OpenAI just made screen recording an AI feature, and the privacy settings actually seem thought through for once.
The Summary
- ChatGPT's macOS desktop app now includes Computer History, a feature that logs your clicks, keystrokes, and screen activity to train automations and resume interrupted tasks
- Opt-in by default with granular controls: exclude specific apps/sites, delete individual entries, auto-ignores incognito browsing
- This is OpenAI's first real move toward persistent context agents that learn your workflow patterns, not just your conversation history
The Signal
ChatGPT isn't just watching what you type into its chat window anymore. Computer History tracks everything you do on your Mac — which apps you open, what you click, what you type, what you leave half-finished at 3pm on a Tuesday. It builds a timeline. When you ask ChatGPT or Codex to do something, they can reference that timeline to understand context you didn't explicitly provide.
This is the infrastructure for agent memory. Not the chatbot kind where it remembers your last five messages. The kind where it knows you always pull data from three specific spreadsheets before writing client reports, or that you abandoned a half-written email to your boss yesterday and probably need to finish it.
"This is OpenAI building the substrate for agents that actually understand your work patterns, not just your prompts."
The privacy approach here is surprisingly competent:
- Opt-in, not opt-out — you have to turn it on deliberately
- App and website exclusions — mark your banking app or therapy journal off-limits
- Manual deletion of individual entries — granular control, not just a nuclear "delete everything" button
- Automatic incognito/private browsing exclusion — if you opened it in private mode, Computer History ignores it
OpenAI's Ari Weinstein confirmed on X that private browser tabs are automatically excluded. That's the right default. It signals OpenAI learned something from the Recall debacle — Microsoft's version of this idea launched with security holes you could drive a truck through and no meaningful privacy controls.
The real play here isn't the feature itself. It's what comes after. Computer History is training data infrastructure. Right now, it helps ChatGPT answer your questions with better context. Six months from now, it's how your agent knows to file your expenses every Friday afternoon without being asked. A year from now, it's how a coding agent picks up exactly where you left off when you closed your laptop at midnight, frustrated and stuck.
This is OpenAI betting that the agents worth having are the ones that can see what you actually do, not just what you say. The Web4 version of automation isn't "do this task when I tell you." It's "you've watched me do this twelve times — just handle it."
The Implication
If you're using ChatGPT for work, this feature is worth turning on with tight boundaries. Exclude anything sensitive. Test it on low-stakes workflows first. The value proposition is clear: agents that can see your patterns will automate better than agents flying blind.
For everyone else, watch how OpenAI iterates the privacy controls. If they stay this granular and transparent, it becomes the template for how agent memory should work. If they start loosening defaults or burying exclusions three menus deep, we're back in surveillance territory with better branding.