# How AI chatbots source information about people

How AI chatbots source information about people: training memory, live search, citations, how sources are weighed, and why wrong answers appear.

---


AI chatbots source information about people from two places: patterns absorbed from a very large body of text during training, and web pages fetched at the moment you ask. Whenever the assistant searches, it reads a handful of pages, writes a summary and, in most products, lists the pages it used as citations. When it does not search, the answer comes from memory alone.

That split explains most of what people find unsettling. A sentence that looks like a fact may be a statistical reconstruction. A sentence with a citation may rest on one weak page. And no provider we read describes how one source wins over another when two disagree. What the providers do publish is enough to describe the mechanism with confidence and to say where the documentation stops.

This article walks through the mechanism in the order an answer is built. It is based on the providers' own help pages and developer documentation, read for this guide and named in the text. Where a page says nothing, we say so. Anything from our experience on files is written as a tendency, never as a rule of the machine.

<div class="box"><b>The short version</b>
<ul><li>An answer about a person is built in stages: what the model learned in training, a decision whether to search, a search that finds pages, a selection of some of them, and a written summary. An error can enter at any stage.</li><li>Providers document the stages but not the weighing. OpenAI says ChatGPT ranks results using multiple factors and does not list them, and no other page we read does better, so any claim about a formula is a guess.</li><li>A wrong answer is usually a thin-evidence problem: few pages, old pages, pages that disagree, or a name shared with someone else. Fixing the evidence beats arguing with the answer.</li><li>Provider request forms exist, and OpenAI says its removal route covers responses, not the websites behind them. Treat them as a supplement to fixing sources.</li></ul></div>

<figure class="keyfacts">
<div class="keyfacts-grid">
<div><b>Three</b><span>kinds of crawler that OpenAI and Anthropic each document: one for training, one for search, one for user requests</span></div>
<div><b>1 in 365</b><span>OpenAI's own odds for a blind guess at a birthday, used to explain why rarely seen facts get invented</span></div>
<div><b>Case by case</b><span>how OpenAI says it assesses requests to remove personal data from responses</span></div>
</div>
<p class="fig-src">Sources: OpenAI crawler documentation, OpenAI's explanation of hallucinations, OpenAI's Help Center page on personal data removal, and Anthropic's crawler help page, all read for this guide.</p>
</figure>

## How AI chatbots source information about people, stage by stage

An answer about a named person passes through up to six stages, and each one has its own failure mode. Naming the stage is the first step in any diagnosis, because a training-memory error and a bad-citation error need different fixes.

<figure class="fig">
<figcaption><b>The path of an answer about a person</b></figcaption>
<ol class="fig-steps">
<li><b>Training</b>Before release, the model learns patterns from a large body of text. It does not keep a lookup table of facts about you.</li>
<li><b>Question</b>You type a name, or a question containing one. The assistant reads the whole conversation, not only your last line.</li>
<li><b>Decision to search</b>The product decides whether web information is needed. If it decides no, the answer comes from memory alone.</li>
<li><b>Search and fetch</b>The system sends one or more searches, finds pages in an index and may open some of them.</li>
<li><b>Selection and synthesis</b>The model picks what to use, writes a fluent summary and blends it with what it already learned.</li>
<li><b>Citation</b>Where the product shows sources, it attaches links to the answer. A link shows what was consulted, not necessarily what produced each sentence.</li>
</ol>
</figure>

A sentence can be traced to a stage only by testing, which the diagnostic table later in this article supports.

## What the model learned before you asked

The model learned patterns, not a file about you, and it learned from whatever was public, widespread and repeated. OpenAI's help page on how its models are developed says they use information that is publicly available on the internet, information from third-party partners, and information from users, trainers and researchers. It says training teaches the model to predict the next most likely word, and that the model does not retain copies of the sentences, images or audio it processes during training.

For reputation, three points from that page matter. The first is that the page says responses may sometimes include personal information about individuals, and that models may learn personal information in order to recognize public figures and well-known entities. So the person who is widely written about is the person the model knows best. The second is that OpenAI says it takes steps to reduce the processing of personal information in training, including excluding some sources that aggregate large amounts of it, and applying filters.  The third is that the page gives no knowledge cutoff or dates for the data, so you cannot read from the provider when your latest news stops being part of memory.

Training memory is a layer you can only observe. The layer you can influence reads pages that exist today.

## When the assistant decides to search

Whether an assistant searches is a product decision made per question, and the providers describe it only in general terms. Microsoft's transparency note for Copilot says that for certain conversations where users seek information, Copilot is grounded in web search results, and that a determination is made whether the request requires grounding data from the web. It does not say what the criteria are. Anthropic's help page says Claude searches when a topic benefits from current information, invoking a search tool to ground its responses with content from the live web. OpenAI's page on searching the web with ChatGPT says it may search on its own when a question would benefit from current information, and that a user can also start a search by hand.

The consequence for a person is practical. Questions about a named individual often look like questions that benefit from current information, so in our experience many name queries do trigger a search. But the same question can be answered from memory in one product, one mode or one session and from pages in another. That is why an audit has to record which assistant, which mode, and whether sources were shown.

OpenAI documents a manual search option, so the same name can get different answers depending on whether the user chose search. In our experience, phrasing such as a request for the latest news also pushes toward search.

## How the search layer finds and picks pages about you

The search layer works like a search engine with a writer on top, and the first rule is that a page the system cannot reach cannot be used. Three documented mechanisms matter: the crawlers, the index and the fan-out of queries.

### Crawlers, and why there are three

[OpenAI documents three bots](https://developers.openai.com/api/docs/bots). GPTBot is used to make its generative AI foundation models more useful and safe, and disallowing it indicates that a site's content should not be used in training. OAI-SearchBot surfaces websites in ChatGPT's search features, and OpenAI says it can take about 24 hours from a robots.txt update for its systems to adjust. ChatGPT-User handles certain user-initiated actions, and OpenAI says that because those actions are initiated by a user, robots.txt rules may not apply. [Anthropic documents](https://support.claude.com/en/articles/8896518-does-anthropic-crawl-data-from-the-web-and-how-can-site-owners-block-the-crawler) the same three-way split: ClaudeBot for possible training, Claude-SearchBot to index content for search relevance, and Claude-User for when a user asks Claude a question. Anthropic says its bots honor robots.txt and that IP blocking may not reliably produce an opt-out.

<div class="tablewrap"><table><thead><tr><th>Job</th><th>OpenAI bot</th><th>Anthropic bot</th><th>What blocking it does</th></tr></thead><tbody>
<tr><td><strong>Training data</strong></td><td>GPTBot</td><td>ClaudeBot</td><td>Signals that your future content should stay out of training</td></tr>
<tr><td><strong>Search index</strong></td><td>OAI-SearchBot</td><td>Claude-SearchBot</td><td>Reduces how often your pages appear in the provider's search answers</td></tr>
<tr><td><strong>User request</strong></td><td>ChatGPT-User</td><td>Claude-User</td><td>May reduce visibility in user-directed lookups; OpenAI says robots.txt may not apply</td></tr>
</tbody></table></div>

A person who wants less said about them may block every AI bot on their own website, then find assistants describing them from a stranger's page. Blocking your own site does not remove other sites. It removes the one source you control.

### The index and the fan-out

[Google's developer page on AI features](https://developers.google.com/search/docs/appearance/ai-features) says its AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources, and that this can display a wider and more diverse set of links than a classic search. It says a page must be indexed and eligible to be shown in Google Search with a snippet to be a supporting link, that there are no additional technical requirements, and that indexing and serving are not promised. It also says AI Overviews appear only when its systems determine they are additive to classic Search.

For a name, fan-out plausibly means one question becomes several: the name plus a profession, a city or a company. A page that answers none of them may never be retrieved. OpenAI's announcement of ChatGPT search says it leverages third-party search providers, and its help page says it sometimes partners with other search providers; the announcement also says any website or publisher can choose to appear. We found no primary source saying which index serves each assistant, so we do not say.

## How sources are weighed: what is documented and what is not

No provider page we read explains how one source is ranked against another in an answer about a person, so what follows is split into the documented and the inferred. Perplexity's help page says it uses AI to search the internet in real time and draws on authoritative sources like articles, websites and journals, with numbered citations; it does not say how it chooses among them. OpenAI's page on searching the web with ChatGPT says it ranks results using multiple factors intended to help users find relevant, reliable information, that placement is not guaranteed, and does not describe the factors; it says responses may include citations. Anthropic's help page says every Claude response includes citations when it searches. Copilot's note says that when grounded in web data the output contains hyperlinked citations below the text. The Gemini help page we read did not describe how sources are shown, so we say nothing about Gemini citations.

So the honest summary is that the products we could confirm cite their sources, none of them documents the weighing in detail, and the citation list is a courtesy, not an audit trail.

What we can say comes from experience on files and from how retrieval works, and it is a tendency, not a law:

- **Agreement beats volume.** When three independent pages state the same fact in the same words, assistants tend to repeat it. When two sources disagree, the answer tends to hedge, pick one, or merge them.
- **Specific beats vague.** A page that states a role, a company and a city in one clear sentence is easier to quote than a long biography that implies them.
- **Reachable beats authoritative.** A well-regarded page that cannot be fetched or indexed is not used.
- **Old pages persist.** An outdated page that remains indexed and well linked can outweigh a newer page that says less.

Treat these as hypotheses to test with your own name. They are why a source-first strategy usually works better than a complaint-first one, and why it fails when accurate pages are absent.

<aside class="callout">
<span class="callout-label">Limit</span>
<h4>A citation is not proof of origin</h4>
<p>A linked source shows that the page was consulted. It does not show that the sentence beside it came from that page. Always open the cited page and check that it says what the answer says.</p>
</aside>

## Why wrong answers about people appear

Wrong answers appear when the evidence is thin, old, contradictory or shared, and when the writing step fills gaps with plausible text. [OpenAI's explanation of hallucinations](https://openai.com/index/why-language-models-hallucinate/) gives the clearest primary account of the last cause. It says language models first learn by predicting the next word in huge amounts of text, that arbitrary low-frequency facts, such as a birthday, cannot be predicted from patterns alone, and that a blind guess at a birthday has a 1 in 365 chance of being right. It adds that most evaluations measure performance in a way that encourages guessing rather than honesty about uncertainty.

For a person with a small footprint, that is the mechanism. The model has seen your name a handful of times. A specific fact such as a degree, a year of birth, an employer or a role is exactly the kind of rarely seen detail that gets filled in with something that sounds right. Microsoft's Copilot note says Copilot could fabricate content that might sound reasonable but is factually inaccurate, and that responses grounded in sources might misrepresent that content in a way that is not completely accurate, up to date or reliable. Google's help page for AI Overviews says they can and will make mistakes, and Gemini's help page says Gemini can hallucinate and present inaccurate information as factual.

The failure modes we see on files fall into a small set. Each one has a different tell.

<div class="tablewrap"><table><thead><tr><th>Symptom</th><th>Likely stage</th><th>How to test</th><th>Where the fix sits</th></tr></thead><tbody>
<tr><td><strong>Wrong detail, no source shown, changes between runs</strong></td><td>Training memory or gap-filling</td><td>Ask the same question three times; ask for sources</td><td>Strengthen accurate pages; provider request for responses</td></tr>
<tr><td><strong>Wrong detail with a citation that says it</strong></td><td>Retrieval of a bad source</td><td>Open the cited page and find the sentence</td><td>The source page: correct, update or remove at origin</td></tr>
<tr><td><strong>Right source cited, wrong sentence</strong></td><td>Summarization error</td><td>Compare the page text to the answer line by line</td><td>Make the page statement shorter and less ambiguous; report the answer</td></tr>
<tr><td><strong>Facts about a different person mixed in</strong></td><td>Name collision</td><td>Add a distinguishing detail to the question and compare</td><td>Disambiguating facts across your owned and third-party pages</td></tr>
<tr><td><strong>An old fact stated as current</strong></td><td>Stale index or memory</td><td>Check the date of cited pages</td><td>Update the source; add dated, current statements</td></tr>
<tr><td><strong>Nothing found, or a vague answer</strong></td><td>Thin evidence</td><td>Ask a narrow question about a known fact</td><td>Publish clear, consistent facts where assistants can reach them</td></tr>
</tbody></table></div>

Two cases deserve extra care. Name collision, where an assistant blends two people, usually means the sources for both names are thin and nothing in the question separates them; the fix is more distinguishing detail in sources. And an accurate but unwelcome answer is not an error: if the assistant is quoting a lawful, accurate, public-interest report, no route in this article is meant to hide it. The work there is to make sure the answer is complete and current.

## What you can change at each stage

You can change very little at the training stage, a lot at the source stage, and some at the answer stage through the provider's channels. The map below is the same one we would use on a file.

**Training memory.** OpenAI's help page on how its models are developed says individuals in certain jurisdictions can object to the processing of their personal information by its models, through its Privacy Portal. Its [page on personal data removal](https://help.openai.com/en/articles/20001057-right-to-be-forgotten-and-personal-data-removal-from-chatgpt) describes a request to remove personal data from ChatGPT responses, open to anyone with or without an account. It says it assesses requests case by case, balancing privacy against freedom of expression and the public interest, and lists factors such as whether you are a private individual or have a public or professional role, how old the information is, how sensitive it is, whether there is a public interest, the impact on you, and whether it concerns children, criminal allegations or official records. It says that if you believe information is wrong, you should explain why and provide reliable evidence, because OpenAI cannot independently investigate disputed facts. It also states that removal prevents the information from appearing in ChatGPT responses but does not remove it from external websites. The page covers responses, and it does not say whether data can be taken out of the trained model itself.

**Search and sources.** This is where the effort belongs. Correct or update the page the assistant read, through that platform's report route or the publisher's correction process. Where accurate pages are missing, publish them. Keep your own site reachable by the search bots of the assistants that matter to you.

**The answer itself.** Providers offer feedback buttons and request forms. Google's help page for AI Overviews describes thumbs-down feedback leading to a report-a-problem option for results that seem unhelpful, inaccurate or biased, and says the feedback includes your most recent search query and its results. Treat a feedback button as a signal, not a correction. We found no provider page promising an individual a corrected answer.

<aside class="callout">
<span class="callout-label">Warning</span>
<h4>Do not game the sources</h4>
<p>Invented reviews, planted pages and hidden text aimed at assistants violate platform rules and create a second problem larger than the first. Publish true things clearly. Nothing else holds.</p>
</aside>

## A worked scenario with a realistic timeline

This is an invented scenario with generic parties, built to show the order of operations and the kind of time involved. A consultant, call her the subject, finds that an assistant says she works at a firm she left four years ago and lists a certification she never earned. Her name is shared with a person in another country who does hold that certification.

She runs the same question three times on two assistants. One run gives the old employer and no sources. Another cites a three-year-old conference page listing her old role. The certification appears only in unsourced answers. That separates two problems: a stale source she can fix, and a name-collision error in memory or gap-filling that she cannot see directly.

She asks the conference organizer to update the speaker page. She updates her own site with her current role, location and one sentence that distinguishes her from the other person, and aligns the same wording on two profiles she controls. She files a provider request for the certification line, with evidence.

<figure class="fig">
<figcaption><b>Realistic sequence for the scenario</b></figcaption>
<ol class="fig-steps">
<li><b>Day 1, record</b>Run the same prompts three times on each assistant. Save the text, the sources and the date. This is the baseline.</li>
<li><b>Days 1 to 3, fix what you control</b>Update your own site and profiles with one consistent, specific statement. Send the correction request to the third-party page owner.</li>
<li><b>Days 3 to 7, file requests</b>Submit the provider request for the unsourced claim with evidence. Do not expect a visible change yet.</li>
<li><b>Weeks 1 to 3, recrawl</b>Search-based answers usually start to reflect changed pages as they are fetched again. Some change in days, others take weeks, and a page that is never re-fetched does not change.</li>
<li><b>Week 4 onward, compare</b>Re-run the baseline prompts. Judge by the pattern across runs, not by one answer.</li>
<li><b>Month 3, review</b>Decide whether the gap that remains is a source problem, a name-collision problem or a memory problem, and act on that layer only.</li>
</ol>
</figure>

The timings are experience ranges, not promises. OpenAI says only that its systems can take about 24 hours to adjust after a robots.txt update, a statement about crawling rules and not about answers. We found no provider that states how long a corrected page takes to change an answer, and anything more precise is a guess.

## What to do this week

Record what three assistants say, as the scenario did on day one, and fix the one source that is wrong or missing. Do that before filing any provider request, because a request is stronger when you can point to the page and the evidence.

<div class="box hot"><b>If this is happening to you now</b>
<p>If an assistant is saying something about you that is inaccurate, outdated or unlawful, start by recording it and finding the stage it came from. Reputation Partners can run that diagnosis and the source work with you, and the AI search page explains how it works.</p>
<p><a href="/ai-search-reputation">See how we handle what AI assistants say about you</a></p></div>

If the issue is wider than one sentence, a standing schedule beats one fix. Our [monitoring service](/reputation-monitoring) repeats the same questions on a fixed prompt set so that you see a change when it happens.

## Questions people ask

### Do AI chatbots look me up live every time?

No. Whether an assistant searches is decided per question by the product. Microsoft describes a determination step in Copilot, and Anthropic says Claude searches when a topic benefits from current information. A name query may be answered from training memory alone in one mode and from live pages in another, which is why you should record the mode and whether sources were shown.

### Does an AI model store a profile of me?

Not in the sense of a file. OpenAI says its model does not retain copies of the sentences, images or audio it processes during training, and describes learning patterns by predicting the next word. A model can still reproduce details about a person, because it may have learned from text that mentioned them. That is reconstruction from patterns, which is why it can be wrong.

### Why does the answer change when I ask again?

Responses can vary. OpenAI's help page says the same question may yield different answers across different queries. If search is involved, the pages fetched may also differ between runs. Change between runs is a useful signal: it usually means the answer rests on memory or gap-filling rather than a stable, cited source.

### Can I stop AI companies from using my website?

For your own site you can set robots.txt rules for each crawler. OpenAI says disallowing GPTBot indicates content should not be used in training, and Anthropic says its bots honor robots.txt. Blocking search bots reduces how often your pages appear in answers. The rules apply only to your site and do not touch what other sites say about you.

### Does an OpenAI removal request delete the information about me?

OpenAI says its request prevents the information from appearing in ChatGPT responses, but does not remove it from external websites. It assesses requests case by case and cannot independently investigate disputed facts, so evidence matters. The page does not say the information is deleted from the trained model, and approval is not promised.

### Are citations a reliable guide to where a claim came from?

Not on their own. Copilot, Claude and Perplexity all describe linked or numbered sources, but none of their pages we read says each sentence is traced to one link. Open the cited page and check that it says what the answer says. Treat any mismatch as a finding.