What does the feature collect?

OpenAI has added a feature called Computer History to the ChatGPT desktop app on macOS. It is opt-in, and once it is on the app builds a timeline out of what the user does. The Verge reports that what it stores are events, with no images, video or audio captured, and that both ChatGPT and Codex can consult the timeline when a request comes in. In the company's own demonstration, Dominik Kundel had the app find the last document he had edited, check whether it had been shared on Slack, and summarise how he had spent his morning.[1]

The controls are real and worth naming: a user can keep individual apps and websites out of the timeline and delete entries, and Ari Weinstein, a product and engineering manager at OpenAI, said on X that content in incognito or private browser tabs is ignored. All of that sits with the person at the keyboard. The question the announcement does not take up is who holds the switch when the Mac belongs to an employer. That is the setting in which a timeline of someone's working day changes meaning.[1]

Handing over is common, finishing is not

There is a measured baseline for how much work is already moving. Epoch AI and Ipsos surveyed 1,106 employed US adults between 10 and 19 July 2026, and 20 per cent of workers said they now give AI at least one task that used to go to a colleague or an outside contractor. Usage runs higher still inside tasks: 57 per cent in software development and 46 per cent in data analysis.[2]

The same survey shows where the work stops. About one in six AI-assisted tasks now takes longer than it used to, and respondents say they use 66 per cent of the output either unchanged or with only minor tweaks. The researchers read the result as a redistribution of tasks between people and AI, and they do not speak of full automation. What remains on the human side of that split is largely checking and correcting, and the sequence of actions Computer History is built to observe is exactly that person's.[2], [1]

Whose pocket does the efficiency get credited to?

In yesterday's column I argued that thinning out entry-level work drains the pool of people able to catch a model's mistakes. Computer History looks at the same problem from the other end: the checking and correcting that survey respondents still do by hand is precisely the work a timeline of clicks and keystrokes makes legible, and therefore automatable. There is a real efficiency gain forming here. At the same point, the question of who it gets credited to becomes concrete.[1], [2], [3]

Two indicators would settle it, and both are observable. The first is whether the timeline stays a personal convenience or becomes an administrator setting on managed devices; a feature the worker can switch off is a different object from one the employer switches on. The second is whether any share of the time the survey measures comes back to the worker: will it show up as shorter hours or higher pay, or only as a larger volume of reviewed output? Neither answer is in this announcement. Both would be in the next enterprise release notes and the next round of the same survey.[1], [2]