What Does AI Know About You? 3 Prompts to Find Out and Take Control

Ask ChatGPT to describe you and you hit something uncomfortable fast: the interesting part of what AI knows about you is never the part you typed. It is the part you didn't. Age range, income bracket, city, family situation, whether you are quietly managing a health problem — none of it volunteered, all of it assembled from the way you phrase things.

Public unease about this is already high. In a 2023 Pew Research Center survey, 81% of Americans who had heard of AI said they expected the information AI companies collect to be used in ways people would be uncomfortable with or never intended. Very few of those people have actually opened their own assistant and looked.

Source: Pew Research Center, 2023 — survey of U.S. adults who had heard about artificial intelligence.

This is a walkthrough, not a study: three prompts you can paste into ChatGPT or Claude in the next five minutes, the settings menus as they stood when we checked both products on 18 September 2026, and a closing section drawn from the AI visibility audits we run on client brands.

What does AI know about you?

AI assistants know what you tell them directly, such as your name, your job, or the company you work for. They also infer what you never shared — age range, income level, location, or health concerns — from how you write and what you ask. If memory is switched on, those details carry across conversations. You can review, pause, or delete them in settings.

Two separate mechanisms are at work here, and confusing them is why most people misjudge the risk. One is storage. The other is deduction.

Why does AI know more than you told it?

Because patterns leak more than sentences do. You control what you state; you don't really control what your phrasing, your recurring topics, and the timing of your questions give away. A chatbot doesn't need you to disclose your income when it can watch which price brackets you compare.

New York Times tech columnist Brian X. Chen tried this on himself. ChatGPT pointed out that he had repeatedly asked about allergic reactions to products and medications, and from occasional questions about toe pain it guessed he might be dealing with an ongoing health issue. He had never said any of that outright.

Source: The New York Times — Brian X. Chen, personal tech column (23 July 2026)

The signals that do the most work are ordinary and easy to miss:

• Vocabulary and register — how formally you write, which language you switch into, and the jargon you use without explaining it.
• Recurring topics — the subject you return to across weeks is a stronger signal than anything you said once.
• The numbers in your questions — budgets, salaries, rent, ticket prices, and product tiers place you in a bracket without you naming one.
• Who else appears in your prompts — a spouse, a manager, a landlord, a child's school. Each one narrows the picture.

Which 3 prompts reveal what AI knows about you?

These three, run in order, in an assistant where memory has been on for a while. Without memory the model can only see the current chat, so the exercise loses most of its point. Treat every answer as a guess — some will land, some will be wrong, and both outcomes are useful.

1. The inference check — what the assistant has guessed about you that you never stated.
2. The sensitive-information check — which of those details would be awkward in front of an employer, insurer, or stranger.
3. The stranger's-eye profile — how a marketer, recruiter, or scammer would use the same data against you.

Prompt 1: The inference check

Based on the way I write and the questions I ask, what have you guessed about me that I never told you directly, such as my age range, income bracket, city, or family and work situation? For each guess, explain which clues led you there.

What it exposes: the invisible profile built out of your writing style and your topics. The second half of the prompt does the heavy lifting — asking for the clues turns a list of assertions into something you can check.

Prompt 2: The sensitive-information check

From our past conversations, what have you learned or inferred about me that could be sensitive or awkward if an employer, insurer, or stranger saw it? Be specific, and tell me where each detail came from.

What it exposes: health, money, and relationship details, which is why this is usually the uncomfortable one. Naming the three audiences matters; a vague ask for "anything sensitive" gets you a vague answer.

Prompt 3: The stranger's-eye profile

Imagine a marketer, recruiter, or scammer could see everything you know about me. Write the short profile they would create, and explain what they could use to target or persuade me.

What it exposes: the same data as the first two prompts, reframed as a targeting brief. That reframing is the point — an abstract list of inferences is easy to shrug off, and a paragraph explaining how someone would work you is not.

Write down whatever surprised you before you close the tab. Then go delete it. If you want more prompts built on the same principle — context, a stated decision, and a demand for sources — we keep a longer set in our competitor analysis prompt library.

How do you delete what ChatGPT and Claude remember?

In ChatGPT, memory lives under Settings > Personalization > Manage Memory; in Claude, under Settings > Memory preferences. Both let you delete individual entries or clear everything, and both keep training controls in a separate menu from memory controls — which is the part people get wrong.

One mistake is worth flagging on its own. In ChatGPT, deleting a conversation does not delete the memories it created. The chat disappears from your history and the saved fact stays in Manage Memory, which is exactly the scenario where someone believes they have cleaned up and hasn't.

For readers in Indonesia, there is a legal floor under all of this: UU PDP, Law No. 27 of 2022 on Personal Data Protection, passed on 17 October 2022, governs how government bodies and private companies collect, use, disclose, and process personal data, and gives you rights over your own. This is not legal advice, and the practical protection is still the same one it has always been — share less to begin with.

What does this mean for brands?

The same inference machinery runs on companies, and the output reaches a far bigger audience. An assistant builds a picture of your brand from reviews, social posts, marketplace listings, and articles, then repeats that picture to everyone who asks about your category.

We run the brand-side version of these three prompts every week, across ChatGPT, Gemini, and Perplexity, for the accounts we audit. Two things show up often enough that we now expect them. The first: the description an assistant gives of a brand is usually assembled from whatever third-party page ranks well — a marketplace listing, an old review round-up, a forum thread — rather than from the brand's own site, so companies routinely find themselves summarized by a source they have never read. The second: the three engines disagree with each other, sometimes on the basics like what the company sells. Checking only one of them and calling it a result is the most common mistake we see.

None of that is visible from inside a marketing dashboard, which is why the audit has to be run as prompts rather than inferred from traffic. We wrote up how many prompts it takes to get a stable read in this piece on tracking AI brand mentions, and the broader mechanics sit on our AI search visibility page.

What this article doesn't cover

• Settings drift — the menu paths above were checked on 18 September 2026 in ChatGPT and Claude. Both products move these controls; if a path does not match, search the settings page for "memory" rather than assuming the feature is gone.
• What providers retain on their side — turning off training and memory changes what the product uses, not necessarily what is stored on a server for safety or legal reasons. Read the provider's own policy for that.
• Gemini, Copilot, and everything else — we covered ChatGPT and Claude because their controls are the clearest. The three prompts work anywhere; the settings table does not.
• Accuracy of the inferences — the model's guesses about you are guesses. A confident answer is not a verified one, and the exercise is about visibility, not truth.

Where should you start?

Open your most-used assistant, run Prompt 2, and read the answer properly. It takes about a minute and it is the fastest way to turn an abstract worry about AI privacy into a specific list of things to delete. Then clear what you don't want kept, and decide which topics will only ever go in a Temporary or Incognito chat from now on.

Worth stating plainly: Intura sells AI search visibility tools, so weigh the brand section accordingly. The method itself is just prompts — you can run the whole thing by hand for free, on your own brand, this afternoon.

If you would rather see the brand picture before you build the habit, our free AI visibility audit runs the category prompts across ChatGPT, Gemini, and Perplexity and shows you what they currently say about you.

Sources

• The New York Times — Brian X. Chen, personal tech column on AI privacy prompts (2026)
• Pew Research Center — public attitudes toward AI and data collection (2023)
• OpenAI Help Center — Memory FAQ and Data Controls documentation
• Claude Help Center — memory preferences, chat search and reference, and incognito chat documentation
• Republic of Indonesia — Law No. 27 of 2022 on Personal Data Protection (UU PDP), 17 October 2022