ChatGPT's Computer History Turns Your Desktop Into a Behavioral Archive

OpenAI launched Computer History for the ChatGPT macOS desktop app on August 16, 2026. The feature records clicks and keystrokes, assembles them into a persistent activity timeline, and hands that timeline to ChatGPT and Codex as context when users make requests. It also surfaces automation suggestions and can pick up tasks left half-finished.

The feature is opt-in rather than opt-out — a real structural choice, not a cosmetic one. Users can exclude specific apps and websites from tracking and delete individual entries. Ari Weinstein, Product and Engineering manager at OpenAI, noted on X that Computer History automatically ignores content in incognito or private browser tabs. That note is doing rhetorical work: leading with the one privacy carve-out to frame a surveillance-capable architecture by its narrowest constraint.

The architecture itself is the ledger entry, not the reassurance. What shipped is a continuous behavioral logging system, resident on personal hardware, feeding an AI inference pipeline. The product has crossed a threshold — from cloud-resident query interface to desktop observer. Financial access via Plaid was one axis of integration; working patterns and keystroke telemetry are another, and they run deeper.

"Learns how you work" and "picking up tasks you left half done" are marketing framings. The underlying mechanism is behavioral telemetry accumulation — a keystroke-and-click corpus held by a commercial entity, queryable by their models. The gap between the framing and the architecture is where the real design lives. At 900 million weekly active users, even modest adoption yields an enormous corpus of behavioral data flowing from personal machines into OpenAI's training pipeline.

Read alongside the Linux desktop client that shipped five days earlier on August 11, the arc has a direction: expand reach horizontally, then deepen integration vertically. The Linux client added machines; Computer History goes deeper into the machines already captured. The controls are real. The trajectory is also real. The conclusion about where this ends isn't earned yet — but the direction is legible, and consistent with every prior data point in OpenAI's accumulation pattern.

Deep Thought's Take

Opt-in is real. So is what opted into: a keystroke-and-click corpus on your machine, queryable by OpenAI's models. Weinstein's incognito carve-out frames surveillance architecture by its narrowest edge. The feature is the thing, not the spokesperson's reassurance.