ChatGPT Privacy: What OpenAI Collects and Who Can See It

Most of what you will read about ChatGPT privacy is out of date, and not by a little. Three claims show up in nearly every guide published this year. Your deleted chats are held indefinitely by court order. OpenAI does not show you ads. You manage what ChatGPT remembers through a list of saved memories. All three were true once. None of them is true now.

This is what OpenAI actually collects in August 2026, who can reach it, and which settings are worth changing this afternoon. That includes the new macOS feature logging your clicks and keystrokes, the quiet link between memory and the ads ChatGPT now serves, and the three separate rules that apply in Europe and the UK. Every figure here comes from OpenAI's own documentation rather than from another blog's summary of it.

The Key Takeaways

• Deleted chats are gone in 30 days again. The court order that forced indefinite retention ended on 26 September 2025. OpenAI resumed normal deletion and confirmed it on 22 October 2025.
• ChatGPT has ads, and from 24 August 2026 so does Europe. The US test began on 9 February 2026 on the Free and Go plans. Ads reach 31 European markets on 24 August 2026. Plus, Pro, Business, Enterprise and Edu stay ad free.
• Your memory feeds those ads. If memory is on and personalization is on, ChatGPT may use saved memories and recent chats to pick which ad you see.
• A Mac feature now logs clicks and keystrokes. Computer History shipped on 13 August 2026. It is off by default and OpenAI warns it raises prompt injection risk.
• Europe and the UK are carved out of all three. Different rules on retention, on Computer History, and on ad personalisation, which OpenAI is not switching on in the EEA or Switzerland to begin with.

What ChatGPT Actually Collects

There are four buckets, and they behave differently. Account data is your name, email address and any accounts you connect. Conversation content is everything you type or paste, plus files and images you upload. Technical data is your IP address, device type, browser and approximate location. Usage data is how often you sign in, which features you touch, and whether you rate a response up or down.

The fourth bucket catches people out. Rating a response is not a neutral gesture. A thumbs up or thumbs down hands that specific exchange to OpenAI for review even when you have turned model training off everywhere else. If you have opted out of training and still rate responses, you have opted back in one conversation at a time.

Files behave differently again. They persist independently of the chat that carried them, so deleting a conversation does not remove the document you attached to it. The same is true of memory, which is where most of the confusion in this topic lives.

Who Can Actually See Your ChatGPT Conversations

Four groups, and it is worth being precise about each. You, through your history. Authorised OpenAI staff and contractors, who can review content flagged by abuse and safety systems. A small audited legal and security team, for the specific set of data still under legal hold. And anyone you send a shared link to, which is the leak most people actually experience, because a shared link stays live until you revoke it.

Some reassurance is warranted. OpenAI does not sell conversations, and advertisers have no access to your chats, history, memories or personal details. Advertisers receive aggregated reporting such as total views and clicks, nothing that identifies you.

On the question people ask most: your employer cannot read your personal ChatGPT account. A Business or Enterprise workspace is a different matter, because those plans include administrator controls and compliance tooling by design. If you are signed into a work account, treat the conversation as work correspondence.

Your Deleted Chats Are Not Held Forever Any More

In May 2025 a federal court in the New York Times case ordered OpenAI to preserve every conversation, including ones users had deleted. That order was real, and for four months the standard advice was correct. It is now wrong, and it is still being repeated.

OpenAI's obligations under that order ended on 26 September 2025, and the company confirmed the return to standard practice on 22 October 2025. Deleted conversations and Temporary Chats are removed within 30 days. API data is deleted after 30 days.

Two caveats remain. A limited set of April to September 2025 data is still stored under legal hold. It is locked to a small audited legal and security team, and it has not been handed to the New York Times or the court. And conversations originating from the European Economic Area, Switzerland or the United Kingdom were excluded from the retention demand in the first place. ChatGPT Enterprise, Edu and Zero Data Retention API customers were never covered at all.

Computer History: What ChatGPT Now Records on Your Mac

On 13 August 2026 OpenAI shipped Computer History in the ChatGPT desktop app for macOS. It turns your activity across apps and websites into memories and a timeline that ChatGPT and Codex can draw on. It replaces Chronicle, the screenshot based experiment from April 2026, and it is a rebuilt system rather than a rename.

What it records, and what it does not

It logs interaction events surfaced through macOS accessibility permissions: clicks, typing, keyboard shortcuts, and switching between apps. It does not capture screenshots, screen recordings, microphone input or system audio, and private browsing activity is excluded. Interaction events sit on your Mac for up to 48 hours, get processed on OpenAI's servers, and are not retained after processing or used for training. The memory files it produces are plain text and stay on your Mac until you delete them.

Who can turn it on

It is off by default and available to Pro, Business and Enterprise accounts. Business and Enterprise administrators have to grant access before anyone in the workspace can enable it. At launch it is not available in the European Economic Area, Switzerland or the United Kingdom. You can choose which apps and websites contribute, pause collection, and review or delete entries whenever you like.

OpenAI's own documentation is unusually direct about the downside. It warns that Computer History raises the risk of prompt injection from malicious content in the apps and websites you visit. It also recommends pausing collection during conversations with people who have not agreed to be recorded. That second warning is the one to take seriously, because the person on the other side of your screen share never consented to anything.

If your Mac app has been behaving strangely since the July rebuild, that is a separate issue covered in our guide to the rebuilt ChatGPT desktop app.

Memory, Ads, and the Link Between Them

ChatGPT started showing ads in the US on 9 February 2026, on the Free and Go plans only. Plus, Pro, Business, Enterprise and Edu accounts do not see them. Ads appear below a response, labelled as sponsored, and they do not influence what ChatGPT tells you.

On 24 August 2026 that test stops being American. Ads switch on across 31 European markets, the 27 EU member states plus Iceland, Liechtenstein, Norway and Switzerland, on the same Free and Go plans. The carve-out that matters is what OpenAI is not turning on: personalised ads are not initially available in the EEA or Switzerland. European ads are picked from the conversation in front of you, your language and a coarse location, not from your memory, your past chats or your ad history. That is the opposite of the American default, where personalisation runs unless you switch it off.

Here is the part that rarely gets mentioned. If ad personalization is on and memory is on, ChatGPT may use your saved memories and reference recent chats when it selects an ad. Your memory is not just personalising your answers. On the free tiers it is helping to target advertising. Turning personalization off does not remove ads, it narrows them to the topic of the current thread. Our guide to using ChatGPT without ads covers every route that does remove them.

Memory itself has been rebuilt, which is why older guides send you to a settings screen that no longer works that way. There is no list of individually saved memories to prune now. There is a memory summary you can read and edit directly, at Settings then Personalization then Memory, which our full guide to ChatGPT memory walks through step by step. If you want context sealed off rather than pooled, project-only memory keeps a workspace out of your account-wide memory entirely. Two behaviours matter more than the interface:

Deleting memory does not delete your chats. Choosing "Delete and turn off memory" clears the summary and stops new memories forming, but your conversation history stays exactly where it was. And if you switch memory back on later, ChatGPT may build fresh memories from the chats still sitting in that history, including old ones. To remove something for good, you have to clear every source it appears in: past chats, archived chats, uploaded files, the memory summary, and any connected apps. Curious what it has already worked out about you? Try the prompts that make ChatGPT show its notes.

Every ChatGPT Privacy Setting and What to Set It To

The training toggle is the one people get wrong most often, and it deserves more room than a table row. We walk through it for ChatGPT, Claude and Gemini in the full guide to opting out of AI training, including the myths about what opting out does and does not reach.

ChatGPT Privacy in Europe, the UK and Switzerland

Three separate rules land on the same set of readers, which is why this deserves its own section rather than a footnote.

Retention. Conversations originating from the European Economic Area, Switzerland or the United Kingdom were excluded from the New York Times retention demand. The indefinite hold that worried everyone in 2025 never applied to them.

Computer History. Not available in those three regions at launch. If you are in Europe and cannot find the setting, nothing is broken.

Ads. The consent model is inverted. In the US, personalized ads run by default and you opt out. In the EEA and Switzerland, where ads arrive on 24 August 2026, OpenAI is not making ad personalization available to begin with, so there is nothing to opt out of and nothing that reads your memory. You still see ads, drawn from the current thread.

The pattern is consistent. European data protection law makes the aggressive default expensive, so the aggressive default does not ship there. It is worth knowing which side of that line you are on before you assume a guide written for a US audience describes your account.

ChatGPT Privacy Compared to Claude, Gemini and Fello AI

The honest summary is that no cloud assistant is private in the way a local text file is private. What differs is how much is collected by default, how much you can switch off, and whether the app on your Mac wants accessibility permissions to watch what you do. Fello AI sits on the quieter end of that scale: one native Mac app that reaches every major model, with no ad system to feed and no activity logger to enable. If you want the broader version of that argument, we cover it in using AI without giving up your privacy.

Conclusion

ChatGPT is not the surveillance machine the scarier headlines describe, and it is not the private notebook the marketing implies. It is a product with defaults set for OpenAI's benefit rather than yours, and you can change almost all of them in about four minutes.

Do these four things now. Turn off "Improve the model for everyone". Turn off ad personalization if you are on Free or Go. Read your memory summary once and decide whether you are comfortable with it. Leave Computer History off unless you have a concrete reason to want it. Then use Temporary Chat for anything you would not paste into a work email, and stop worrying about the rest.

FAQ