How to fix what AI says about me?

There's no edit button to fix what an AI model says about you. But you can change the sources that it reads. Wrong answers about a person usually derive from one of four places: the model's training memory, an outdated or wrong page it retrieved, a person with the same or a similar name whose details graft onto yours, or the absence of any clear source. Each has a different fix. And if the answer is true, there is no fix at all.

1. Find out the origin of the wrong answer

The first step before fixing an AI answer is finding out where it came from. The same wrong answer can have four different causes, and the fix for one does nothing for the others.

Ask the same question across ChatGPT, Gemini, Claude, Perplexity, and Google AI Mode. There are other models to test, sure, but these will give you a good baseline. Note whether each one searched the web and which pages it cited. If no sources are cited (no citation chip after a paragraph), the answer probably came from training memory. If sources are shown, click on each one. My own research data shows cited sources that failed to load, links that led to an index page rather than a content page, and sources that said something entirely different from what the answer claimed.

How an AI model decides what to say about you from these sources is a different question, answered here.

2. When the wrong answer comes from the model's memory

If you ask an AI assistant a question and it answers right away without searching, it is reciting information it absorbed during training, and no one outside the company can update that. Each AI company ships updated versions on its own schedule, and each version carries a newer snapshot of the internet. Users don't control that schedule, so the practical remedy is indirect: make sure the sources you control are accurate and findable when the next snapshot is taken, particularly the ones a search-based answer would reference when looking for you.

In my study, The Machine's Version, the no-search condition was far less accurate than the search conditions and produced the majority of the fabricated claims. On 11 of my subjects, ChatGPT searched the web anyway, against explicit instructions not to. Users cannot switch that behavior off, even by turning off web search in settings. OpenAI has not said why it happens. It could be the system prompt, the hidden instruction set OpenAI attaches to every conversation. It could be a routing error, or something else entirely.

3. When the wrong answer comes from a page the model found

If the model cites an actual source, then you can improve the AI answer by fixing that source. Once you correct the page the AI is pulling its information from, the answer usually follows when the page is recrawled. This is easiest to do with the pages you control directly. In my own research, I've seen a company's own director page give a current title in its first sentence, followed by a superseded title three sentences later. Ten of 24 AI answers about that person repeated the old title. The same tool answered with the wrong title for one question, but when asked a different question a few minutes later, it answered with the correct title because it pulled from a different page on the same website. Simple consistency errors like these are not the AI's fault and are easily corrected by fixing the source.

Your pages also need to be findable. ChatGPT's web search draws heavily on Bing's index, so a page Bing has not indexed is much harder for ChatGPT to find. Check your main pages in Bing Webmaster Tools.

Third-party sources take more time and effort, but they are just as important for correcting incorrect and potentially damaging information. You cannot edit another publisher's page. You can ask each publisher for a correction, one at a time, and some will not respond. A correction at the original source doesn't travel to syndicated copies on other outlets. In one case I documented, a wire service corrected its story, and four smaller outlets still carried the original wording years later. Two of them ran it under a local staff byline with no wire credit, so one corrected error now reads as four independent reports.

Lastly, recent changes lag. Answers can take weeks to update, depending on when the page is recrawled. Sometimes a change that happened months ago still doesn't show up in AI answers. For one of my subjects, a leadership change from eight months earlier appeared in zero of 24 answers. When that happens, publish a new, dated page that states the change, rather than relying on an update to an existing source.

4. When AI confuses you with someone else

When two people share a name, AI models can either merge them into one person with a shared record, or treat them as separate people but attribute the wrong facts. Both errors can sound equally confident to a reader. In The Machine's Version, two business leaders each had a sibling with the same last name. In one case, the models wrongly attributed facts to the sibling instead of the subject. In the other, the two siblings' records merged into one person. Opposite failures from the same problem.

To fix this, you have to keep your name immutable. It needs to be the same form on every source you control, with no nicknames or variations. You have to pick one version of your name and own it completely. A canonical page that your other profiles link to helps here, as does keeping identifiers like company, city, role, and dates consistent. The simpler it is for an AI model to match you to your name, the less likely it is to confuse you with someone else.

5. When the answer is accurate but damaging

If an AI answer is true, no legitimate method will make it disappear. But you can change whether it arrives with the proper context. Accuracy and fairness are separate problems. In my research, one AI tool scored perfectly on accuracy across six answers about the same person and still ranked among the worst on tone, because it framed true facts in the least favorable way. For another person, four tools gave four different readings of the same facts within 21 minutes.

Context comes from sources. If the only page describing an event is the original report, the model repeats that report. If a dated, factual account of the outcome also exists (a settlement, a dismissal, a correction, what happened next), the model has a second source to draw on. Write that account plainly, without spin. A defensive page gives the model nothing new to quote.

Reviews and Reddit posts work differently. AI models act as aggregators of opinion, so one damaging review or Reddit post will not undo a wider consensus of positive experience. But a pattern of similar complaints will end up weighing the scales against you, and the model has less reason to hedge in defense of your reputation. Posting more positive content does not fix a pattern. You need to fix the thing people are actually complaining about.

6. The formal route

OpenAI accepts requests to keep personal information about you out of ChatGPT answers. The request removes the answer but not the source.

The request is called "Remove my personal data from ChatGPT responses," and you submit it through OpenAI's Privacy Portal, with or without a ChatGPT account. It derives from privacy laws that give individuals rights over their personal data, the best known being the EU's General Data Protection Regulation (GDPR). You state why the information is inaccurate, excessive, irrelevant, or no longer appropriate. OpenAI then weighs your request against the public interest in the information, and approval is not guaranteed.

Three caveats apply. First, a successful request affects ChatGPT answers only. Gemini, Perplexity, Claude, and Google's AI answers are run by other companies, each with its own process for managing personal data. Second, the information stays on every website that published it and in every search engine that indexed it, so the next search will still find it. Third, the rights behind this type of request are strongest for Europeans. A US founder or executive should not assume the same rights apply to them.

There is also a cost. If ChatGPT declines to discuss you while a Google search shows a detailed account, that discrepancy becomes the thing a reader notices, and it makes it look like you have something to hide. A slightly wrong answer you can correct is often a better position than no answer at all.

7. What you can do yourself

Most of this work is free, and you can start it this week. Run the AI response check from section 1, and save the answers with the date. Correct the pages and profiles about yourself that you control. Use one name form everywhere. And don't fall for anyone claiming they can guarantee what ChatGPT or any AI model will say about you by a certain date. They are promising something they don't control. For records involving a name collision, adverse material, or conflicting sources across countries, a private review is the place to start.