The Verge is reporting that OpenAI has rolled out a feature called Computer History to the macOS desktop version of ChatGPT, one that monitors user clicks and keystrokes to build a running timeline of activity on the machine. That timeline can then be drawn upon by ChatGPT and Codex to suggest automations, resume unfinished tasks, and, according to The Verge's summary, effectively learn the behavioral patterns of whoever is sitting at the keyboard.
To understand why this lands with such weight, it helps to situate it within a longer arc. OpenAI has spent the past two years methodically expanding ChatGPT from a conversational novelty into something closer to an operating-system-level assistant. The release of custom GPTs, the gradual integration of browsing and code execution, and the push into voice and vision have all pointed in the same direction: a product that does not simply answer questions when asked, but anticipates needs and embeds itself into the daily rhythm of work. Computer History, as described by The Verge, is a significant step further along that road. Where previous features responded to prompts, this one watches and records before any prompt is issued.
The competitive context matters here too. Microsoft has been attempting something structurally similar with its Recall feature for Copilot Plus PCs, which generated a firestorm of criticism when it was revealed that Windows would be taking continuous screenshots of user activity and making them searchable. Microsoft was eventually forced to delay the rollout and rework the feature's privacy architecture under sustained pressure from security researchers and regulators. OpenAI is now entering essentially the same territory on a rival platform, and the parallels are close enough that the company will almost certainly face the same line of questioning.
The key tension in any feature of this kind is the gap between its genuine utility and the surveillance infrastructure it requires to deliver that utility. The pitch is persuasive on its face: a system that knows a user left a spreadsheet half-built or a draft email unsent, and can step in to help complete it, addresses a real friction in knowledge work. The mechanism that makes it possible, however, is persistent logging of input at a granular level. Keystrokes and clicks are not abstract metadata. They can reconstruct passwords entered into fields that do not mask input, capture the content of private messages typed in any application, and reveal behavioral patterns that carry genuine security implications if that data were ever exposed or misused.
OpenAI occupies a particularly scrutinized position when it comes to training data practices. The company has faced sustained criticism and several ongoing legal challenges over the data used to train its models, and regulators in the European Union and elsewhere have made clear they are watching closely. Framing a keystroke-logging feature as something that turns user actions into training data, as The Verge's summary puts it, is likely to attract immediate attention from privacy advocates and, potentially, from data protection authorities. The distinction between data used to personalize a session and data used to improve the underlying model is one that will matter enormously to regulators, and the current description does not make that line obviously clear.
For enterprise users and businesses that have standardized on ChatGPT through OpenAI's commercial tiers, the implications are sharper still. Corporate IT and security teams generally maintain strict policies about what software may observe and log activity on managed devices, and a feature like this would typically require explicit review and sign-off before deployment. The likely reading is that OpenAI will offer opt-in or opt-out controls, as it has done with similar sensitive features, but the default state of those controls will be the thing privacy advocates watch most closely.
For individual consumers, the consequences are more diffuse but no less real. Users who adopt the feature in good faith and find genuine value in its automation suggestions will be extending a significant degree of trust to a company whose data handling practices remain an active subject of public and regulatory debate. That is not a reason to assume the feature is malicious, but it is a reason to read the fine print more carefully than most people do.
Several things are worth watching as this develops. How OpenAI describes the data handling in its terms and privacy documentation will tell much of the story, particularly whether keystroke and click data is processed locally or transmitted to OpenAI's servers, and how long any resulting timeline is retained. Regulatory reaction in Europe will be an early indicator of whether this feature survives in its current form internationally. And Microsoft's response is worth noting too: if OpenAI normalizes this category of ambient monitoring without the backlash Recall attracted, it may clear a path for Microsoft to revive its own ambitions in the space with considerably less friction.
