OpenAI Has a Cookie on 1,029 Websites and a Timeline of Your Desktop. Only One of Them Is Opt-In.

Two stories about OpenAI collided this week, and the internet has mashed them into one alarming sentence: ChatGPT is watching your browsing.

That sentence is not quite right, and the reason it is not right matters. There are two separate systems here. One is advertising infrastructure that follows you around the web whether or not you ever open ChatGPT. The other is a feature on your Mac that records what you do all day, which is off until you turn it on and then records considerably more than a cookie ever could.

Only one of them asked your permission.

The cookie nobody voted for

The first story starts with a researcher publishing months of traffic analysis. Buchodi monitored OpenAI’s advertising pixel in the wild and found it on a lot more of the web than most people would guess: 936 distinct pixels spread across 1,029 hostnames.

The mechanism is ordinary adtech. When you use ChatGPT, OpenAI can write an identifier into a cookie called __obi. That cookie then travels with you to any site running OpenAI’s advertising pixel. The pixel reports back what you looked at, what you read and whether you bought anything. If you are signed in to ChatGPT, that identifier can be connected to your account rather than floating anonymously.

The detail doing the real work here is the classification. OpenAI’s cookie policy lists __obi as an analytics cookie with a one year duration. Analytics is the category people wave through. It sounds like page counts and load times. Marketing is the category people reject. A cookie that connects your shopping behavior on a thousand other sites to your account is doing marketing work while sitting in the analytics drawer, and consent banners are built to respect the label rather than the behavior.

It is worth being precise about the limits of the finding, because the alarmed version of this story overshoots. Nothing in the research shows advertisers reading your conversations. OpenAI says its advertising system gives advertisers no access to chat history, and the traffic analysis produced no evidence to the contrary. What it shows is a conventional cross-site ad network, built by a company most people do not think of as an ad network, running at meaningful scale.

That last part is the story. OpenAI started putting sponsored agents inside ChatGPT conversations earlier this year, and an ad business needs measurement the same way any other ad business does. The pixel is what measurement looks like once it is deployed.

Computer History is the bigger one, and it is opt-in

The second system has almost nothing to do with the first, except that both involve a record of what you did.

Computer History arrived in OpenAI’s changelog on August 13, 2026 for the ChatGPT desktop app on macOS. It records interaction events through macOS accessibility features: clicks, typing, keyboard shortcuts, switching between apps. From that stream it builds summaries and a searchable timeline that ChatGPT and Codex can draw on when you ask them something.

The obvious comparison is Microsoft’s Windows Recall, and it is the comparison OpenAI invites. The difference is that Recall works by taking periodic screenshots of your screen. Computer History does not take screenshots. It reads the structure of what you are doing rather than photographing it.

Whether that is meaningfully less invasive depends on what you were doing. A screenshot of a password manager is a picture of a password manager. A log of which app you opened, what you typed and where you clicked is a different kind of exposure, and for some workflows it is a more useful one to an attacker, not a less useful one.

What it enables is genuinely convenient. You can ask what you were working on before lunch. You can ask where the document you had open yesterday actually lives. You can ask for a summary of your week for a standup. When the system spots a workflow you repeat, the timeline can offer to turn it into a reusable skill or a scheduled automation that Codex then builds for you.

The two systems, side by side

Rollout for Computer History is global for Pro, Business and Enterprise, with the European Economic Area, Switzerland and the United Kingdom following later. If you are in the EU and cannot find the setting, that is why.

And then there is the one that moves your mouse

A third thing gets folded into these headlines and deserves its own paragraph. On Windows, Computer Use lets ChatGPT operate the active desktop directly. OpenAI is blunt about the tradeoff in its own documentation: it cannot work in the background while you keep using the same Windows session, so expect ChatGPT to move the pointer, type, and take over the foreground while a task runs.

That is not surveillance. It is the opposite, in a sense, because you are sitting there watching it happen and cannot do anything else. But it is the clearest illustration of where this is going. The assistant is no longer a text box. It is a process with hands.

The same logic applies to the desktop app reading local files and interacting with installed software in ChatGPT Business. It requires explicit permission, workspace controls still apply, and actions that need approval surface on screen for you to review. Those are real guardrails. They are also guardrails that depend entirely on people reading the approval dialog rather than clicking through it, which is not how approval dialogs have historically gone.

How to turn each one off

Why the two got confused

These stories merged because they arrived within weeks of each other and both answer the same anxious question: how much does this thing know about me. The honest answer is that the amount it knows has less to do with any single feature than with how many separate doors you have left open, and most people have no inventory of those doors.

The cookie is the one you did not choose. It is also the smaller of the two, and the one with an existing, boring fix that also solves a hundred other problems. Blocking third party cookies is not a response to OpenAI. It is basic browser hygiene that happens to cover this.

Computer History is the one you do choose, and it is the one that would actually hurt if it leaked, which is a strange asymmetry to sit with. It is also the more defensible design, because it is off by default, it requires administrator sign-off in workplaces, and the record is yours to read and delete.

The pattern is familiar from the last time this came up. Human reviewers can read some ChatGPT conversations for quality work, and that opt-out also lives two menus deep rather than in front of you. None of this is hidden. It is just never on the path you walk. If you want the broader picture of what OpenAI retains and who can read it, we went through that in detail in our guide to whether ChatGPT is actually confidential.

What to take from this

• ChatGPT is not reading your browser history. An OpenAI advertising cookie is reporting your activity on sites that carry the pixel, which is a different and more ordinary thing
• The analytics label is the real finding. A cookie doing cross-site attribution filed under analytics is how consent tooling gets routed around, by design or by accident
• Computer History records far more, and asked first. No screenshots, but clicks and keystrokes at the structural level, and it stays off until you say yes
• Block third party cookies once and stop thinking about it. It is the highest value setting in this entire story
• If you work on a managed Mac, your administrator decides whether the timeline is even available, so that is the person to ask

The uncomfortable part is not that any one of these is outrageous. It is that the company building the assistant you talk to about your job, your health and your money has now also built the measurement layer that watches you shop. Those two businesses want very different things from the same identifier.