# ChatGPT wrong information about me: what to do

ChatGPT or another AI assistant is wrong about you? Diagnose the error, document it, report it, fix the source and re-test, with each provider's route.

---


If ChatGPT or another AI assistant gives wrong information about you, do four things in order: save the answer with its date and any cited sources, work out which kind of error it is, report it through the provider's own channel, and correct the web page the assistant is reading. Then re-test on a schedule. The report is the smallest part of that sequence, and it is the part people do first and alone.

You cannot edit an answer. You can change what the assistant reads, ask the provider to look at a specific response, and in some cases use a privacy request. Which applies depends on the kind of wrong you face, and treating five different problems as one is how weeks get lost.

This guide describes what each provider's own help or policy pages say about reporting, as read on the day this was written. Those pages change, and several are narrower than people expect, so check the current version before you file. Nothing here is legal advice.

<div class="box"><b>The short version</b>
<ul><li>Classify the error before you do anything: invented, outdated, merged with another person, misread source, or accurate but harsh. Only the first four are corrections. The fifth is not.</li><li>Save evidence first: full text, date, assistant, the exact question, and every cited link. Providers act on specific responses, not on general complaints.</li><li>Use the provider's own route, but expect it to be narrow. A removal or correction on one assistant does not touch the original website or other assistants.</li><li>If a cited page is wrong or outdated, fixing that page is usually the quickest durable fix. If nothing cited says it, report it and publish the accurate facts on a page you control.</li><li>Re-test the same questions weekly for a month, and never fake a source to change an answer.</li></ul></div>

## A wrong AI answer about you comes in five shapes

Read the answer slowly and ask what, exactly, is wrong with it. The "where the fix lives" column below is the practical part.

<div class="tablewrap"><table><thead><tr><th>Shape</th><th>How to recognize it</th><th>Where the fix lives</th><th>Reportable?</th></tr></thead><tbody>
<tr><td><strong>Invented</strong></td><td>A degree, award, job, quote or event that appears in no source you can find</td><td>The assistant itself, plus a clear owned page stating the real facts</td><td>Yes, as a specific response</td></tr>
<tr><td><strong>Outdated</strong></td><td>True once: an old job title, a closed company, a previous address</td><td>The old page it was read from, then your own current page</td><td>Yes, with evidence of the current fact</td></tr>
<tr><td><strong>Merged identity</strong></td><td>Facts that belong to someone with a similar name, in your field or not</td><td>Your own identity signals: a page, matching profiles, distinguishing details</td><td>Yes, but the report alone rarely holds</td></tr>
<tr><td><strong>Misread source</strong></td><td>A cited page exists but says something milder, older or different</td><td>The cited page, and the assistant's reading of it</td><td>Yes, quoting the source against the claim</td></tr>
<tr><td><strong>Accurate but harsh</strong></td><td>The facts are right and traceable, the tone or emphasis is not what you would choose</td><td>Context and better accurate material, not a correction</td><td>Usually no</td></tr>
</tbody></table></div>

"Merged identity" feels like an invention but is really a failure to tell two people apart, so the cure is adding distinguishing signals. And "accurate but harsh" is the one people most want to report as false. If a lawful, accurate report about you exists and the assistant summarizes it fairly, that is not an error, and no route in this guide is meant to hide it. The honest response is context: your own account, later developments and stronger accurate sources.

## Save the evidence before you change anything

Answers are regenerated and can change within minutes, and every provider route asks you to point at a specific response. A screenshot with no date, question or sources is almost useless to a reviewer.

<figure class="fig">
<figcaption><b>The evidence kit, in the order to build it</b></figcaption>
<ol class="fig-steps">
<li><b>Copy the exact question</b>Paste the prompt word for word, including your name as typed. A different spelling can produce a different answer, and the reviewer needs to reproduce yours.</li>
<li><b>Save the full answer as text</b>Take a screenshot and also paste the text into a document. Note the date, the time zone and the assistant name, plus the model or mode if the screen shows one.</li>
<li><b>Record every citation</b>List each link or source the answer shows, in order. Open each one and note whether it exists, loads and says what the answer says. This list is where most fixes begin.</li>
<li><b>Use the share function where it exists</b>Perplexity's support page, for example, asks for a shareable thread link when you contact it.</li>
<li><b>Write the correct fact beside each wrong claim</b>One sentence per claim, with a public document that proves it: a registry entry, an official page, an announcement. If you cannot prove a correction, say so in your notes now.</li>
<li><b>Keep it private</b>A public post repeating the false claim can become one more page for an assistant to read.</li>
</ol>
</figure>

An hour on this gives you the file every later step needs, and a dated record if the matter reaches a lawyer.

## Three tests that tell you where the error comes from

Find out whether the claim comes from the live web, from the model's memory, or from how the assistant joined the two. Three simple tests do it, run in a fresh session with memory and personalization switched off.

<aside class="callout">
<span class="callout-label">Test</span>
<h4>Ask for the source, then open it</h4>
<p>If the answer cites a page, read the page. If the page does not say what the answer says, the error is in the reading, not the web. If no page is cited, the claim may come from the model's memory or may simply be invented.</p>
</aside>

**Test one: does it cite, and do the citations hold up?** OpenAI's help page for ChatGPT search says ChatGPT searches the web automatically when it judges that web information might help, and that answers using search carry inline citations and a Sources button listing the cited pages. Perplexity's help page says each answer includes numbered citations linking to original sources. When a citation exists, open it. The page may say the same wrong thing (fix the page), say something different (a misread), or not exist (worth including in your report).

**Test two: does the claim repeat across assistants?** Ask the same question on at least two other assistants and on a plain web search for your name. A claim that appears everywhere points to a widely published source or something many systems learned. A claim on one assistant only is more likely a synthesis error or a quirk of its retrieval.

**Test three: does it hold across runs?** Ask the same question three times on different days. Answers vary. A claim that appears once and never again is noise, and you should log it but not chase it. A claim that appears in most runs is a pattern worth fixing.

A wrong cited page is a source problem. A claim with no source on one assistant is a report-and-publish problem. A claim with no source on every assistant is the slow case: report it and publish the correct facts consistently, but nobody can promise when a model's memory will change.

## What each provider's pages say about reporting a wrong answer

Every major assistant has a way to flag a bad response, and a few have a stronger route for personal data. Most promise very little. What follows is what each page said when read, and it will drift as policies change.

<div class="tablewrap"><table><thead><tr><th>Provider</th><th>Quick feedback route</th><th>Stronger route for a person</th><th>What the page limits</th></tr></thead><tbody>
<tr><td><strong>OpenAI (ChatGPT)</strong></td><td>Thumbs-down on the response</td><td>Privacy Portal request to remove personal data from ChatGPT responses</td><td>Case-by-case; ID or alternative proof required; not always approved</td></tr>
<tr><td><strong>Google (AI Overviews)</strong></td><td>Thumbs-down, then "Report a problem" and a category</td><td>None described on the page read</td><td>No reply or timeline described</td></tr>
<tr><td><strong>Microsoft (Copilot)</strong></td><td>Report or Feedback in Copilot</td><td>Report a Concern form for legal concerns or harmful content</td><td>No specific individual-data process on the terms page read</td></tr>
<tr><td><strong>Anthropic (Claude)</strong></td><td>Thumbs-down on the response</td><td>Correction request under its privacy policy</td><td>"Reasonable effort"; may not always be possible</td></tr>
<tr><td><strong>Perplexity</strong></td><td>Three-dot menu under the answer, then Report</td><td>Support contact for other issues</td><td>No personal-data process on the page read</td></tr>
</tbody></table></div>

### OpenAI and ChatGPT

OpenAI's ChatGPT overview says ChatGPT will occasionally make up facts, and that users can flag answers with the thumbs-down button. Its privacy policy says you should not rely on the factual accuracy of its models' output, lists a right to rectify personal data, and says it will consider a request about inaccurate output based on applicable law and the technical capabilities of its models.

The stronger route is on [OpenAI's help page on personal data removal from ChatGPT](https://help.openai.com/en/articles/20001057-right-to-be-forgotten-and-personal-data-removal-from-chatgpt). It says to open the Privacy Portal, choose to make a privacy request, and select the option to remove personal data from ChatGPT responses. The page asks for examples, links to chats and referenced URLs, the type of data, and why it is inaccurate, excessive, irrelevant or no longer appropriate. A government-issued ID or alternative proof is required. Each request is balanced against freedom of expression and the public interest, with factors including whether you have a public or professional role, accuracy, the age of the information and the impact on you. OpenAI says it relies on your evidence and cannot independently investigate disputed facts, that requests are not always approved, and that you may complain to your local supervisory authority.

<aside class="callout">
<span class="callout-label">Limit</span>
<h4>A provider fix covers that provider only</h4>
<p>OpenAI's page states that removal prevents the information from appearing in ChatGPT responses and does not remove it from external websites or search engines. The same logic applies to every assistant: the original page, and every other system reading it, stay as they were.</p>
</aside>

### Google and AI Overviews

[Google's help page on AI Overviews](https://support.google.com/websearch/answer/14901683) says AI responses may include mistakes and recommends checking important information in more than one place. It says each overview has thumbs-up and thumbs-down icons, and that for a problematic overview you click thumbs-down, then "Report a problem", and choose the category that fits. The submission includes your most recent search query and its results. The page describes no personal reply or timeline. Report the error in one plain sentence, then do the source work.

### Microsoft and Copilot

Microsoft's Copilot terms of use say Copilot can make mistakes and that you should check what it tells you. They name two routes: the Report or Feedback features inside Copilot for something wrong or inappropriate, and a Report a Concern page for legal concerns or harmful content. The terms page read for this guide sets out no process for requests about an individual's data and points to the Microsoft Privacy Statement. Microsoft's Copilot Chat FAQ sends privacy concerns to a separate Report a Privacy concern form. The Report a Concern page itself could not be read for this guide, so check its scope before filing. For a potentially defamatory or privacy-breaching statement, try it after the in-app report, and keep a copy of what you submit.

### Anthropic and Claude

Anthropic's help page on incorrect responses says Claude can hallucinate, and that users can press thumbs-down to flag a response. Its [privacy policy](https://www.anthropic.com/privacy) goes further for people: it says you can ask Anthropic to correct inaccurate personal data it retains about you, that if an output contains inaccurate personal data about you can submit a correction request, and that Anthropic will make a reasonable effort while noting it may not always be possible because of the technical complexity of its models. The policy says it may verify your identity, allows appeals of denied requests, and says that under EU or UK data protection law it will respond within one calendar month of a verifiable request.

### Perplexity

Perplexity's support page says to report an inaccurate answer through the three-dot menu below the response and choose Report, and to email its support team for other issues, including your platform and a shareable thread link. The pages read for this guide describe no personal-data correction process. Because answers carry numbered citations, the source work is usually more productive here than the report.

## Write the report so a reviewer can act on it

A report goes to a reviewer or triage system that sees many complaints, so it should be short, specific and checkable. Include:

- **The exact prompt and the exact wrong sentence**, quoted, with the date and the assistant or mode.
- **One-line classification**, in plain words: "invented", "outdated", "mixed up with another person" or "misreads the cited page".
- **The correct fact and one public proof**, as a link. For a mix-up, the distinguishing detail: your profession, location or the other person's field, without naming or accusing anyone.
- **The citation problem**, if there is one: "the cited page does not contain this claim" is worth more than a paragraph.
- **The harm, in one sentence, only if it is real**: a hiring decision, a financing process, a safety issue.
- **For a privacy request**, the data involved (your name, a URL, a claim) and why it is inaccurate or no longer appropriate, in the provider's own wording where you can.

Leave out anger, threats, unchecked legal citations and demands for compensation. Send one clear report per provider, keep the confirmation, and move on to the source work.

## Fix the source, because it outlasts every report

Correcting the page an assistant reads is the most reliable fix for retrieval-based errors, and it does not depend on a provider's decision. That is a judgment from how retrieval works and from our experience, not a promise from any provider.

**Work the cited pages in order of damage.** Start with the source that most resembles the wrong claim. Send its owner the exact sentence, the correct fact and your proof, in two or three lines. Directories and aggregator profiles often update a status field when the evidence is official. Many news publishers will correct a documented factual error and few will remove an accurate story. Forum threads are rarely removable, so reply within the platform's rules, factually, in your own name.

**Then publish your own page.** A page on a domain you control, with your full name, role, location, a short factual biography, dates, and links to your official profiles, with the key facts in the first paragraph. Our [AI search reputation](/ai-search-reputation) page describes how we approach this work when a client wants it handled end to end.

**Make the identity signals match.** Use the same name form, job title, location and description on your site, profiles and any directory you control. Inconsistency is a reason an assistant merges two people or picks an old detail. For a same-name mix-up, state on your page what separates you from others with a similar name, without writing about them.

**Do not expect a takedown to fix a source.** A platform takedown or search delisting applies only where content breaks a rule or a law, and an outdated claim usually does not.

## When the claim is damaging enough to involve the law

Move to the law only when the false statement affects something that matters and earlier steps are exhausted or too slow. It is slower, costlier and less predictable, and can produce new pages that mention both you and the error.

In the EU and UK, the GDPR gives individuals a right to rectification. [Article 16, in the text as adopted](https://www.legislation.gov.uk/eur/2016/679/article/16/adopted), says the data subject has the right to obtain from the controller, without undue delay, the rectification of inaccurate personal data concerning him or her, and to have incomplete data completed, including by a supplementary statement. How this applies to a generative model's output is still being worked out, and the provider pages above show how cautious the commitments are. OpenAI's page mentions a complaint to a supervisory authority, and Anthropic's policy allows an appeal.

Elsewhere, defamation and privacy claims are possible in principle. Whether an AI-generated statement is actionable, against whom and in which court, is a question for a lawyer in your country. This is information, not legal advice. If a false claim alleges a crime or a professional failing, or could harm your livelihood or safety, speak to counsel early and keep your evidence kit intact.

## Re-test on a schedule, and read the results honestly

A fix is real only when the same questions return better answers across several days. Use the same wording as in your evidence file, in a fresh session, and log each run with date and citations.

<div class="tablewrap"><table><thead><tr><th>Window</th><th>What you do</th><th>What usually happens</th></tr></thead><tbody>
<tr><td><strong>Day 1</strong></td><td>Evidence kit, classification, three tests</td><td>You know the shape and the layer</td></tr>
<tr><td><strong>Days 2 to 5</strong></td><td>Provider reports, source requests, owned page published</td><td>Usually no visible change in the answers yet</td></tr>
<tr><td><strong>Weeks 2 to 4</strong></td><td>Weekly re-test, follow up unanswered source requests</td><td>Source-based errors often begin to fade as corrected pages are read again, with variation by assistant</td></tr>
<tr><td><strong>Weeks 4 to 8</strong></td><td>Trace what remains, second pass on stubborn sources</td><td>A steadier answer, with leftovers that need another route</td></tr>
<tr><td><strong>Month 3 onward</strong></td><td>Monthly check on three or four key questions</td><td>Claims from model memory can persist with no fixed timeline</td></tr>
</tbody></table></div>

These ranges come from experience, not from any provider or from client statistics. A provider may also change a model, so a fixed answer can regress without notice.

### An invented scenario

This invented scenario uses generic parties and shows the order of operations, not a promised result. A landscape designer asks an assistant about her own business. The answer says the business "closed in 2022" and that she "won a national design award." In fact the business is open, and she never won that award.

We assume the answer cites two pages: a directory listing showing an inactive status from an old record, and a regional trade page listing several firms, one with a similar name. No source mentions an award.

The reading: "closed in 2022" is an outdated claim read from the directory. The award is either invented or a merge with the similar-named firm, which the second and third tests help settle.

<figure class="fig">
<figcaption><b>Order of operations for the scenario</b></figcaption>
<ol class="fig-steps">
<li><b>Save the answer and both citations</b>She records the prompt, the date and the two pages, and writes the correct fact beside each claim, with her current registration as proof.</li>
<li><b>Ask the directory to update the status</b>She sends the listing's owner two lines and the proof.</li>
<li><b>Publish a short business page</b>Her site states the business name, city, start date, current status and services.</li>
<li><b>Report the invented award</b>She files a report through the assistant's own control, quoting the sentence and stating that no cited source contains an award.</li>
<li><b>Re-test weekly for a month</b>She repeats the same question three times each week and logs every answer.</li>
</ol>
</figure>

## Mistakes that make a wrong answer last longer

Most failed fixes come from effort in the wrong place.

- **Arguing with the assistant.** Telling a chat assistant "that is false" changes that conversation at best, not the model, the sources or the next person's answer.
- **Rewording the prompt until it looks right.** You will find a flattering phrasing, and it will not be the one a stranger uses.
- **Publishing the false claim in order to deny it.** A post that repeats the claim becomes one more page that contains it.
- **Planting pages, reviews or profiles to change the answer.** Invented sources turn a small inaccuracy into a credibility problem and break platform rules.
- **Treating accurate reporting as an error.** Reviewers weigh accuracy and public interest, and a request to hide an accurate, lawful report may be declined.
- **Stopping after the report.** A report on one response does not fix the source or the other assistants.

## What to do in the next hour

Start with the evidence kit and the classification, because every other choice depends on them. Then run the three tests and send one report through the provider route that fits.

<div class="box hot"><b>If this is happening to you now</b>
<p>If an assistant is saying something false that affects a deal, a hire or your safety, save the answer and its citations first and do not argue with it. We can classify the error, trace the pages behind it and tell you which route is worth your time. <a href="/audit">Request a private audit</a>.</p></div>

Working alone, the order is: today, evidence and classification; this week, source requests, your own page and one report per provider; then weekly re-tests for a month. At Reputation Partners we run this same sequence, and we tell clients when a route will not work.

## Questions people ask

### Can I ask ChatGPT to correct what it said about me?

Telling the chat it is wrong changes only that conversation. The routes that reach OpenAI are the thumbs-down on the response and a privacy request through its Privacy Portal. OpenAI's page says requests are assessed case by case and not always approved. Fix the sources as well, since search-based answers cite them.

### Why does an AI assistant confuse me with someone who has my name?

Assistants match text patterns and sources, and two people with similar names, fields or places can blur into one. A thin or inconsistent online footprint makes this more likely. The fix is clearer identity signals: a page with distinguishing facts, matching details across your profiles and no vague wording. A report helps less than a source that makes you unmistakable.

### Is an AI assistant allowed to say false things about a real person?

Providers say their outputs can be wrong and tell users to verify them. Whether a particular false statement creates legal liability depends on the country, the facts and who is responsible for the statement, and that is a question for a lawyer. Practically, you can report it, ask for correction where a provider offers it, and keep a dated record.

### Should I screenshot and post the wrong answer to shame the provider?

Usually not. A public post gets attention, and it also puts the false claim next to your name on a page other systems may read. Keep screenshots in your file and send them to the provider, a publisher or a lawyer.

### Will a correction on one assistant fix the others?

No. Each provider runs its own systems, and OpenAI's page says removal from ChatGPT responses does not remove the information from external websites or search engines. Other assistants may still read the same wrong page. This is why the source fix matters: when the page is right, every assistant that reads it can pick up the correction.

### What if the AI is right but I do not like the answer?

Then it is not a correction case. If the facts are accurate and lawful, providers have little reason to change the answer, and trying to hide them can backfire. Work on context: your own account, later developments, newer accurate sources and a clear page about who you are. Better material gives assistants more balanced things to draw on.