OpenAI's latest ChatGPT feature records your keystrokes and saves them in unencrypted text.

      OpenAI has introduced Computer History, a feature for ChatGPT that tracks clicks, keystrokes, keyboard shortcuts, and app switches on macOS, creating a searchable memory. These memory files are stored locally as unencrypted plain text and are not accessible in the EEA, Switzerland, or the UK.

      ChatGPT now has access to your Mac activity through Computer History, which logs various user interactions via the operating system’s accessibility framework, transforming them into a memory that the assistant can reference.

      However, the feature is more limited than it may appear. It does not capture screenshots, screen recordings, microphone input, or system audio, and private browsing sessions are excluded.

      This limitation is a purposeful adjustment. Computer History is intended to replace Chronicle, a Codex feature that recorded the screen for context, and it uses input events instead of images.

      One aspect that remains unchanged is that memory files are plain-text Markdown files stored locally without encryption, meaning any program running under the same macOS account can access them.

      Consent for this feature is managed more carefully than in previous launches. In Business and Enterprise workspaces, an administrator must enable it, individual users must provide separate consent, and the Memories feature needs to be activated.

      Users have more control over the feature. They can include or exclude specific apps and websites, and they have the option to pause recording from the menu bar.

      The retention period for raw data is brief. Temporary event files will remain on the machine for 48 hours before being deleted, and OpenAI states it processes these files on its servers without retaining them afterward, unless legally obligated.

      The situation becomes less clear regarding training. While OpenAI states that memory files are not utilized for training models, memories that are referenced in future chats might still be included in training data, depending on user settings.

      OpenAI’s documentation explicitly highlights the potential risks. It notes a greater risk of prompt injection and advises against using the feature with communication apps unless all parties involved have consented, reiterating an industry-wide issue concerning recording conversations with others and suggests excluding anything related to health, financial, or personal data.

      For users in Europe, the feature is not available, as it cannot be used in the European Economic Area, Switzerland, or the United Kingdom.