AI That Watches Your Screen So You Stop Repeating Yourself

OpenAI has shipped Computer History, a feature that watches what happens on your Mac so the assistant already knows what you are working on. The problem it solves is real and irritating: a large share of the time people spend with AI goes on re-establishing context it should already have. Solving that is genuinely valuable. It is also a categorically larger grant of access than a chat box, and the interface makes it feel like a small toggle. This is worth an actual decision, not a reflex in either direction.

The gap between remembering and watching

It is easy to file this alongside AI memory, and that undersells the change. Memory across conversations retains what you deliberately typed into the assistant. Screen observation captures what happens to be on your display, and on a working computer that is a much wider net: a client file open in another window, an email thread you had nothing to do with, a payroll spreadsheet minimized in the background, someone else’s personal information visible while you were helping them. The step from remembers what I told it to sees what I am doing is large, and it deserves more thought than the toggle implies. This is a different question from the one raised by AI persistent memory, and a closer cousin to always-on recording in the workplace.

Whether it matters depends on your screens

There is no universal answer here, which is precisely why it needs a decision rather than a default.

The right-hand column is where Canadian obligations bite. Privacy duties attach to client and personal information regardless of the route it took out of your control, so a screenshot is not meaningfully different from a form submission, a point we cover in PIPEDA-compliant AI.

Four questions before you switch it on

Answer these in writing, from the vendor’s own documentation. What exactly gets captured, and is it processed locally or sent to the provider? How long is it kept, and can you actually delete it? Who can access it, including whether an employer can review an employee’s captures, which is its own conversation. And what happens to third-party data that appears incidentally, the client details that were simply on screen? If you cannot answer all four, that is your answer for now. This is the same discipline that the reasoning-trace leak taught the hard way: know what the system retains before you feed it.

Tell your team, or lose them

If you do enable this, say so plainly and say what it means. Discovering after the fact that their screen was being captured is the kind of thing that permanently changes how people feel about the tools you give them, and about you. Openness costs nothing here and buys a great deal. It also tends to surface the practical objections early, the person who works on the most sensitive files pointing out that their machine should be excluded, which is exactly the input you want before rollout rather than after.

Capture the benefit on purpose

Do not let the privacy question obscure the real insight underneath the feature: an enormous amount of AI overhead is people re-explaining their business to a machine. You can address a good chunk of that without any screen capture at all, by writing your context down once and reusing it, which is exactly the case for treating prompts and context as company property. Ambient observation is where assistants are heading, and it will get more useful. Adopt it where your screens are safe, exclude it where they are not, and get the underlying benefit either way.

Frequently Asked Questions