01 Trigger Event
OpenAI has launched the Computer History feature on the macOS version of the ChatGPT desktop application, in opt-in mode. It continuously records user clicks and keystrokes, building an activity timeline that both ChatGPT and Codex can reference in subsequent requests. Product lead Ari Weinstein confirmed on X that the system automatically ignores incognito / private browsing content, users can manually exclude specific apps and websites, and individual entries can be deleted.
Source: The Verge. I haven't personally tested this feature on macOS — the analysis below is based on publicly available information.
02 What This Really Means
The question isn't that OpenAI added a productivity feature — it's that OpenAI is running a dual flywheel of distribution + training data, and this time spinning both directions simultaneously.
Surface narrative: Make ChatGPT understand your workflow better and deliver more accurate suggestions. Underneath:
- Product dimension: Computer History feeds Codex, upgrading the agent from "waiting for your command" to "seeing where you're stuck and proactively stepping in." This is a completely different species from Anthropic's Claude Code / Cursor's agent loop.
- Data dimension: Users' long-term click + keystroke streams are a gold mine for fine-tuning agents. OpenAI pushed the o-series / GPT-5 series toward agents in 2024; now they're going back to build the data infrastructure.
This is in the same lineage as Ben Thompson's aggregation theory. Google became the gateway to the internet through Chrome + Search; now OpenAI wants to become the gateway to your entire digital life through the desktop app. Once Computer History accumulates months of context, the switching cost of moving to Claude / Gemini rises sharply — not because of model capability differences, but because your workflow memory isn't there.
03 Historical Analogy
The most direct comparison is Microsoft Recall from June 2024. Recall was designed as opt-out with screenshot-level surveillance, was besieged by the security community and media, and Microsoft was forced to delay it, remove screenshots, and add BitLocker encryption. A year later, adoption remains dismal.
OpenAI clearly absorbed that lesson:
- opt-in rather than opt-out
- ability to exclude specific apps / websites
- supports entry-level deletion
- automatically skips incognito
This is more sophisticated privacy design. But Recall's lesson has a second half: even with the design done right, the category itself still has a trust threshold. Users' instinctive resistance to "AI watching you continuously" won't disappear just because it's opt-in.
A further analogy is Google Now from 2012. Back then, Google used Gmail / Calendar / Search / Location to assemble an ambient assistant, hyped as "the next stop after search." The result: Google Now was replaced by Assistant, Assistant by Gemini. The reason: data flywheels sound great in theory, but the daily driver product experience is extremely hard to build.
Whether OpenAI will repeat this misstep is the real question to answer over the coming year.
04 What This Means for AI Builders
Decisions to make this week:
- Product positioning: If you're building a productivity / dev tool on macOS, your users now face a control group of "AI already sees everything." Your counter-positioning can be local-only / no activity stream uploads — this has leverage in privacy-sensitive industries like legal / healthcare / finance.
- Exclude list audit: Assess whether your app / website will be actively excluded by users. If so, it means OpenAI's AI can't see your users using your product — that's a data disadvantage in itself.
To observe this quarter:
- Codex's agent capabilities will climb rapidly because OpenAI is building closed-loop real user behavior data — this data line Anthropic and Google can't close in the short term.
- The Agent SDK war will further bifurcate: closed-source labs get real user behavior data; the open-source camp (Qwen / DeepSeek / Llama) can only rely on synthetic + public corpus. This gap will widen in the agent category because agent training depends most heavily on real environment interaction data.
- The protocol layer — MCP / A2A / OpenAI Apps SDK — will gain a new dimension of local timeline: future agent-to-agent negotiation will not only be about tool calls, but also about "can I read your computer history."
One concrete build idea: even if your product has a small user base, vertical depth behavioral data has value. An agent focused on legal / design / gaming, even with only a few thousand power users, accumulates click streams worth more than generic crawler data.
05 Counterarguments
I may have misjudged in the following areas, listing them:
First, users simply won't turn this feature on. Recall's story reminds me that even with sufficiently sophisticated privacy design, users' trust threshold for "AI watching me constantly" remains high. Computer History's opt-in model means most users will keep it off, and the real data volume OpenAI gets may be far lower than my assumption.
Second, I'm overweighting the training data flywheel. Directly using user clicks / keystrokes to train the next generation of models faces significant restrictions under the EU AI Act and various state privacy laws. This feature may only exist in a product-layer personalization loop and won't flow back into pretraining / fine-tuning datasets. I haven't seen OpenAI publicly confirm this, and may have misjudged.
Third, macOS sandbox is a hard constraint. Apple's restrictions on accessibility / keystroke hooks are stricter than Windows. Computer History's actual observation depth on macOS may be far less than Recall's free run on Windows. If so, its strategic value to OpenAI would be substantially diminished.
Fourth, I'm overestimating switching cost. The real moat isn't data, it's workflow integration. Users change computers / change accounts / clear storage, and accumulated data disappears. Anthropic / Google wanting to use similar local timeline features (Google has been doing parts of this with Pixel / Chrome for a while) doesn't face insurmountable barriers. The switching cost OpenAI builds here may not be as thick as imagined.
Final point: I haven't actually run Computer History on macOS. All my UX experience judgments above are extrapolated from reporting. If the experience is poor (lag / battery drain / false triggers), the entire flywheel narrative extinguishes early. This needs to be revised after the first wave of user feedback.