ChatGPTの「Computer History」機能がクリックやキー入力を記録・学習ChatGPT’s Computer History tracks your clicks and keystrokes

OpenAI appears to be pushing further into agentic computing with a new feature in ChatGPT's macOS desktop app called Computer History, which logs a user's on-screen activity and turns it into data the assistant can learn from. According to The Verge, the feature records actions such as clicks and keystrokes to construct an activity timeline, then uses that record to suggest automations, learn how a person works, and even resume tasks that were left half finished. The move is notable because it shifts ChatGPT from a tool you actively prompt toward one that continuously observes what you do, a change that carries clear privacy implications.

The core mechanism is the timeline. Rather than treating each conversation as an isolated session, Computer History builds an ongoing record of a user's behavior that both ChatGPT and Codex, OpenAI's coding-focused system, can read and reference. In practice, that means the assistant is meant to develop a working memory of routines and workflows: which applications you move between, the steps you repeat, and the tasks you commonly perform. From that history, the app can propose automations for recurring work or pick up an unfinished job where it was left off. OpenAI frames the captured activity as training data, indicating the behavior is intended to improve how the assistant models a given user's habits.

That framing is also the source of the most significant concern. Continuously logging clicks and keystrokes means the app is capturing a granular stream of what a person types and interacts with, which can include sensitive material well beyond the boundaries of any single task. The description provided does not detail whether the data is processed locally on the Mac or sent to OpenAI's servers, how long timelines are retained, or what controls users have to pause, review, or delete the history. Those specifics matter a great deal for evaluating the feature's risk, and until they are clarified, characterizations of exactly how the data is stored and used should be treated cautiously.

The concept has close and instructive precedents. Microsoft drew heavy criticism in 2024 when it announced Recall, a Windows feature that periodically captured screenshots to create a searchable timeline of user activity; the backlash over security and privacy led Microsoft to delay the rollout, add encryption and authentication safeguards, and make the feature opt-in. Computer History invites similar scrutiny because it addresses the same fundamental tension: an assistant becomes more useful the more it observes, but broad observation of personal computing activity is inherently sensitive. How OpenAI handles consent, transparency, and data handling is likely to shape how the feature is received.

The feature also fits a broader industry direction toward agents that can act on a computer rather than merely answer questions. OpenAI has previously experimented with agentic capabilities, including tools designed to browse and operate software on a user's behalf, and Codex reflects its interest in systems that can carry out multi-step technical work. Rivals are pursuing comparable goals, with Anthropic having demonstrated a "computer use" capability that lets its Claude models move a cursor and interact with interfaces, and Google building agentic functions into its Gemini products. Computer History extends this trajectory by grounding an agent in a persistent, personalized understanding of how an individual actually works, which is the kind of context automation systems have generally lacked.

For now, the feature is described as specific to the macOS desktop app, and it is unclear from the available information whether it is broadly available, limited to a subset of users, or gated behind a particular subscription tier. Readers should also note that details can change as OpenAI refines the feature and responds to feedback. What is clear is the design intent: to convert everyday computer use into a resource the assistant can draw on for proactive help. Whether users embrace that trade-off will likely depend on the granularity of the controls OpenAI provides and how transparently it explains where the recorded activity goes and how it is used to train its models.