Is ChatGPT safe? What happens to your data, and how to keep it private

The short answer

ChatGPT is secure in the sense that matters to an engineer: traffic is encrypted, the infrastructure is audited, and OpenAI has a security team that responds to incidents. It is not private in the sense that matters to you. On the free and paid personal plans, what you type is stored on OpenAI's servers, can be read by authorised staff, and is used to train future models unless you switch that off.

So the honest answer is this: ChatGPT is safe to use, as long as you treat every prompt as something you are handing to a third party. Anything you would not paste into an email to a stranger should be removed or replaced before it goes in. The rest of this guide covers what is actually stored, who can see it, what has gone wrong before, and the seven settings that change the picture.

What ChatGPT does with your conversations

Four things happen to a prompt after you press Enter on a personal plan.

It is stored. Every conversation is saved to your account history until you delete it. Deleted conversations disappear from your account immediately, and OpenAI says they are removed from its systems within 30 days unless a legal obligation requires a longer hold. Temporary Chats are the exception: they never enter your history, and OpenAI keeps them for up to 30 days for safety review before deleting them.

It may be used for training. On Free, Go, Plus and Pro, your conversations and any memories ChatGPT has saved about you can be used to improve OpenAI's models. This is on by default. You can turn it off in Settings, under Data Controls, with the switch called "Improve the model for everyone". Business, Enterprise, Edu and API customers are opted out by default.

It may be read by a person. OpenAI's policy allows a limited number of authorised employees and contracted service providers to view conversations for abuse investigations, support requests, legal obligations and, on personal plans, model improvement. This is the part most people underestimate: the model is not the only reader.

It may be remembered. With Memory on, ChatGPT extracts facts from your chats, such as your job, your family and your projects, and reuses them in later conversations. Memories are stored on your account, listed in Settings under Personalization, and can be deleted one by one or turned off completely.

Who can see your ChatGPT conversations

• You, in your history and in a full data export.
• OpenAI staff and vendors, under the conditions above.
• Anyone with a share link. A shared conversation is readable by whoever holds the URL, and it stays live until you delete the link under Data Controls. In July 2025 roughly 4,500 shared chats showed up in Google results because an optional "make this chat discoverable" toggle let search engines index them. OpenAI removed the feature within days, but the lesson stands: a share link is a publication.
• Your workspace administrators, if you use ChatGPT through a Business, Enterprise or Edu workspace. Admins control retention and can export conversation data for compliance.
• Courts and regulators. In May 2025 a US court ordered OpenAI to preserve all consumer chat logs, including deleted ones, as evidence in the New York Times copyright case. The order was lifted for new conversations in late September 2025 and never applied to accounts in the EU, UK or Switzerland, but logs collected during those months are still held. Deleting a chat is a request, not a guarantee.

Nobody else can see your conversations by default. Other users cannot, and neither can the websites you paste text from.

Is ChatGPT secure?

Technically, yes. Data is encrypted in transit with TLS 1.2 or higher and at rest with AES-256. OpenAI holds a SOC 2 Type 2 report, runs a bug bounty programme, and offers multi-factor authentication on every account.

The incidents that have happened were mostly not break-ins:

• March 2023. A bug in a caching library briefly showed some users the titles of other people's conversations, and exposed partial payment details of about 1% of Plus subscribers.
• April 2023. Samsung engineers pasted internal source code and meeting notes into ChatGPT. Nothing was breached; the data simply became OpenAI's to store. Samsung banned the tool internally.
• December 2024. Italy's data protection authority fined OpenAI 15 million euros over the legal basis for training and the age checks in place at the time.
• July to August 2025. The shared-link indexing incident described above.
• Ongoing. Fake "ChatGPT" browser extensions that steal session cookies, and phishing pages that imitate the login screen. These target you, not OpenAI.

The pattern is clear. The risk is rarely that someone breaks into OpenAI. The risk is what you put in, and what a link, a setting or a subpoena later does with it.

Is ChatGPT safe for confidential information?

On a personal plan, no. Client data, patient records, employee details, unreleased financials, source code under NDA and anything covered by GDPR, HIPAA or a confidentiality agreement should not go into a Free, Plus or Pro conversation in its raw form. You would be transferring the data to a US processor without a contract, for an undefined retention period, with human review possible.

Three things change that answer:

1. A business plan. ChatGPT Business, Enterprise and Edu come with a data processing agreement, no training by default, admin-controlled retention and, in Enterprise, the option of EU data residency. If your organisation has one, use it, and never your personal login.
2. The API with zero data retention. Developers building on the API can request zero data retention for eligible endpoints, so prompts are not stored at all after the response is returned.
3. Redaction before the prompt leaves your browser. If you are on a personal plan, or you do not fully trust the plan you are on, remove the identifying details before you send. Replace names, emails, account numbers and addresses with placeholders, and put the real values back into the answer yourself. This is exactly what the Caviard extension automates for ChatGPT and DeepSeek: detection and substitution happen on your machine, and the originals never leave it. To see what that looks like on your own text first, run it through the free prompt anonymizer, which works in the browser without sending anything anywhere.

A useful rule for teams: ChatGPT may see the shape of your problem, but not the identities in it.

How to make ChatGPT more private: seven settings

Each of these takes under a minute. All of them live in Settings, which opens from your profile picture.

1. Turn off training. Data Controls, then switch off "Improve the model for everyone". This applies to new conversations only.
2. Use Temporary Chat for anything sensitive. Click "Temporary" at the top of a new chat. No history, no memory, no training, deleted after 30 days.
3. Review Memory. Personalization, then Memory. Delete what you do not want kept, or turn it off. Saved memories are used for training if the training switch is on.
4. Audit your shared links. Data Controls, then Shared links. Delete any you no longer need, and treat every new share link as public.
5. Turn on multi-factor authentication. Security, then Multi-factor authentication. Your history is only as private as your login.
6. Export, then delete, old history. Data Controls, then Export data if you want a copy, then delete the conversations you no longer need. Bulk deletion is under Data Controls as well.
7. Check which account you are in. If your employer provides a workspace, make sure the workspace name is showing before you paste work data. The personal and business versions look almost identical.

For screenshots of each step, see the ChatGPT privacy settings walkthrough.

What happens to your data, by plan

Plan	Used for training by default	How long it is kept	Who can access it
Free, Go, Plus, Pro	Yes, until you turn it off	Until you delete it, then up to 30 days	You, OpenAI authorised staff
Temporary Chat (any personal plan)	No	Up to 30 days	OpenAI safety review only
Business (formerly Team), Edu	No	Until deleted; admin controls apply	You, workspace admins, OpenAI under the DPA
Enterprise	No	Set by your admins; EU data residency available	You, workspace admins, OpenAI under the DPA
API	No	Up to 30 days for abuse monitoring, or none with zero data retention	The developer, OpenAI abuse monitoring

When you have to use real data anyway

Most of the value of ChatGPT comes from giving it real context: the actual email thread, the actual contract, the actual spreadsheet. Telling people to never paste anything sensitive fails within a week. What works is making the safe path the easy path.

• For individuals and small teams. Keep the personal plan, turn off training, and install a redaction layer in the browser. Caviard highlights personal data as you type, swaps it for realistic placeholders before the prompt is sent, and swaps the originals back into the reply with one keyboard shortcut. The free plan covers up to 10 protected values per document, and a one-time $9 lifetime licence removes the limit.
• For companies. Buy Business or Enterprise, sign the DPA, set retention, and still redact. A DPA governs what OpenAI may do; it does not stop an employee pasting a customer list into the wrong account. Our enterprise redaction guide covers the rollout.
• For regulated data. Run a local model, or use the API with zero data retention inside your own infrastructure. ChatGPT's consumer product is not designed for patient records or payment card data, whatever plan you are on.

If you only remember one thing: the settings decide what OpenAI may do with your data, but only you decide what goes in.