OpenAI's latest ChatGPT feature records your keystrokes and saves them as plain text.

      OpenAI has introduced a feature called Computer History for ChatGPT, which tracks clicks, keystrokes, keyboard shortcuts, and app switches on macOS to create a searchable memory. The memory files are stored locally in unencrypted plain text, and this feature is not accessible in the EEA, Switzerland, or the UK.

      ChatGPT now keeps a record of actions performed on your Mac. Computer History logs various user interactions through the operating system’s accessibility framework and compiles them into a memory that the assistant can reference.

      However, it is more limited than it may seem. The feature does not capture screenshots, record video, take audio input, or include private browsing sessions.

      This limitation is intentional. Computer History serves as a replacement for Chronicle, the previous Codex feature that used screen captures for context, and it opts for input events over images.

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

      The consent process is more structured compared to typical launches. In Business and Enterprise workspaces, an administrator must enable the feature, the individual user must provide separate consent, and the Memories option must be activated.

      Users are also equipped with real controls. Applications and websites can be added to lists to include or exclude data, and users can pause recording directly from the menu bar.

      The retention period for raw data is short. Temporary event files remain on the device for 48 hours before being deleted, and OpenAI states they process these files on their servers without retaining them afterwards unless legally mandated.

      The area of training introduces some ambiguity. OpenAI claims that the memory files are not used to train models, but when memories appear as context in future chats, that chat content could potentially become training data based on user settings.

      The company’s documentation candidly highlights potential risks. It points out increased vulnerability to prompt injection, recommends against using the feature with communication applications unless all parties have consented (highlighting an ongoing challenge in the industry regarding recording others), and advises excluding any sensitive data related to health, finance, or personal matters.

      Individuals in Europe cannot access this feature at all, as it is not available in the European Economic Area, Switzerland, or the United Kingdom.