OpenAI's Computer History turns keystroke logging into a ChatGPT setting

OpenAI announced a feature called Computer History on 13 August that records activity across approved macOS apps and websites, building a searchable timeline that ChatGPT and Codex can reference ^[1]. It tracks typing, clicks and keyboard shortcuts through macOS accessibility APIs and converts that raw stream into structured memories ^[2], which means the question of whether keystroke-level activity capture runs on company laptops is now settled inside an app's settings pane rather than by a security procurement cycle.

The pitch is continuity: pick up an abandoned half-finished task, get automation suggestions based on observed workflow patterns, and issue requests without restating context ^[13]. It replaces Chronicle, an earlier research preview that took periodic screen captures to do roughly the same job ^[3]. Cryptobriefing, whose write-up credits Gizmodo, frames the switch from screenshots to discrete interaction data as less intrusive by design, while noting it raises its own questions about keystroke logging and about what is stored, where, and for how long ^[14].

The controls are better than the category's track record. The feature is opt-in and off by default ^[6], users can pause collection, exclude specific apps and websites, and delete individual timeline entries ^[7], and administrators on team plans get their own access controls over how it operates across a workforce ^[8]. Ari Weinstein, a product and engineering manager at OpenAI, said on X that Computer History automatically ignores content viewed in incognito or private browser tabs ^[4]. Weinstein previously worked at Apple and led the team that built the feature as an evolution of the Memories system OpenAI introduced in 2024 ^[5].

Availability is where the governance work starts. The initial rollout covers Pro, Business and Enterprise tiers on macOS ^[9], so Plus and free users are not in scope at launch ^[1], and two of the three eligible tiers are organisational ones where admin controls apply rather than individual consent alone ^[2]. Users in the European Economic Area, Switzerland and the United Kingdom do not get access at launch, which Cryptobriefing attributes to the pattern of AI features colliding with GDPR ^[10]. That gives compliance teams in those jurisdictions a delay, not an answer.

The obvious comparison is Microsoft Recall, which also captures desktop activity into a searchable timeline and which drew enough privacy backlash after its 2024 announcement to force delays and redesigns ^[11]. Cryptobriefing's assessment is that OpenAI learned from that episode by shipping opt-in from day one and using accessibility APIs instead of screen grabs ^[12]. The operator's read should be less generous: a default-off toggle is a distribution strategy, not a data protection posture. Once the capability exists on every Pro seat, the exposure depends on how many employees flip it on inside apps that hold client data, credentials or regulated records, and on whether anyone has written a rule about it.

Three things to watch. First, retention and storage specifics, which the announcement as reported does not settle ^[14]. Second, whether admin controls on Business and Enterprise let a firm disable the feature outright and prove it, rather than merely manage it ^[8]. Third, whether EEA, Swiss and UK availability arrives and on what legal basis ^[10], because that will show what OpenAI thinks it can defend rather than what it can ship.