CNBC, in an April 23 follow-up to Reuters's MCI scoop two days earlier, named the third-party sites Meta's Model Capability Initiative will log: Google, LinkedIn, Wikipedia, GitHub, Slack, Atlassian — plus Meta-owned Threads and Manus. [1] The list was widely circulated inside Meta after a Superintelligence Labs researcher posted a memo intended to defuse worker concerns. The paper's Wednesday brief treated the methodology as the disclosure; today's standard treats the named tracked-site list as the corporate-surveillance escalation the methodology was always going to require.

The tool itself, as Reuters first described it, captures mouse movements, click locations, keystrokes, and periodic screenshots on US-based employees' work computers. [2] Meta's Andrew Bosworth, in a separately leaked memo Monday, told employees the "AI for Work" program had been rebranded as the "Agent Transformation Accelerator," and that the goal was for agents "primarily do the work" while employees "direct, review and help them improve." [3] Spokesperson Andy Stone acknowledged MCI data would feed the training pipeline, that it would not be used for performance reviews, and that "safeguards" would exclude unspecified "sensitive content."

The named-sites list is what makes the policy concrete. Tracking how an engineer navigates a GitHub pull request, a Slack thread referencing that pull request, and a Wikipedia tab held open during code review is the workflow data agent models cannot get from public training corpora. The Superintelligence Labs memo, reviewed by CNBC, said the team needed "a big and unbiased" dataset and that capturing on-screen context was the way to do that. [1] Employees called the framing "dystopian" in internal threads — the word recurs in BBC, Reuters and CNBC reporting. [4] [2] [1]

The most important word in the rollout is "US-based." Reuters reported the tool will run on US employees' computers; Meta's spokesperson confirmed the scope to multiple outlets. [2] That scope is not a technical limitation. It is a regulatory tell. Italy's labor code explicitly prohibits electronic monitoring used to assess productivity; Germany's courts have held keystroke logging permissible only in narrow criminal-suspicion cases; and Europe's General Data Protection Regulation, per legal analysts Yale's Ifeoma Ajunwa and York University's Valerio De Stefano, would almost certainly read sustained keystroke and screenshot capture as a data-processing activity requiring explicit consent and a documented purpose. [2] [3] Meta is not running MCI in Europe because Meta does not think MCI clears GDPR.

Bosworth's internal response to employees clarified the structure. When workers asked about opting out, Bosworth replied on the internal forum that "there is no option to opt out of this on your work provided laptop" — a sentence Business Insider reproduced from internal threads. [5] Workers who do not want their activity captured "can control what shows up on your screen by not doing personal work on your work computer," the launch memo said. The labor implication is direct: the corporate device is now a data-collection sensor whose continuous operation is a condition of employment.

The macro frame on X is the layoff sequencing. Meta has telegraphed that roughly 20 percent of the workforce — building on earlier cuts of 2,000 and a substantial pullback in open job listings — will be reduced beginning May. [6] MCI was announced to Superintelligence Labs in the same week those layoff plans were being reported inside the company. The causal vector that the labor accounts on X describe — employees training the agents that will replace them — is harder to falsify when the timelines run inside the same calendar week.

The third-party sites are the part the regulators will read most closely. Google's Workspace administrators, LinkedIn's product owners, Atlassian's enterprise customers, and Wikipedia's contributors did not consent to be the substrate of Meta's AI training pipeline. None of them are direct counterparties; each is a place where Meta employees conduct day-to-day work that intersects with content they did not author. The legal question — whether Meta's screen capture of a third-party SaaS dashboard creates a derivative-data obligation upstream — has not been litigated. The first lawsuit will not arrive from a Meta employee. It will arrive from a vendor whose product appears in screenshots.

The cleanest reading of MCI is structural. The methodology — keystroke, click, screenshot — is the cost of running an agent program at Meta's scale. The named-sites list is the company's admission that the training surface is the open web filtered through employee workflows. The US-only scope is the company's admission that the training surface is not yet legal in Europe. The "Agent Transformation Accelerator" rebrand makes the goal explicit. The agents will do the work. The data to make them do it comes from the people whose work is now being captured one keystroke at a time.

-- THEO KAPLAN, San Francisco