The productivity claim came first. The methodology was in the footnotes.

As the paper's May 12 account turned Meta's AI efficiency layoffs into a workplace surveillance story, the frame was productivity numbers masking something else — what Wednesday's reporting makes clear is that the something else is a formal program, confirmed by the CTO, with no opt-out for American employees.

Meta has spent the past year arguing that AI assistance makes its engineers significantly more productive — shipping code faster, resolving bugs sooner, compressing the cycle time between idea and deployment. Mark Zuckerberg has cited the productivity numbers in investor calls. The company has pointed to them to explain headcount reductions. Fewer people, the argument goes, can now do more work because AI is doing some of it.

What emerged in late April is how Meta knows this. The Model Capability Initiative, first reported by Reuters, records employee mouse movements, clicks, and keystrokes across work-related applications. It periodically captures screenshots of employees' screens. CTO Andrew Bosworth confirmed the program and confirmed there is no opt-out for U.S. employees [1].

European employees are exempt. GDPR prohibits it.

That jurisdictional carve-out is the detail that warps the productivity story. Meta has constructed a behavioral data collection program whose coverage map is sorted by the strength of legal privacy protection rather than by any organizational logic internal to the company. European engineers, doing the same work, generate no surveillance data. American engineers generate continuous surveillance data. The distinction is a legal one, not a performance one.

The stated purpose of MCI is to train AI systems. Meta says the keystroke logs and mouse movement data help its models understand how humans navigate software — the specific choices people make in menus, the shortcuts they use, the sequence of actions that constitutes completing a task. This is genuinely useful training data. Human-computer interaction at the microscopic level is a hard problem for AI to learn from synthetic data, and real-world behavioral traces from knowledge workers doing real work would represent a meaningful dataset.

What the stated purpose does not explain is why Meta needed to build that dataset from the same employees it is simultaneously offering AI productivity tools to, without telling them in any meaningful way [2].

The traditional compact in employer-employee monitoring has been functional and visible: badge readers know when you arrive, time tracking systems know when you log off, VPN logs know what systems you access. Keystroke logging at the character level, with screenshot capture, is categorically different. It knows what you typed before you deleted it. It knows the order in which you opened windows. It knows how long you paused before clicking. It knows, in short, what you were thinking about — insofar as thought leaves a trace in input device behavior.

Meta says this data will not be used for performance evaluation. The company's word is the only enforcement mechanism for that commitment [1].

American labor law does not prohibit private employer keystroke monitoring on company-owned devices. Several states have notification requirements. Federal law has none. The monitoring is legal. The American worker who objects has no legal remedy and no opt-out.

The productivity claims and the surveillance methodology are not separate stories. The metrics that Meta cites — engineers shipping more code, cycles accelerating, fewer people doing more — are themselves likely products of this data collection. The company is measuring productivity gains from AI assistance using a measurement apparatus that involves monitoring the humans doing the work. The methodology is inside the claim [3].

A workplace where your employer knows every keystroke and has said it will not use that knowledge to evaluate you is still a workplace where your employer knows every keystroke. What that does to how people work — the pauses, the exploratory wrong turns, the thinking-on-the-keyboard that characterizes cognitive labor — is not a question Meta appears to have studied. It is building models of how humans work by watching humans work in conditions that may no longer be normal.

That is the productivity footnote.

-- MAYA CALLOWAY, New York