An AI system that watches a screen and records meetings does not collect only its user's life. It collects everyone who appears in the workflow.
Sam Altman's reported six-month forecast for a deeply contextual ChatGPT makes the privacy issue immediate. The value comes from continuous context. The risk comes from the same place.
The user cannot consent for everyone
A screen can contain customer records, private messages, health information, credentials, confidential documents, and content licensed for limited use. A meeting can include employees, candidates, partners, and customers operating under different policies or laws.
A click by the account owner does not resolve those rights. People need to know when capture is active, what the system receives, why it is processed, where it runs, how long it stays, and who can retrieve the derived memory.
Ambient access needs session boundaries
Continuous capture is the easiest technical model and the weakest governance model. Access should be limited by purpose, application, meeting, participant, and time.
A trustworthy design makes recording visible, pauses on sensitive surfaces, separates personal and work contexts, and lets users inspect and delete both raw capture and derived memory. Local processing can reduce exposure, but it does not remove consent or ownership questions.
Derived memory may be more sensitive than the recording
The system may infer relationships, performance, health, financial pressure, or internal disagreement from ordinary activity. Deleting a recording is incomplete if the summary, embedding, profile, or agent memory remains.
Organizations need an owner for the complete lifecycle: collection, processing, access, correction, retention, deletion, incident response, and vendor change. They also need a usable path for people who cannot or will not participate.
Always-on context can make AI substantially more useful. It should arrive as a bounded capability, not a default surveillance layer hidden behind convenience.