Altman bets ChatGPT will watch screens in six months

OpenAI CEO Sam Altman told a room of Silicon Valley interns that a descendant of ChatGPT could watch a user’s computer screen all the time, sit in on every meeting, record every call and hold perfect context of their digital life within about six months. The comments, made to Z Fellows founder Cory Levy at Internapalooza, arrived days after OpenAI shipped its first Mac activity-tracking feature and make the always-on assistant feel less like science fiction than the next product step.

Altman framed the system as a collaborator, not a decision-maker. Users would choose which services it can reach. The assistant would then surface forgotten details, flag possible errors or offer to finish small tasks while the person keeps working.

What Altman actually described

Speaking with Levy, Altman laid out a continuous-context future rather than another chat window. He said the technology is close.

I think we are close to a world where you can have a descendant of ChatGPT watch your computer screen all the time, watch every meeting you’re in, record every call, everything that has perfect context of your whole life, everything you see.

Altman said the line in the interview with Cory Levy. He added that people would connect the system to texts, email, documents or Slack and decide the boundaries. The assistant would not seize control of important choices. It would instead act like a sharp colleague who has already read the background.

“If you’re like the CEO of a startup, there’s always like more stuff to do than you can do,” he said. “And context, you can’t all keep track of.” In that setting the AI could pull customer feedback the founder never had time to read, spot a gap in a strategy draft or draft the next email from material already in view.

He estimated genuine usefulness could arrive “sometime in the next six months” and that the leap may be only one model generation away. OpenAI has not named a product, set a ship date or published a specification that matches the full description. The remarks remain a directional bet, not a roadmap commitment.

Computer History already watches Mac activity

The timing is precise. On August 13 OpenAI released Computer History inside the ChatGPT desktop app for macOS. The feature is available to Pro, Business and Enterprise users (with EEA, UK and Switzerland access still pending). It is off by default.

Unlike earlier experiments that grabbed screenshots, Computer History records interaction events exposed by macOS accessibility: clicks, typing, keyboard shortcuts and app switches. It turns those events into text summaries and local memory files that ChatGPT and Codex can reference. Private browsing is excluded. No microphone or system audio is captured. Temporary event files live on the Mac for up to 48 hours; generated memories stay local until the user deletes them.

Users and workspace admins get granular levers:

• Turn the feature on only after explicit opt-in (admins must first enable it for Business and Enterprise seats).
• Include or exclude specific apps and websites.
• Pause and resume collection from the menu bar.
• Inspect the daily timeline, open memory files in Finder, or clear the last 10 minutes, hour, day or everything.
• Require Memories to be enabled so context can travel across chats.

OpenAI’s opt-in Computer History controls and timeline documentation stresses that the company processes events on its servers only long enough to build the summaries and does not retain the raw stream for training. The same page warns that the feature can raise prompt-injection risk from malicious content in apps or sites and advises pausing it during conversations with people who have not consented.

That shipping date sits roughly a week before Altman’s interview circulated widely. Computer History is narrower than the full vision he sketched: no continuous screen pixels, no call recording, no automatic meeting attendance. Yet it already gives ChatGPT a live feed of what the user is doing outside the chat window. The six-month claim therefore reads as an extension of work already in users’ hands rather than a clean-sheet invention.

How the full vision differs from today’s tools

Computer History still waits for the user to ask or to open a relevant chat. Altman’s descendant would maintain ongoing awareness and proactively offer help while a document is open or a customer call is under way. It would pull from email threads, Slack channels and prior meetings without fresh prompting.

OpenAI already offers conversation memory that personalizes replies from earlier chats. Computer History adds computer-level activity. The next jump would fuse both with real-time multi-modal input (screen content plus audio) and deeper service connectors. Altman said the system could become “incredibly useful” once the underlying model generation improves enough to reason reliably over that firehose of context.

The company has also been expanding workplace tooling. Its recent ChatGPT Work workplace agent launch points in the same direction: agents that operate inside company workflows rather than isolated prompt boxes.

Microsoft already tried continuous screen memory

Microsoft’s Recall feature for Copilot+ PCs offers the closest public parallel. Recall periodically captures encrypted snapshots of the user’s screen, runs local OCR and builds a searchable timeline of everything viewed on that device. Microsoft made it opt-in after early previews drew heavy criticism over the ease of extracting snapshot databases.

The company responded with deeper protections. Snapshots and the vector index stay encrypted, keys live in the TPM and are usable only inside a Virtualization-based Security enclave, and Windows Hello Enhanced Sign-in Security is required to search or view history. Sensitive content filtering is on by default. Users can delete snapshots, set retention limits, pause capture and remove the feature entirely. Microsoft’s September 2024 blog on Microsoft’s updated Recall security architecture spells out the four design principles: user control, encryption with protected keys, isolated processing and intentional presence via Hello.

Recall still faces skepticism years later. Security researchers continue to probe residual risks. The episode shows that continuous digital memory is technically feasible today and that the hard problems are policy, encryption boundaries and user trust, not raw capability.

The table makes the progression clear. OpenAI is already past pure chat memory. Microsoft proved the snapshot approach at OS scale. Altman is simply stating the obvious next combination: multi-modal continuous intake plus deep service connectors once models can handle the load without constant hallucination or leakage.

Privacy and control questions that will decide adoption

Giving any model a continuous feed of a knowledge worker’s screen, inbox and calls creates a high-value target. A single compromised account or malicious prompt could expose months of client data, strategy drafts or personal communications. Enterprise admins will treat this as a governance problem first and a productivity feature second.

OpenAI’s current design choices try to keep the blast radius small: local-first storage for raw events, short retention, explicit app exclusions, easy pause and delete, and no training use of the temporary stream. The official Computer History documentation repeats those limits. Still, once the system can also join Zoom or Teams and read every Slack channel, the permission surface expands dramatically.

On X the reaction split cleanly. Productivity accounts celebrated the end of context-switching and the arrival of a true second brain. Privacy-focused users and local-AI builders called it surveillance packaged as convenience. One YC founder promoting an open-source local alternative wrote that users would be “rent[ing] your own memory to an evil corporation.” The line “perfect context of your whole life” spread quickly as both promise and warning.

Altman has previously argued that governments, not labs alone, should set the hard boundaries for advanced AI. His Altman’s earlier call for governments to set AI rules sits in the same tension: the company wants to move fast on capability while acknowledging that continuous personal context requires rules larger than any single firm’s terms of service.

Who feels the change first

Startup founders and overloaded managers are the obvious early users. They already drown in customer tickets, investor updates, Slack threads and half-finished docs. An assistant that has quietly absorbed all of that can surface the one forgotten commitment or the contradictory number without another status meeting.

Knowledge workers who live inside browser tabs and video calls form the next wave. The same system that helps a founder could also generate standup summaries, reconstruct a lost thread of decisions or turn a repeated three-app workflow into a one-click skill.

The losers, at least initially, are people and organizations that refuse the trade. They keep the privacy of air-gapped context and pay for it with slower coordination and more manual recall. Companies that mandate the feature without airtight controls risk employee pushback or regulatory scrutiny in jurisdictions still excluded from Computer History today.

Model quality remains the gating item. Altman said the current generation is not quite there. One more step in reasoning, long-context reliability and multi-modal grounding could flip the experience from “occasionally clever” to “I would not work without it.”

The six-month clock is now running

OpenAI has not promised a ship date. Altman’s language was careful: “close,” “sometime in the next six months,” “one model generation away.” Yet the company just put the first production building block on Mac desktops and immediately let its CEO describe the completed house. Competitors will treat the comments as a public commitment to accelerate their own ambient agents.

By early 2027 the market will show whether continuous-context assistants became genuinely useful on schedule or whether privacy, security and model limits forced another delay. Users who enable Computer History today are already training themselves, and their employers, on the permission model that larger systems will inherit. The screen is no longer just a display. It is becoming the primary sensor for the next generation of AI.