5 Things to Never Tell ChatGPT: A Practical Privacy Guide
Key Takeaways
- ChatGPT is not private in the way a diary is private: OpenAI stores your conversations by default, uses them to improve its models unless you opt out, and can be legally required to hand them over.
- The five things to never type into ChatGPT are: your full name and home address together, bank and financial details, government ID numbers, passwords and login details, and photos or scans of identity documents.
- A US court ordered OpenAI to hand over about 20 million de-identified ChatGPT conversations in a copyright case, and OpenAI filed on 13 November 2026 asking a judge to reverse that order.
- In July 2025, more than 4,500 shared ChatGPT conversations appeared in Google search results after users ticked a "make this chat discoverable" box; OpenAI removed the feature within days.
- Deleting a chat normally removes it within 30 days, but a legal hold can pause that clock, and data already used in training cannot be extracted.
- You can reduce the risk with Temporary Chat, turning off model training and memory, and redacting sensitive details before you paste — but no setting can defeat a valid legal order.
The One Rule Behind the List
ChatGPT is genuinely useful, and for everyday tasks — asking questions, drafting letters, planning, brainstorming — it is safe enough. The privacy problem is not that ChatGPT is dangerous. It is that most people treat the chat box like a private diary when it is, in the words of one privacy researcher quoted in the coverage, a postcard: anything you type into it is stored by OpenAI and could, in theory, be reviewed. As Jennifer King, a privacy researcher at Stanford University, puts it: "Once you type something in, you lose possession of it."
That is the rule behind everything in this article. Before you paste anything into ChatGPT, ask whether you would be comfortable emailing it to a stranger. If the answer is no, do not paste it. The five things below are the cases where that rule matters most, because the harm from exposure is concrete: identity fraud, stolen money, compromised accounts. Each section explains what the risk is, why it is real, and what to do instead.
Two things this article does not do. It does not claim that your chats are public: by default they are not, and OpenAI says they are not shared with advertisers. And it does not invent incidents. Every risk in this article is grounded in OpenAI's own published policy, in documented settings, or in events that were reported at the time and are cited below. Where a claim is disputed or the details are still developing, the article says so.
What ChatGPT Actually Collects
Before the list, it helps to know what you are handing over. OpenAI's privacy policy, updated on 6 February 2026, says it collects your prompts, uploaded files, images, audio and video, your IP address, your device details, your general location, and your date of birth. It also collects account information: your name, contact details, account credentials, payment information and transaction history. If you connect your contacts, it uploads your device address book too.
On consumer plans — Free, Go and Plus — your conversations are used to help train OpenAI's models by default. That means the words you type can become part of the data that shapes future versions of the model. You can turn that off, and the settings section below explains how, but the default is on, and many users never change it.
Your chat history is protected while it is being sent and stored, but it is not protected with zero-access encryption. That means OpenAI can access your past conversations, and a limited number of its employees and outside contractors can read specific conversations, for example for safety review or when you submit feedback. The code of ChatGPT's apps is not open source, so there is no public accountability into what is logged or how it is processed behind the scenes.
None of that is secret. It is all in OpenAI's own help pages and privacy policy. The problem is that very few people read those pages, and the chat interface gives no visible sign of any of it. The box looks private. The policy says it is not.
1. Your Full Name and Home Address Together
Your name alone is low-risk. Your address alone is the same. Together, they give an identity fraudster enough to open accounts, apply for credit, or commit fraud in your name. That is why the combination is the first item on the list, and why it appears in every version of it, from the Wall Street Journal's original list in March 2025 to the guides published since.
The risk is not that ChatGPT will publish your address. It is that the combination sits in a database that can be accessed in ways you do not control: by OpenAI staff, by a contractor, by a court order, or by an attacker in the event of a data breach. No database is perfectly secure, and a name-plus-address pair is one of the most reusable pieces of personal data there is. It is also the piece that makes every other piece more dangerous: a fraudster who has your name and address can use them to support a fake application built from other details gathered elsewhere.
What to do instead: if you need help with a letter, a form, or an address-related task, use placeholders. Write "123 Example Street" or "[my address]" and fill in the real details yourself afterwards. If you are asking about a local service, a neighbourhood, or a planning question, give the area, not the street. The advice ChatGPT gives you will be the same, and your address will not be in its logs.
2. Bank and Financial Details
Card numbers, sort codes, account numbers, PINs, and online banking credentials should never go into ChatGPT. None of these should ever be typed into a chat box, for any reason, including to ask for help with a transaction or a budget.
The reason is simple: ChatGPT cannot access your accounts, so sharing your bank details achieves nothing except putting them in a database. There is no task that requires ChatGPT to know your card number. If you want help with a budget, use rounded figures or percentages. If you want help understanding a bank statement, redact the account number, sort code and any transaction references that identify you, and paste only the lines you need help with. If you want help with a loan or a mortgage, use example figures and adjust the answer yourself.
The same applies to investment account details, cryptocurrency wallet keys, and any financial login. A password or a seed phrase typed into a chat is a password or a seed phrase stored on someone else's server. There is no setting that makes that safe, and no task that makes it necessary.
3. Government ID Numbers
Your National Insurance number, your NHS number, your passport number, your driving licence number — in the United States, your Social Security number — are unique identifiers. They are protected under UK GDPR as personal data, and fraudsters specifically target them for identity theft. The UK's data regulator, the Information Commissioner's Office, has published guidance on your rights when using ChatGPT, and the point it makes is the same one this article makes: these numbers are the keys to your identity in official systems, and they do not belong in a chatbot.
A National Insurance number, in the format AB 12 34 56 C, is what you use for tax and benefits. An NHS number is the unique reference on your NHS card and in letters from your GP surgery. Either one, combined with your name and date of birth, is enough for a fraudster to impersonate you to a government service, a bank, or an employer. A passport or driving licence number is the same: it is a credential, not a piece of trivia.
What to do instead: if you need help with a form that asks for one of these numbers, describe the form and the question, not the number. Ask what the field means, what format it expects, or what to do if you have lost the document. ChatGPT can answer all of those questions without ever seeing the number itself. If you are drafting a letter to a government department, leave a placeholder and fill it in by hand.
4. Passwords and Login Details
Passwords should never go into ChatGPT. Neither should login details, security question answers, one-time codes, or recovery keys. ChatGPT cannot log in to your accounts for you, so sharing a password achieves nothing except putting it in a database.
This is the item people are most likely to get wrong by accident. It happens when someone pastes a block of text that contains a password — a set-up guide, an error message, a configuration file — without checking what is in it. It happens when someone asks ChatGPT to remember a password, or to generate a password and then pastes the old one for comparison. It happens when someone shares a screenshot that includes a logged-in session.
The fix is a habit, not a setting: before you paste anything, scan it for credentials. If you are sharing code, check for API keys and tokens. If you are sharing an error message, check for session IDs. If you are sharing a screenshot, crop out anything that shows you logged in. And use a password manager for your passwords — not a chatbot, not a note, and not your memory. If you have already pasted a password into ChatGPT, change it immediately, and change it everywhere else you used the same one.
5. Photos and Scans of Identity Documents
A photo of your passport, driving licence, national ID card, or birth certificate contains a concentration of personal information: your full name, your date of birth, your document number, your photograph, and sometimes your address and signature. A passport photo combined with a date of birth and a document number is more than enough for identity theft.
People paste these into ChatGPT for understandable reasons: to ask what a document says, to check whether a photo meets the requirements, to get help filling in a form that asks for the details on the document. In every case, there is a safer way. Describe the document and the question. Ask what the requirements are for a passport photo, rather than uploading your passport photo. Ask what a field on a form means, rather than uploading the form with your details on it. If you need a document translated or summarised, redact the identifying fields first, or use a local tool that does not send the image to a server.
The same applies to photos of other people, and especially to photos of children. A child's face, name and date of birth, pasted into a chat, is a record that you cannot take back. Do not upload it.
Beyond the Five: What Else to Keep Out
The five items above are the highest-risk ones, but they are not the only things worth keeping out of a chat box. The general rule — if it is highly confidential, sensitive, regulated or proprietary, keep it out — covers a wider set of material, and the reporting on ChatGPT privacy names several categories that deserve the same treatment.
Medical records and detailed health information are one. ChatGPT can help you understand a condition in general terms, but your own test results, diagnoses and treatment history are health data, and they belong with your doctor, not in a chat log. Legal strategy is another: what you tell a lawyer is protected by privilege; what you tell a chatbot is not, as the Heppner case below shows. Work secrets are a third: source code, contracts, customer lists, pricing, and unreleased plans are all things your employer would not want on someone else's server, and many employment contracts prohibit sharing them with third-party tools. Details about your children, your exact location in real time, and intimate personal disclosures belong in the same category.
None of this means you cannot use ChatGPT for work or for personal questions. It means you should use it the way you would use a helpful colleague in an open-plan office: fine for general questions, drafting, and public information, and not the place for anything you would not want overheard.
The Two Incidents That Prove the Risk Is Real
Sceptics sometimes ask whether any of this has ever actually happened. Two documented incidents answer that question, and both are worth knowing in detail because they show different kinds of exposure.
The first is the court case. In the New York Times copyright lawsuit against OpenAI and Microsoft, which dates to December 2023, a federal court ordered OpenAI to preserve ChatGPT output logs, including chats users had deleted. The preservation order came from Magistrate Judge Ona Wang on 13 May 2025. On 7 November 2025, Judge Wang ordered OpenAI to produce about 20 million de-identified consumer ChatGPT logs — a random sample of complete conversations from December 2022 to November 2024 — to the news plaintiffs. The order was affirmed on review by District Judge Stein on 5 January 2026. Enterprise, Education, Business and API customers were excluded; the sample covered consumer conversations only.
The story did not end there. On 13 November 2026, OpenAI filed a request asking a federal judge in New York to reverse the order, arguing that producing the chat logs would expose confidential user information and that 99.99% of the transcripts have no connection to the copyright claims in the case. OpenAI's chief information security officer, Dane Stuckey, described the order as "an invasion of user privacy". The judge had previously ruled in favour of disclosure, with safeguards for de-identification, and OpenAI faced a deadline to comply. The outcome of the latest filing is not yet known.
What the case proves is not that your chats are public. It is that "deleted" does not always mean deleted, and that a legal order can override your settings. OpenAI's published policy schedules deleted chats for permanent removal within 30 days, with exceptions for legal obligations. A legal hold pauses that clock. The 20 million logs were de-identified first, meaning names and account details were stripped out — but de-identified is not the same as anonymous. If you typed your own name, your town, or your company into a chat, that is still sitting in the words themselves.
The second incident is the shared-links episode of July 2025. ChatGPT's share feature creates a shareable link to a conversation. In 2025, users who clicked "share" were offered a checkbox labelled "Make this chat discoverable", with the subtitle "Allows it to be shown in web searches". Many did not realise what that meant. On 30 July 2025, Fast Company reported that a plain Google site search against the share URL path returned nearly 4,500 conversations, some of which described mental health disclosures, addiction, drug use, and traumatic experiences. TechCrunch confirmed the finding the next day, and reported that the material it surfaced was largely mundane — but the point was not what the chats contained. It was that they were findable at all.
OpenAI removed the feature within hours of the TechCrunch report. Dane Stuckey said on 31 July 2025 that the feature "introduced too many opportunities for folks to accidentally share things they didn't intend to", and the company worked with search engines to remove the indexed pages. The discoverability option is gone. Shared links themselves still exist, and anyone with a link can open the chat and forward it. The lesson survives the fix: a shared link is a published page, and deleting the original chat does not automatically revoke every copy that was made of it.
A third ruling, from February 2026, makes the legal point sharper still. In United States v. Heppner, a criminal defendant used the public version of Anthropic's Claude to work through his defence after a grand jury subpoena. Judge Jed Rakoff of the Southern District of New York ruled that the chats were protected by neither attorney-client privilege nor the work-product doctrine. The case concerned Claude, not ChatGPT, but the principle is the same: a conversation with a chatbot does not carry the legal privilege that protects what you tell a lawyer, a doctor, or a therapist. If you are facing a legal problem, talk to a lawyer — not to a chatbot, and not about the chatbot conversation afterwards.
The Settings That Actually Help
There are settings that reduce the risk, and they are worth using. They do not make ChatGPT private, and they do not defeat a court order, but they do reduce what is stored and what is used for training.
Turn off model training. On consumer plans, your conversations are used to help train OpenAI's models by default. In ChatGPT's settings, under Data Controls, you can turn off "Improve the model for everyone". That stops new conversations from being used for training. It does not remove data already used, and it does not stop OpenAI from storing the conversations.
Use Temporary Chat. Temporary Chat does not save the conversation to your history and does not use it for training. It is the closest thing ChatGPT has to a private mode. Your conversations are still logged in the short term, and a legal hold can still apply, but nothing is kept in your history and nothing is used to train the model.
Turn off memory, or manage it. ChatGPT's memory feature remembers details about you across conversations. You can turn it off, or review and delete what it has stored. If you have told ChatGPT things you would rather it forgot, check the memory settings and clear them.
Review your shared links. In Data Controls, you can see the links you have shared. Delete any you no longer need. Remember that deleting the original chat also removes its link, according to OpenAI's shared-links FAQ, but a link that was copied or archived elsewhere is beyond your control.
Turn on two-factor authentication. Your account is the front door to every conversation you have ever had. Two-factor authentication is the single most effective step you can take to stop someone else from reading them. Our earlier report on a Danish man who sued OpenAI after ChatGPT advice led to his eviction is a reminder that the consequences of a chatbot's words can be real, and that the record of what was said matters.
None of these settings changes the fundamental fact: your chats sit on OpenAI's servers, without zero-access encryption, and can be read by the company, reviewed by contractors, or produced under a legal order. The settings reduce the exposure. They do not eliminate it.
What to Do If You Have Already Shared Something Sensitive
If you have already pasted something sensitive into ChatGPT, you are not powerless, but you should be realistic about what can be undone.
Delete the chat. In ChatGPT, delete the conversation from your history. OpenAI says it removes the data from its systems within 30 days, except where a legal obligation requires it to be kept. Deleting the chat also removes its shared link, according to OpenAI's FAQ. Do this first, because it stops the conversation from sitting in your history where anyone with access to your account can read it.
Turn off training and memory, as above, so that the conversation is not used to shape future models and is not remembered across chats.
Ask OpenAI what it holds. Under UK data protection law, you can ask OpenAI what personal data it holds on you, and you can ask for it to be deleted under the right to erasure in GDPR Article 17. OpenAI must respond within 30 days, usually with a downloadable file. The same right exists, in different forms, in other jurisdictions. This is a real remedy, and it is worth using if the material is sensitive.
Accept the limit. If your words were already used in a training run, they cannot be extracted. Researchers at Stanford's Human-Centered AI institute have flagged exactly this: large language models can memorise personal details and, in rare cases, surface them to someone else later. Think of it like a spoonful of sugar stirred into batter — you cannot fish the sugar back out once it is mixed in. That is not a reason to panic, and it is not a reason to stop using ChatGPT. It is a reason to be careful next time, and to change any password or credential that was exposed, rather than hoping the deletion was enough.
What Is Fine to Share
A privacy guide that only says "don't" is not much use, so here is the other side. ChatGPT is safe for general questions, drafting, brainstorming, explanations, and anything involving public information. Ask it to explain a concept, draft a letter from a template, plan a trip, summarise a public article, or help you learn a skill. Ask it about a neighbourhood in general terms, a medical condition in general terms, or a legal process in general terms. Use it to practise a language, to write code that contains no secrets, or to work through an idea.
The test is the postcard test. If you would be comfortable with the words on a postcard — visible to the postal service, to anyone who handles it, and to anyone who finds it if it is lost — then it is fine for a chat box. If you would want it in a sealed envelope, keep it out of the chat box. That is the whole guide, in one sentence.
For readers who want to compare ChatGPT with its rivals on privacy and features, our guide to ChatGPT alternatives worth trying covers the options, and our ChatGPT, Claude and Gemini comparison puts the three assistants side by side. Some alternatives offer local processing or stronger privacy guarantees. None of them makes the postcard test unnecessary.
What We Could Not Verify
- The outcome of OpenAI's 13 November 2026 filing asking a judge to reverse the 20-million-log production order is not yet known. We report the filing as reported by Reuters and trade coverage, not as a decided outcome.
- The figure of "more than 100,000 shared chats saved by the Internet Archive's Wayback Machine" after the July 2025 discoverability episode is reported by one outlet and we have not verified it independently. We cite only the 4,500 figure from Fast Company's contemporaneous reporting.
- OpenAI's ad programme, which began testing in the United States in February 2026, is described by OpenAI as keeping conversations private from advertisers and not selling user data. We have not independently verified how ad personalisation draws on chat history.
- Whether a specific conversation of yours was included in the 20 million produced logs cannot be verified. The sample was a random selection of consumer conversations from December 2022 to November 2024, and OpenAI has not published a way for individuals to check.
Sources
- All About Cookies, "Is ChatGPT Safe? What OpenAI Keeps After You Hit Delete": https://allaboutcookies.org/is-chatgpt-safe
- WellWired, "5 Things to Never Tell ChatGPT": https://wellwired.ai/blog/what-not-to-tell-chatgpt
- Proton, "Is ChatGPT safe?": https://proton.me/lumo/ai/is-chatgpt-safe
- TechCrunch, "Your public ChatGPT queries are getting indexed by Google and other search engines," 31 July 2025: https://techcrunch.com/2025/07/31/your-public-chatgpt-queries-are-getting-indexed-by-google-and-other-search-engines/?utm_campaign=WIR
- YourStory, "ChatGPT disables tool that made private chats publicly searchable": https://yourstory.com/ai-story/openai-chatgpt-search-index-share-feature-removal
- Jaroha, "Can ChatGPT Chats Be Used in Court? What Rulings Show": https://jaroha.com/articles/main/can-chatgpt-chats-be-used-in-court-what-rulings-show-so-far
- Marketing Interactive, "OpenAI pushes back against order to hand over millions of ChatGPT convos": https://www.marketing-interactive.com/openai-pushes-back-against-order-to-hand-over-millions-of-chatgpt-convos