Title: OpenAI Launches Computer History to Track Desktop Activity via ChatGPT

OpenAI rolled out Computer History, a new feature inside the macOS ChatGPT app that records clicks, keystrokes and other event-level signals. The model can reference a user’s recent desktop activity without ever taking screenshots. By turning raw interaction data into a timeline, the feature aims to shift ChatGPT from a reactive chatbot to a proactive assistant that can fetch the last edited document, confirm whether it was shared on Slack, or deliver a morning productivity recap—all while keeping visual capture out of the loop.

How the feature works

Computer History builds a chronological log the moment the app runs. Each entry notes which window was active, which button was pressed, and what text was typed, but it never stores images, video or audio. The log stays on the device and is exposed to the large language model only when a user starts a ChatGPT session. In practice, the model can query the timeline to answer “What file did I just edit?” or “Did I send that draft to a colleague?” without the user having to copy-paste files or describe the context manually.

OpenAI’s developer team demonstrated a workflow where the AI located the most recent document, checked its sharing status on a messaging platform, and then produced a concise summary of the day’s activities. By stitching together signals from multiple apps, the assistant can suggest automations for repetitive steps or nudge the user to finish an unfinished task.

Privacy-first architecture

The most visible difference from competing offerings, such as Microsoft’s Recall feature, is the data type collected. Recall leans on frequent screenshots to reconstruct a visual history; Computer History relies exclusively on metadata-rich “events.” OpenAI says the system automatically skips any activity that occurs in incognito or private-browsing modes, and it never captures pixel data that could reveal sensitive visual information.

Ari Weinstein, OpenAI’s product and engineering lead, emphasized that the feature is opt-in. Users must explicitly enable the history logger, and they can fine-tune what gets recorded:

  • App exclusion – blacklist specific programs or websites so they never appear in the log.
  • Manual deletion – prune individual entries to erase particular actions from the model’s memory.
  • Granular controls – toggle the logger on or off for particular time windows, ensuring the assistant only sees data when the user wants it to.

These safeguards aim to reassure both individual users and enterprise IT departments that the assistant will not silently harvest visual content or overhear private conversations.

Why it matters for productivity and the market

If the timeline is reliable, the AI can cut the friction of moving between apps. Instead of opening a file, copying its path, and pasting it into a prompt, a user can simply ask, “What did I work on after lunch?” and receive a ready-made answer. For power users who juggle code editors, design tools, and communication platforms, that reduction in context-switch overhead could translate into measurable time savings.

From a market standpoint, the move signals OpenAI’s push toward “agentic AI” – systems that act on behalf of users based on an ongoing view of the operating environment. Competitors that rely on screen captures face higher regulatory scrutiny and user wariness; event-based logging skirts many of those concerns while still delivering rich context. Enterprises that have blocked screen-capture tools for privacy reasons may find Computer History a viable path to AI-assisted workflows.

The other side of the coin

Even without screenshots, logging every click and keystroke raises red flags. Critics note that metadata can still reveal sensitive intent: a series of searches for medical terms, a pattern of file names, or the timing of interactions with confidential documents can be pieced together to infer private information. The opt-in model mitigates risk, but it also relies on users remembering to enable or disable the feature at appropriate moments—a habit that can be hard to maintain.

データレジデンシー（データの所在）の問題もあります。この機能はイベントログをローカルに保存しますが、ユーザーがチャットを開始すると、関連する履歴の一部が推論のためにOpenAIのサーバーに送信されます。同社は、送信は暗号化されており、リクエストに必要な最小限のコンテキストのみが送信されると述べていますが、企業がこのツールを大規模に導入する前には、オンプレミス処理やより厳格なデータ取り扱い契約を求める可能性があります。

今後の注目点

OpenAIはmacOS以外の展開スケジュールを明らかにしていませんが、アーキテクチャはプラットフォームに依存しないため、WindowsやLinux版も続くことが示唆されています。次の論理的なステップは、ポリシー制御、監査ログ、および推論を完全に企業のファイアウォール内に留める機能を備えた、エンタープライズグレードの提供です。

EUや米国の規制当局は、ユーザーのアクティビティデータを取り込むAIシステムに注目し始めています。「透明性のあるAI」に関する将来的な義務付けにより、OpenAIは具体的にどのイベントがログに記録され、それらがどのように整理されるかを公開する必要が生じるかもしれません。プライバシーポリシーの更新や、「視覚的なキャプチャは行わない」という主張を検証する第三者監査に注目してください。

結論

Computer Historyは、視覚的な監視の代わりにイベントレベルのロギングを採用することで、ChatGPTにユーザーの最近のデスクトップ環境を垣間見る窓を提供しつつ、より強力なプライバシー保護を約束しています。もしこのバランス（有用なほど豊かなコンテキストでありながら、プライバシーの懸念を招かないほど限定的であること）が維持されれば、この機能は次世代のパーソナルAIアシスタントの礎となり、私たちが自身のマシンと対話する方法を再構築する可能性があります。