Is ChatGPT safe? For most everyday questions, yes: it is a mainstream service run by a large company, it does not install anything on your device, and the chat window is not a way for strangers to reach you. The real question is narrower and more useful: what happens to the words, files and photos you type or upload into an SI (AI) chat assistant, who can read them, and for how long. This guide answers that from OpenAI's own documents, the incidents that have actually happened, and the advice of a national cyber security agency. It is not legal advice, and the settings described here can change, so check the current version of each page before relying on it.
The short answer
ChatGPT is reasonably safe to use if you treat it like an email to a company rather than a private diary. Four facts sum up the situation:
- OpenAI receives everything you send. Its privacy policy lists your prompts and uploaded files as personal data it collects, alongside your account details.
- By default, consumer chats can help train future models. You can switch that off, and the switch is easy to find.
- Deleting is real but not instant. Deleted chats are removed from OpenAI's systems within 30 days, with exceptions for legal and safety reasons.
- Leaks have happened, but they were bugs and sharing mistakes, not the model repeating your secrets. We cover each case below.
"Safe" also covers other risks that have nothing to do with privacy: wrong answers given with confidence, and scams built with SI tools. We come back to those at the end.
What happens to what you type
OpenAI's privacy policy (version updated on 6 February 2026) calls everything you put into the service "Content": your prompts, the files and images you upload, and the rest of what you give it. Content sits next to account information such as your name, contact details, date of birth and payment information, and next to technical data about how you use the service.
The same policy says OpenAI may use that Content "to improve our Services, for example to train the models that power ChatGPT", and points to instructions for opting out. It also lists the people who may see personal data outside the company: hosting, cloud and customer service providers working on its instructions, affiliates, a buyer in case of a sale or restructuring, and government authorities or other third parties when the law requires it.
One line matters a lot at work. If you join a ChatGPT Enterprise or business account, the policy says the administrators of that account may access and control your account, including your Content. A work login is therefore not private from your employer.
Does ChatGPT learn from my chats and repeat them to others?
This is the fear most people have, and the answer is more reassuring than it sounds. The UK National Cyber Security Centre (NCSC) explained in a March 2023 blog post that a chat model is trained first and then queried, and that adding information to a question does not automatically put it into the model for other people to retrieve. The risk the NCSC does flag is different: the provider can see your queries, stores them, and may use them to build later versions, and stored queries can be hacked, leaked or accidentally made public. Its two recommendations are blunt: do not include sensitive information in queries to public chat models, and do not ask anything that would cause problems if it were made public.
The NCSC adds a point people rarely think about: a question can be sensitive because of who asks it and when. A chief executive asking how to lay someone off, or a person asking about a health or relationship problem, reveals something even without names. Many questions from the same login also add up to a profile.
How long OpenAI keeps your chats
According to the privacy policy, chats you keep stay in your account until you delete them. Once you delete a chat or your whole account, OpenAI says it removes the data from its systems within 30 days, unless it must keep it longer or the content has already been de-identified for training. The exceptions are listed plainly: content tied to a ban for breaking the usage policies, data under a legal obligation such as a subpoena, payment records for paid plans, and an audit record of your deletion request itself.
Temporary Chat is the closest thing to an off-the-record mode. OpenAI's data controls page says temporary chats do not appear in your history, do not create memories, are not used to improve its models, and may be retained for up to 30 days for safety purposes. If you choose to save one, it becomes a normal chat with your normal settings.
The court order that kept deleted chats
The 30-day rule was suspended for a while. In 2025, in the copyright lawsuit brought by The New York Times, a court ordered OpenAI to preserve consumer ChatGPT and API content, including chats users had deleted. OpenAI fought the order publicly. In an update dated 22 October 2025, the company wrote that its obligations under that order ended on 26 September 2025 and that it had returned to deleting chats within 30 days. It also wrote that it would keep a limited set of historical user data from April to September 2025, locked and accessible only to a small audited legal and security team. The order never covered ChatGPT Enterprise and Edu customers, and OpenAI says it is not keeping conversations from the European Economic Area, Switzerland or the UK under it.
The lesson is general: a deletion promise is only as strong as the legal context around it. A privacy policy can always be overridden by a court.
When ChatGPT data actually leaked
Two incidents show what the real risks look like. Neither involved the model "remembering" a user's secret.
March 2023: other users' chat titles and payment details
On 20 March 2023 OpenAI took ChatGPT offline after a bug in redis-py, an open-source library it used for caching, let some users see the titles of another active user's conversations, and possibly the first message of a new chat. In its write-up, the company said the same bug may have exposed payment-related information of 1.2% of ChatGPT Plus subscribers active during a nine-hour window: name, email address, payment address, card type, the last four digits of the card and its expiry date. Full card numbers were not exposed, according to OpenAI. The Italian data protection authority later criticised OpenAI for not notifying it of this breach.
July 2025: shared chats in search results
ChatGPT lets you create a public link to a conversation. TechCrunch reported on 31 July 2025 that searching Google and other engines for the chatgpt.com/share address turned up strangers' conversations, some with personal details. The links existed only because users had clicked "share" and then "create link", and an option to make them discoverable by search engines was part of what OpenAI called an experiment. Hours after the article, OpenAI said it had removed the feature because it introduced too many opportunities for people to share things they did not intend to. The general rule stands: anything you turn into a public link is public.
Using ChatGPT at work
Companies worry less about the model and more about employees pasting in confidential material. In May 2023, TechCrunch reported, citing a memo seen by Bloomberg, that Samsung temporarily restricted generative SI tools on company devices after internal, sensitive data was accidentally leaked to ChatGPT in April. Samsung told staff using such tools elsewhere "not to submit any company-related information or personal data", and noted that data sent to external servers is hard to retrieve and delete.
OpenAI separates consumer and business plans on this point. Its data controls page says that by default it does not use content from ChatGPT Business, Enterprise, Edu or ChatGPT for Healthcare workspaces to train its models, and that those workspaces can have their own retention and access settings. If your employer provides one of those plans, use it for work instead of a personal account, and follow the company's policy on what may be pasted in.
What regulators have said
Europe has been the most active. In March 2023, the Italian data protection authority, the Garante, temporarily blocked ChatGPT in Italy, and it then opened an investigation. In a decision dated 2 November 2024, it found that OpenAI had trained ChatGPT on users' personal data without first identifying an adequate legal basis, had fallen short on transparency, had not notified the March 2023 breach, and lacked age verification to protect children under 13. It imposed a fine of 15 million euros and ordered a six-month public information campaign. That decision did not stand: the authority's own page now says it was removed after a judgment of the Court of Rome, published on 18 March 2026, upheld the appeal against it.
So the fine was annulled, but the questions raised (legal basis for training, transparency, age checks) are the same ones every regulator now asks about chat assistants. OpenAI's privacy policy says its services are not intended for children under 13.
Five settings and habits that make ChatGPT safer
| What to do | What it changes | What it does not change |
|---|---|---|
| Turn off "Improve the model for everyone" | New chats are not used to train OpenAI models | Chats stay in your history and OpenAI still receives them |
| Use Temporary Chat for one-off questions | No history, no memory, no training | May be kept up to 30 days for safety |
| Delete chats you no longer need | Removed from OpenAI's systems within 30 days | Legal holds and ban records are exceptions |
| Avoid public share links for anything personal | Keeps the chat off the open web | Anyone with the link can still read it |
| Keep secrets out of the prompt | Nothing sensitive is stored at all | Nothing: this is the only complete protection |
Two details from OpenAI's help page are easy to miss. Turning off model training does not delete or hide your saved chats; deleting is a separate action. And if you give feedback with a thumbs up or thumbs down after opting out, the whole conversation linked to that feedback may still be used to train models. Archiving a chat changes nothing about training either.
Never paste passwords, full card numbers, identity documents, medical records, client files or anything covered by a confidentiality agreement into a consumer chat assistant. If you need help with a document like that, remove names and numbers first.
Safe in other senses: wrong answers and SI scams
Privacy is not the only way a chat assistant can hurt you. The NCSC's post lists other serious flaws of large language models, including that they can get things wrong and present incorrect facts confidently. Our guide to SI hallucination explains why chat models invent sources, quotes and numbers, and how to check them. For health, legal or money decisions, treat an answer as a starting point and verify it with a qualified source.
The other risk comes from people using these tools against you. The NCSC already warned in 2023 that criminals could use language models to write convincing phishing emails, including in languages they do not speak well. Fake profiles, cloned voices and fabricated photos follow the same logic, and we describe the common patterns in our article on SI scams. When a message, a picture or a document looks too convenient, a quick check with SI or Not can tell you whether it was probably made with SI before you act on it.
So, is ChatGPT safe?
For drafting, brainstorming, explaining a topic or fixing a spreadsheet formula, yes, with sensible settings. It is not a vault: what you type travels to a company, may be read by its systems and, in specific cases, by people or courts, and may be kept for a while after you delete it. The biggest leaks so far came from software bugs and public links, not from the model. Keep sensitive data out, switch off training if you prefer, use Temporary Chat for one-off questions, and verify any answer that matters. For the vocabulary behind all this, from large language models to training data, see our guide What is SI?
