summaries in search are not a separate library with their own opinion of you. They are built by pulling pages from the same search index, using the same ranking systems, and linking to the pages that support the answer.
“To be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet.”
“There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary.”
— Google Search Central, “AI Features and Your Website,” last updated December 10, 2025, read August 15, 2026
It adds that no special markup is needed and that its generative features are rooted in its core ranking systems. The other assistants describe the same shape: each runs a crawler for its search feature and fetches pages on request.
So when a machine says something current about a private person, the likeliest explanation is not that it remembers you. It looked you up. The pages ranking for your name are the raw material it had, which makes the article the object and the machine downstream of it. Not comforting, but usable, and the only pipeline anyone has documented.
The operator has written down that it may be wrong
People arrive here with a particular fear: a machine has pronounced on them, and machines get treated as neutral. The operators say otherwise in their own help pages.
“AI responses may include mistakes.”
— Google Search Help, “Find information in faster & easier ways with AI Overviews in Google Search,” read August 15, 2026
“Because generative AI is experimental and a work in progress, it can and will make mistakes,” including that models “may make things up.”
— Google Search Help, “Learn about generative AI features,” read August 15, 2026
That is the operator, in writing, saying the answer about a person may be false and may be invented outright. If somebody has waved an AI answer at you as if it were a record, those sentences are worth having to hand.
What I will not do is take the next step. Whether a false machine-generated statement about a person is actionable, and against whom, is unsettled law. That is an attorney's work and I am not one. What follows is only what the operators publish about their own processes.
Not sure which of these applies to you? Send me the address of what you found and what you have already tried. I will tell you which category it falls into and whether there is a route worth using — including when the answer is that there is not one. How I help.
What the search operator publishes for a person, and what it does not
The feedback control. Under an AI answer sits a thumbs-down and a report a problem option with a category and a free text box. Nothing is published about what happens next: no case number, no decision, no timeframe, no appeal.
The likeness route. Real, narrow, and one of the few individual-facing AI complaint mechanisms any operator publishes. It covers images and video where your likeness has been generated or altered by another user through the operator's own generative tools. Per “Report Google AI policy violations that involve your likeness,” read August 15, 2026, you must be signed in with a personal account, be 18 or older, and be somewhere its selfie verification service operates. Audio is out of scope. If a complaint is substantiated it may block or restrict the account that generated the likeness. Another company's tools, intimate images not made with AI, and general content removal all route elsewhere. On what you will hear back:
“we cannot provide you with further updates on the status of our investigation”
— Google, on that same help page, read August 15, 2026
And the gap. For the ordinary case — the summary said something false about me, drawn from web pages — no correction or removal request is published at all. The thumbs-down is the action, and no fee changes that.
The two operators who will take a request from you
These exist, they are free, and almost nobody uses them.
The company behind ChatGPT publishes a personal data route. Through its privacy portal you make a request and choose to remove personal data from ChatGPT responses. The line worth reading twice:
“Whether or not you have a ChatGPT account, you can ask OpenAI to prevent information about you from appearing in ChatGPT responses.”
— OpenAI Help Center, “Right to be forgotten and personal data removal from ChatGPT,” read August 15, 2026
Requests may be made by the person or a legally authorized representative. Note the wording: information appearing in responses. That is not a claim to remove anything from a trained model. How often requests are granted, how long they take, and what happens when one is refused are not published.
Anthropic publishes a privacy rights process for requests relating to its training data, asking for your name, email, screenshots of the response, the prompts used and the URLs referenced. That list is worth having before you write to anyone.
Either way, an assistant's answer is not stable. Capture the exact prompt, the answer and the date. A description of what a chatbot said last Tuesday is not something anyone can act on.
The ones that publish nothing, and the ones I have not checked
This is where such a page usually goes quiet and implies coverage it does not have. The state of it on August 15, 2026:
- Microsoft Copilot. Its privacy FAQ documents controls for the person using it — conversation history, memories, opting out of training. There is no published process for someone else to ask that information about them be removed from its answers.
- Bing. Its guidance on reporting a concern routes everything through a general form and states that it limits removal of search results to a narrow set of circumstances.
- Perplexity. Its privacy notice documents access, deletion and correction rights and a data request form, framed around your data as a user. Whether it accepts a request from a non-user about what its answers say is not addressed, and I am not going to imply otherwise.
- Everything else. Other assistants were not surveyed for this page. If somebody tells you what one of them will do about you, ask to see the page that says so.
Read that list beside the previous section and the pattern running through this subject appears: where a statute stands behind a category there is a form, criteria and staff. Where there is not, there is a mailbox, a feedback button, or nothing.
Is it stuck in the model forever?
This is the question people actually ask, usually at midnight. The honest answer has three parts, and none is the one being sold.
Models can retain training text. Published security research has shown that training data, including personal details, can be pulled back out of a language model. That is peer-reviewed work, not a vendor claim. It does not establish that the finding carries over to the assistants in use now, which are larger, trained differently and wrapped in filters and lookup layers. Nobody can tell you an assistant has memorized the article about you.
Removing one person from a trained model is an open research problem. It has a name in the academic literature — machine unlearning — and is surveyed as an active area with contested results, not shipped as a feature. No operator surveyed claims to delete a person from a trained model on request, and Anthropic publishes the constraint directly:
“Models do not store text like a database, and they do not have access to or pull from the original training data once the models have been trained.”
— Anthropic Privacy Center, “Privacy rights requests relating to Anthropic's training data,” last updated July 8, 2026, read August 15, 2026
In practice the mechanism is lookup, not memory. Which gives the one genuinely useful conclusion here: if the source page comes down and stops being retrievable, the answer will generally change — not because anything was unlearned. How fast, and whether anything lingers in caches, nobody publishes. Both confident stories are unsupported: it is in the AI forever, and somebody can scrub you out of the model.
The small ask that targets the sentence itself
The levers here belong to whoever runs the page about you, not to you and not to me. They are documented, and worth knowing because they make a much smaller request possible.
A publisher can mark a page or a passage so its text may not be shown in a snippet, cap how much may be used, or take the page out of the index altogether. A separate control governs whether the site's content may be used to train future models, and the operator states it does not affect search inclusion or ranking. Where the line falls is stated plainly:
“AI is built into Search and integral to how Search functions, which is why robots.txt directives for Googlebot is the control for site owners.”
— Google Search Central, “AI Features and Your Website,” last updated December 10, 2025, read August 15, 2026
Translated: a site cannot stay in the search engine and opt out of its AI features. Leaving is the only opt-out.
Why it matters to you. When a small blog or a former employer will not delete a page, would you mark the paragraph about me so it cannot be quoted in a snippet is a far smaller thing to ask. It targets the exact text a summary would lift and leaves the article intact. Hardly anybody asks, because hardly anybody knows it is there.
What honest work looks like when the machine is the surface
Begin with what does not exist. The operator states there are no additional requirements and no special optimizations needed to appear in its AI features, no special markup, and that they run on the same ranking systems as ordinary results. A paid service to optimize you for AI is selling a lever its own documentation says is not a separate thing.
What is left is unglamorous and real. Capture what is being said, with the prompt and the date. Work out which live pages the statement is being assembled from, since the documented pipeline is retrieval from indexed pages. Check whether any of them crosses a written policy line and has a route of its own. File the personal data requests that are published, which cost nothing. And where the statement is an invention with no source page at all, say so and keep the operators' written admissions in the record.
Nobody can tell you what a model will say about you tomorrow. Anyone who does is describing a control no operator has published, on a surface where the honest answer to most questions is still: nobody has published this.
Frequently Asked Questions
ChatGPT said something about me that is not true. Can I get it fixed?
There is a published route, and it is free. Through OpenAI's privacy portal you can make a privacy request to remove personal data from ChatGPT responses, and its help article states that you can ask whether or not you have an account. Read the wording carefully: it is about information appearing in responses, not about deleting anything from a trained model. No decision time is published, and how often requests are granted or what happens when one is refused is not published either. Capture the exact prompt, the answer and the date first.The AI summary at the top of my name search is wrong. Who do I complain to?
There is no dedicated route, and that is the finding rather than an evasion. For a factual statement about a person in an AI answer, the search operator publishes no correction or removal request. What exists is the thumbs-down and a report-a-problem box with a category and free text, and nothing is published about what happens after you use it — no case number, no decision, no timeframe, no appeal. The operator does publish that its AI responses may include mistakes, which is worth citing to anyone treating the answer as a record.Is what it said about me stored inside the AI forever?
Nobody can honestly tell you that. Published research has shown that models can reproduce training text, including personal details, but whether that carries over to the assistants in use today is not established. Removing one person from a trained model is an open research problem, not a product, and no operator claims to do it on request. In practice, current statements about a private person usually come from the assistant looking pages up rather than remembering them, which points at the source page as the thing that matters.If the article comes down, will the AI stop repeating it?
Generally, yes, and for a plain reason: the documented pipeline for current information about a person is retrieval from live indexed pages, so a page that is gone and no longer retrievable stops being available to build an answer from. That is not the same as anything being unlearned. How quickly the change appears, and whether anything lingers in caching, is not published by any operator, so treat it as a direction rather than a schedule. It is also why removal at the source is still the highest-leverage action available.Someone made a fake AI picture of me. Is there anything published for that?
There is one narrow route worth knowing. The search operator publishes a complaint process for images and video where your likeness was generated or altered by another user with its own generative tools. You have to be signed in with a personal account, be 18 or older, and be somewhere its verification service operates. Audio is out of scope, and so is anything made with another company's tools. If a complaint is substantiated it may act against the account that generated the likeness, and it states it cannot give you updates on the investigation.A company offered to optimize how AI sees me. Is that a real service?
The operator's own documentation says the mechanism being sold is not a separate thing. It states there are no additional requirements and no special optimizations needed to appear in its AI features, that no special markup is required, and that those features run on the same ranking and quality systems as ordinary search results. So AI optimization as a distinct paid lever is being sold against published documentation. What genuinely changes the raw material an AI answer is built from is what ranks for your name, slowly, with nothing promised.Can I make an AI stop using a particular page about me?
Not directly, but the person who runs that page can, and it is a much smaller favor to ask than deletion. A publisher can mark a passage so its text may not be shown in a snippet, cap how much text may be used, or remove the page from the index entirely. A separate control governs whether a site's content may be used for training future models, and the operator says that one does not affect search inclusion. Asking a reluctant publisher for the small change sometimes works when asking for deletion did not.Published