ChatGPT for Mac swaps screenshots for keystroke logs, and hands IT a policy problem

OpenAI shipped Computer History in the ChatGPT desktop app for macOS, announced in a Thursday changelog entry ^[1], and it replaces the Chronicle research preview with a system that reads macOS accessibility events instead of taking screenshots ^[4][5][6]. The recorded events include clicks, typing, keyboard shortcuts and switching between apps ^[6], which makes this an endpoint activity log with an assistant on top, and therefore something IT, security and privacy owners have to write policy for rather than evaluate as a feature.

The mechanism matters more than the naming. Chronicle relied on screenshots ^[5]; Computer History subscribes to the interaction events macOS exposes through its accessibility features ^[6], and OpenAI describes the result as activity across chosen apps and websites turned into memories and a timeline that ChatGPT and Codex can use, with per-source selection, a pause control, and review or deletion at any time ^[2]. It is off by default and each user has to enable it ^[3]. The pitched uses are mundane: asking what you were doing before a break, finding a recently viewed document, generating a summary of the day, and suggestions to turn repeated workflows into skills or automations ^[18].

Eligibility and geography shape the work. The feature is limited to Pro, Business and Enterprise ^[7], and in Business and Enterprise workspaces an administrator must grant access before a member can turn it on ^[8], so two sequential authorizations sit between a managed Mac and collection ^[19]. Initial availability excludes the European Economic Area, Switzerland and the United Kingdom ^[9]. It arrives a month after Codex joined ChatGPT on the desktop ^[10].

On retention, OpenAI says raw interaction events are stored on the Mac for up to 48 hours ^[13], processed on its servers to generate memories, and not retained after processing or used for training ^[14], while the resulting plain-text memory files stay local until deleted ^[15]. The buffer expires; the derived record does not, so the artifact with the longest life is a local plain-text summary of what a person did, with no stated expiry ^[20]. Users can inspect timeline entries, reveal memory files in Finder, delete specific items, or clear the last 10 minutes, hour, day, or all time ^[12]. According to OpenAI, no screenshots, screen recordings, microphone input or system audio are captured, and private browsing is excluded ^[11].

Two of OpenAI's own warnings are the real policy inputs. The company says Computer History can increase prompt-injection risk from malicious content in apps or websites ^[16], which puts a text-consuming agent downstream of whatever a browser renders. And it recommends pausing collection during communications with other people without their express consent, and excluding apps that hold sensitive information ^[17]. That leaves the consent problem for other people's messages with the individual user, with pause and per-source exclusion as the listed controls ^[21].

Watch whether admin controls gain app-level or domain-level allow and deny lists instead of a single workspace switch ^[8][2]; whether the EEA, UK and Switzerland exclusion is lifted and on what basis ^[9]; and whether accessibility permission for an assistant becomes routine on managed fleets now that the requester is a productivity tool ^[6].