# AI hallucinations about people: why they happen
AI hallucinations about people explained: why assistants invent facts about named individuals, how to document them and which correction routes exist.
---
AI hallucinations about people happen because an assistant is built to produce a fluent, confident answer, and a person with little written about them gives it facts it cannot recall and a gap it fills with something plausible. OpenAI's own explanation is that training and testing reward guessing over admitting uncertainty, and that rare, arbitrary facts are where the guessing fails most. A named individual's degree, award, birthday or job history is exactly that kind of fact.
That is why the assistant can be right about a famous person and wrong about a mid-career professional in the same minute. It is also why the error looks so convincing: the invented detail has the right shape. A university that exists, an award with a believable name, a job title that fits your field.
This guide explains the mechanism in plain terms, then turns it into practice: how to document a false statement so that it is usable, which correction routes each provider publishes on its own pages (read for this guide), what those routes can and cannot do, and what wastes time. It is not a legal analysis and nothing here is legal advice. Provider pages change, so check the current version before you file.
The short version
A hallucination about a person is usually a guess filling a gap. The thinner your documented footprint, the more room there is for a guess, so being obscure does not protect you.
You cannot edit an answer, and one good answer after a retry is not a fix. Answers vary between runs, so document several runs, not one.
Every provider we read separates a quick feedback click from a formal privacy or correction request. The formal request is the one that names you and your data.
Providers weigh requests case by case and none promises a result. Evidence that you can hand over, such as an official page stating the true fact, is what moves a request.
Arguing with the chatbot, regenerating until it behaves, and mass feedback clicks do nothing durable.
1 in 365the chance OpenAI says a model has of guessing a birthday right
Case by casehow OpenAI says it assesses requests to remove personal data from ChatGPT responses
One monththe ICO's standard time to respond to a request to correct inaccurate personal data
Sources: OpenAI's page on why language models hallucinate, OpenAI's Help Center page on personal data removal from ChatGPT, and the ICO's guidance on the right to rectification, all read for this guide.
## Why AI assistants state false things about named people
The core reason is incentive, not malice and not a bug in one product. OpenAI's own page on why language models hallucinate says that standard training and evaluation procedures reward guessing over acknowledging uncertainty. Its multiple-choice comparison is simple: leaving an answer blank scores a zero, while a guess has some chance of being scored correct. A model tuned and graded that way learns that a confident guess beats "I don't know".
The same page explains why people are a weak spot. It says arbitrary, low-frequency facts, and it uses a pet's birthday as the example, cannot be predicted from patterns alone. Spelling and punctuation follow rules a model can learn. Your birthday, your thesis title or the year you joined a firm follow no rule. Either the model saw them enough times to hold them, or it has nothing to go on.
The page then gives a case about a person. Asked for the doctoral dissertation title of a named researcher, a widely used chatbot produced three different answers, none of them correct. Asked for his birthday, it gave three different dates, all wrong. For the birthday, the page notes that a guess has about a 1-in-365 chance of being right. Three answers, three failures, each delivered as fact. That is the pattern to expect when you ask about an individual with a thin record.
Three consequences follow, and they shape everything below.
- **The error is often fresh, not copied.** A hallucination is generated at the moment of the answer. It may exist on no web page at all, which is why searching for the false claim often finds nothing.
- **The error is unstable.** Because the answer is generated, asking again can produce a different version. OpenAI's help page on how its models are developed says the same question may yield different answers across different queries.
- **Being known helps, up to a point.** The same help page says ChatGPT responses may sometimes include personal information about individuals, particularly those who appear often online, such as public figures. Frequent documentation raises the odds that a fact is held rather than guessed. It does not make an answer reliable.
Providers say this about themselves too. Google's Gemini Apps help page says that large language model experiences, Gemini Apps included, can hallucinate and present inaccurate information as factual, and that Gemini Apps may produce inaccurate or offensive information, including about people. Anthropic's privacy policy does not promise the factual accuracy of outputs. None of the pages we read treats a false statement about a person as impossible.
## Who gets hallucinated about most: the thin-footprint problem
Many people assume the risk belongs to the famous. For specific facts the mechanism points the other way. The more reliably documented you are, in more than one independent place, the more likely a fact is held and not invented. The table below is our reasoning from the mechanism and from file experience, not a measurement. It describes tendencies.
Profile
Typical failure
What helps most
Risk of a repeat
Widely documented public figure
Old stories weighted too heavily; confident mixing of dates
Fresh authoritative pages that state the current facts
Medium
Established professional, modest coverage
Invented credentials, awards, employers that fit the field
One strong owned page plus matching profiles
High
Private person with a few mentions
Generic filler, or a stranger's facts attached to the name
Often nothing to add; a privacy request may be the better tool
High
Common or shared name
Two people blended into one record
Distinguishing facts on pages the assistant reads
High
The third row deserves attention. If very little exists about a private person, adding public pages to fill the gap is not always the right answer. Sometimes the right answer is to ask the provider not to show information about you at all, which is the privacy route covered below.
## Three failures that look the same on screen
An invented answer, a stale answer and a wrong-person answer all read as "the AI is wrong about me". Their causes differ, and so do their remedies. The test is where the claim could have come from.
- **Fabrication.** The claim exists nowhere you can find. No page, no cached copy, no archive. Fix lives with the provider's route and with giving the assistant a better source to read.
- **Retrieval error.** The assistant searched the web and summarized a page. The page is wrong or was misread. Fix lives with the page. Where an assistant shows its sources, you can open them and check.
- **Identity error.** The facts are real but belong to someone else. This is a matching problem, not an invention problem.
You can usually separate them with one habit: ask the assistant the same question three times, then ask it which sources it used, and open each cited page. If a cited page contains the false claim, you have a retrieval error. If a cited page contains a different person's version, you have an identity error. If nothing contains it, you have a fabrication. Treat that sorting as a working diagnosis, not proof. Assistants can cite pages that do not support the sentence beside them.
## What to record before anything else: documenting a false statement
Providers act on specific responses and specific data, and every route we read asks for examples. OpenAI's removal page tells you to give specific examples of where the information appears, including chat links and referenced URLs. Perplexity's help page asks for the URL of the query and a description of the error and the result you expected. A vague complaint that "AI says false things about me" gives a reviewer nothing to open.
Record these in a single dated file.
How to document a false AI statement
Capture the exact wordsCopy the full answer as text and take a screenshot that shows the question, the answer, the assistant's name and the date. Save the chat link if the product offers one.
Repeat the question three timesAsk the identical question in fresh sessions on different days. Save every answer, including the ones that are correct. A claim that appears once and a claim that appears every time are different problems.
Record the sources shownWhere the assistant cites pages, list each URL and note whether the page really says what the answer says. Mark each as supporting, contradicting or unrelated.
Write the true fact next to the false oneTwo columns: what the answer said, what is correct. Beside each correct fact, name the authoritative document that proves it: a registry entry, an employer page, a publication, an official profile.
Note who could have seen itIf the answer reached anyone but you, such as a client who told you, a recruiter, or a shared chat, write down who and when. This matters if a lawyer ever looks at the file.
The fourth step is the one most people skip. OpenAI says on its removal page that it cannot independently investigate disputed facts and assesses accuracy based on what you provide. So the quality of your proof is the quality of your request.
## The correction routes each provider publishes
Each provider we read offers a lighter lane and a heavier one. The names differ; the split does not. The table summarizes what each provider's own page said when we read it.
Provider
Light lane
Formal lane
What the page says it can do
OpenAI (ChatGPT)
Feedback on a response
Privacy Portal: remove personal data from ChatGPT responses
Stops information appearing in ChatGPT responses; does not remove it from websites or search engines
Google (Gemini Apps)
Feedback on a response; some data is reviewed by people
Report legal issue under the response; a Help Center form for privacy objections or corrections of inaccurate personal data
The page describes the routes; it promises no outcome
Anthropic (Claude)
Not covered on the page we read
Correction request to the privacy contact named in the policy
A reasonable effort to correct; the policy says it may not always be possible
Perplexity
Flag icon below the answer, or a support ticket
Privacy policy rights; not read in detail for this guide
The page lists misinformation and outdated information among the issues to report; it gives no outcome
### OpenAI
OpenAI's personal data removal page says anyone can ask, with or without a ChatGPT account, to stop personal information appearing in ChatGPT responses where it is inaccurate, excessive, irrelevant or no longer appropriate. You use the Privacy Portal, choose to make a privacy request, then choose to remove personal data from ChatGPT responses. OpenAI says a government-issued ID or alternative proof will be required, and that it may ask follow-up questions.
The page lists what OpenAI weighs: role in public life, accuracy, ongoing relevance, sensitivity, public interest, impact on the individual, official sources, criminal matters and whether the person was a child. It says no single factor is decisive, and that clearly incorrect or outdated information may weigh in favor of removal, especially where authoritative sources clarify the position. It also says information tied to a public or professional role may be more likely to remain relevant. The page did not describe a separate process for invented claims, and it did not say how removal is carried out technically.
### Google Gemini
Google's Gemini Apps help page says some collected data is reviewed by people, and that reviewed data is kept for up to three years, disconnected from your Google Account. It also says you can report a response in the app: below the response, choose More, then Report legal issue. For privacy objections or corrections to inaccurate personal data, it points to a Help Center request form. Only the second of those names you as a data subject.
### Anthropic and Perplexity
Anthropic's privacy policy says that if an output contains inaccurate personal data about you, you can submit a correction request, that Anthropic will make a reasonable effort to correct it, and that it may not always be possible because of the technical complexity of its models. It adds that you may be asked to verify your identity and may appeal a denial.
Perplexity's help center says you can report incorrect answers with the flag icon below the answer or through a support ticket, and asks for the URL of the query and a description of the error and the expected result. We did not read Perplexity's privacy rights section closely enough to describe it, so check it before filing a formal request.
## What a privacy request can and cannot do
A formal request is built on data protection law where it applies, and on the provider's own policy elsewhere. The useful reference point for a UK reader is the ICO's right to rectification guidance. It says an organization must respond without undue delay and at the latest within one month, extendable by up to two further months if the request is complex or there are several, and that the organization must tell you within the first month if it extends. It says personal data is inaccurate if it is incorrect or misleading as to any matter of fact. Opinions are harder to call inaccurate. If the organization refuses, it must tell you why, and that you may complain to the ICO or another supervisory authority, or go to court. The page is marked as under review following changes from the Data (Use and Access) Act, so check it before relying on it.
Two limits matter for hallucinations.
First, a request does not oblige a provider to confirm that your version is right. OpenAI's page says it assesses accuracy from what you provide, and the ICO page separates factual error from opinion. A claim such as "unreliable" or "controversial" is not a matter of fact in the way "holds a degree from X" is.
Second, the ICO's response to its generative AI consultation, on accuracy of training data and model outputs, says accurate training data will not stop generative AI models from hallucinating but can constrain the margin of error, and that the accuracy required is linked to the specific purpose of the model. So a regulator is not saying a model must never be wrong. A request is a way to ask a controller to deal with your data, not a way to demand a perfect model.
## What does not work
The failures are predictable, and each wastes time you could spend on the routes above.
- **Arguing with the assistant.** Telling a chatbot that it is wrong may change the next sentence in that chat. It does not correct the model or the sources it reads, and the next session starts again. Since answers vary between runs, a good reply after a protest proves little.
- **Regenerating until it behaves.** A single good screenshot feels like a result. It is one sample from a variable process, and it is useless as evidence of a fix.
- **Mass feedback clicks.** None of the pages we read promises a reply or an outcome for a feedback click. Coordinated clicking from many accounts is also manipulation, and we do not suggest it.
- **Writing pages aimed at the assistant instead of the reader.** Hidden text, keyword stuffing, planted claims or fake profiles to push a fact into answers breaks platform rules and can put your name in a worse position. A true fact on a clear page that a person would also want to read is the lawful version of the same idea.
- **Starting with a lawyer's letter.** A legal letter has its place, but it is slow and costly for a statement that may differ on the next run, and a provider route often costs nothing to try first. If the statement is plainly damaging and has been seen by others, take legal advice early, in your own country, and keep the documentation file ready.
- **Asking to remove accurate public-interest information.** If the information is accurate, lawful and tied to a public role, a request is likely to fail. OpenAI's own factors include public life, public interest and criminal matters. We do not help hide accurate information about wrongdoing.
## The narrower cases that sit next to this one
Four related situations have their own detail, and this guide mentions each once. Checking what ChatGPT says about you is an audit job: ask neutral and field-specific questions, save the answers, and compare across sessions. Perplexity is its own case: its help page asks for the URL of the query, and a wrong claim there is best handled by opening the sources the answer shows. Gemini has its own legal-issue and privacy lanes, covered above. A same-name case is an identity problem, solved by giving the assistant facts that separate you from your namesake and by never asking to remove facts that belong to them. Pick the case that matches your documentation before choosing a route.
## A worked scenario and a realistic timeline
Invented scenario, generic parties, to show the order of work. An independent consultant asks an assistant for a short biography. The answer says she won an industry award and previously worked at a firm she has never worked for. Neither claim appears in any search result.
Assumptions: she has a personal website with a short bio, a professional profile and two conference listings. She is in the UK. She wants the false statements gone and the correct ones used.
- She sorts it: nothing contains the claims, so it is a fabrication, not a retrieval error.
- She runs the same prompt three times on three days. The award appears in two of three answers. The firm appears in one.
- She writes the two-column file, citing her website bio, professional profile and conference listing as proof.
- She files the formal route on the assistant where it appeared most, with chat links and the file, and uses the light lane on the other assistants.
- In parallel she makes her own bio page state her actual employers and recognitions plainly, with dates.
The timeline below is what we usually see in files of this kind. It is experience, not a promise, and no provider page we read gave response times.
Stage
Usually takes
What you are waiting for
Documentation
One week
Three runs on different days, the evidence file
Formal request filed
Same day as the file is ready
Identity verification and any follow-up question
Provider reply
A few weeks; in the UK a one-month standard applies to data controllers
A decision that may be partial or a refusal with reasons
Source page changes
Weeks, depending on the site
Your page updated, pages that repeat the error corrected
Re-test
Monthly for a quarter
Whether the false claim still appears across several runs
The re-test is the honest measure. If the claim is gone across several runs in several sessions, the route worked. If it appears once in ten, it is reduced, not fixed, and the evidence file is still the thing you need.
## The next step if one is wrong about you now
If this is happening to you now
Save the answer, run the same question three times, and write the true fact beside each false one with the document that proves it. Then choose the formal route of the provider involved and keep the file for every other route. If you want a second pair of eyes on what the main assistants say about you, see our work on AI search reputation.
At Reputation Partners we start every AI file the same way: classify the error, document it across several runs, then decide whether the provider, the source page or a lawyer comes first.
## Questions people ask
### Why does the AI make up facts about me when it knows nothing about me?
Because it is built to answer, and a guess scores better than silence in the way many models are trained and tested, according to OpenAI. When a fact about you is rare and unpredictable, such as a degree, a date or a prize, the model fills the gap with something plausible. Little coverage of you makes this more likely, not less.
### Is an AI hallucination about a person the same as defamation?
Not automatically. A hallucination is a technical description of a generated error. Defamation is a legal question about a false statement of fact that harms reputation, published to others, and it depends on the country and the facts. Many hallucinations never reach anyone but the person who asked. If you think one is defamatory, take advice from a lawyer where you live.
### Will asking the assistant to correct itself fix it for other people?
No. Telling an assistant that an answer is wrong may change the rest of that chat, but the providers we read publish their correction routes as feedback tools and separate formal requests. A conversation is not a ticket. Use the provider's own route, and keep evidence in case a later answer repeats the error.
### Can I ask an AI company to delete what it holds about me?
OpenAI says anyone can ask for personal information to stop appearing in ChatGPT responses, and it assesses each request case by case. Anthropic's policy describes a correction request for inaccurate personal data. Outcomes vary, a refusal is possible, and in the UK you may complain to the ICO if a controller refuses.
### Why did the answer change when I asked again?
Because answers are generated, not looked up. OpenAI's help page says the same question may yield different answers across different queries. That is why you should record several runs on different days. One correct answer does not show the problem is fixed, and one wrong answer does not show it is permanent.
### Do I need a lawyer to get an AI answer about me corrected?
Usually not for a first attempt. Provider routes and a privacy request cost nothing and need only your evidence. A lawyer becomes useful when a statement is clearly damaging, other people have seen it, a provider has refused, or a deadline may be running. Bring the documentation file, because it shortens every conversation.