OpenAI shipped a feature this week that records your screen activity, your keystrokes, your app switches, and your document text into a local timeline it calls Computer History. The idea is that ChatGPT can now answer questions about what you were doing two hours ago, what you copied from that PDF, or which tab you had open when you wrote that email.
It is genuinely useful. It is also the kind of thing you should think about before you turn it on.
Computer History runs as a background process. It captures periodic screenshots and logs application events. That data is stored locally on your machine, not sent to OpenAI's servers in real time. The AI reads from the local store when you ask it something that requires context from your recent activity.
On paper, local storage sounds reassuring. In practice, the reassurance is thinner than it looks. The timeline is stored unencrypted by default. Anyone with physical or remote access to your machine can read it. That includes the IT team at a client's office, a device management tool your employer runs, or malware that got in while you were doing something else. The data is not protected at rest in any meaningful way unless you add that layer yourself.
There is no selective recording. The feature does not know the difference between your personal banking session and a competitor's pitch deck you opened for research. It records both.
The most privacy-sensitive thing about this feature is not what OpenAI does with the data. It is what sits unencrypted on your hard drive, readable by anyone who gets there first.
Agency work is context-dense by nature. At any given moment, a creative director's screen might hold an unreleased brand identity, a client's internal revenue figures shared under NDA, a recruitment conversation, and three browser tabs of competitor research.
All of that becomes AI context if Computer History is running.
The immediate risk is not a data breach in the dramatic sense. The risk is that a feature designed to make you more productive quietly accumulates a record that was never meant to exist. Client confidentiality agreements rarely contemplate this kind of ambient logging. If you work under a data processing agreement with a client, you have obligations about what systems touch their data. A local AI timeline that ingests their documents may not be covered. That is worth a legal conversation before you enable the feature, not after.
There is also the question of what happens when that machine leaves your hands. A laptop returned to a client, sold, repaired, or simply lost takes the timeline with it unless you have explicitly cleared it.
I want to be clear. I understand why this feature exists and why people will use it. The problem of lost context is genuine. How much time does a PM spend reconstructing what was decided in a document they cannot find? How often does a designer repeat a research step because they cannot remember where they saved the output?
A local AI that can answer "what did I read about that competitor last Tuesday" is not a trivial thing. It addresses a real friction point. The productivity gain is not imaginary.
But productivity and privacy are not automatically at odds here. The issue is that OpenAI shipped the feature without encryption at rest as a default. That is a product decision, not an inevitability. Encryption does not prevent the AI from reading the local timeline. It just means that reading requires authentication. The tradeoff is almost entirely on the implementation side, not the conceptual side.
Until that changes, using Computer History at work is a calculated risk, not a free lunch.
If you run an agency or manage a team that uses AI tools, this is a good moment to get ahead of it.
First, check whether anyone on your team has already enabled Computer History. It launched quietly and some people will have turned it on without reading the details.
Second, decide whether your current client agreements and data processing arrangements allow ambient screen logging. If you are not sure, assume they do not and ask legal.
Third, if you do want to use the feature, add disk encryption at the OS level. FileVault on macOS and BitLocker on Windows both encrypt the local drive, which closes the unencrypted-at-rest gap without requiring any changes to how ChatGPT works.
Fourth, consider a usage policy. Not a policy in the sense of a long document nobody reads. A short, direct team norm: what categories of client data are off-limits for AI context features, full stop.
None of this requires abandoning the feature. It requires treating it as infrastructure rather than a free plugin.
Computer History is one feature in a wider shift. AI tools are moving from discrete actions, you ask, it answers, toward continuous ambient presence. They observe, accumulate, and surface context you did not think to save. That shift has real value. It also changes the threat model in ways that standard cybersecurity advice was not written for.
The industry's default posture has been to assess AI tools feature by feature: is this safe to use? The better question is systemic: what does this feature know, where does that knowledge live, who can reach it, and what happens when something goes wrong?
For agencies in particular, that question is not abstract. It sits inside every NDA, every client relationship, every device policy. Getting ahead of it is not a compliance exercise. It is part of doing the work properly.
