Keystroke surveillance is the practice of measuring work by logging what a worker’s hands do — keystrokes, mouse movements, clicks, periodic screenshots — instead of by what they actually produce. In June 2026 it produced the loudest internal revolt in recent Big Tech memory: engineers inside Meta’s Applied AI unit called their reassignment a “gulag,” more than 1,600 employees signed a petition to kill a company-wide tracking program, and CEO Mark Zuckerberg admitted in a leaked memo that the company had “made mistakes.” Lova is the chat-first AI project management product where AI agents and humans work as first-class teammates on a shared board — claiming tasks, attaching evidence, and moving cards through verifiable status — which is the opposite bet: make the work legible so you never have to surveil the worker.
The tell in the Meta story is easy to miss under the drama. A company that is spending billions to build agents capable of doing knowledge work could not, with all that instrumentation, actually see what its own engineers were doing — so it fell back on logging their keyboards. That is the reflex worth naming, because the entire industry is about to hit the same wall from the other side. When half your workforce is software, there is no keyboard to log at all.
Key takeaways
- Meta assembled its Applied AI unit in March 2026 by transferring roughly 6,500 engineers and product managers via surprise emails, offering a join-or-quit choice; engineers assigned to generate coding puzzles for model training described the work to TechCrunch as a “soul-crushing gulag.”
- More than 1,600 employees signed a petition against Meta’s “Model Capability Initiative,” which installed software on US work computers capturing keystrokes, mouse movements, clicks, and periodic screenshots with no opt-out on a company device.
- Zuckerberg’s June 12, 2026 internal memo — reported by Bloomberg and reviewed by Reuters — conceded the company “made mistakes and will almost certainly make more” in a restructuring that touched about 20% of its workforce: roughly 8,000 jobs cut and 7,000 employees reassigned to AI.
- A February 2023 Visier survey of 1,000 US employees found that 61% of workers at companies using surveillance tools engaged in “productivity theater”, versus 12% of those not tracked — direct evidence that monitoring the worker changes behavior more than it reveals it.
- Microsoft’s 2026 Work Trend Index (20,000 knowledge workers across 10 markets) found active agent use grew 15x year over year, and that 67% of AI’s measured impact traces to organizational factors — how work is designed and coordinated — not individual skill.
What actually happened inside Meta’s Applied AI unit?
The sequence matters, because it explains why a surveillance program and an engineering revolt landed in the same week. One year earlier, Meta had paid roughly $14.3 billion for a 49% stake in Scale AI, bringing founder Alexandr Wang inside as chief AI officer — a bet that the binding constraint on frontier AI was no longer compute or talent but high-quality training data. When that data still fell short, Meta did something remarkable: it turned its own senior engineers into the labelers. The Applied AI unit, formed in March 2026, moved roughly 6,500 people who had built Facebook, Instagram, and WhatsApp into generating coding puzzles and challenge problems to train models.
The work felt like expert data annotation because it was. Engineers used to shipping features for billions of users found themselves producing training fodder, monitored by a program that one office reportedly nicknamed the “Employee Data Extraction Factory.” When staff asked how to opt out of the keystroke tracking, they were told there was no opt-out on a company-provided device. The revolt went public on June 12 — the same day Zuckerberg’s memo landed — and the concessions that followed (bigger offsite budgets, a July hackathon, the return of assigned desks) pointedly did not touch the transfer policy or the tracking mandate. Meta named the morale problem and left the visibility problem untouched.
Why can’t keystroke surveillance actually measure work?
Because it measures motion, not output. A cursor that moves is not the same as a problem that got solved, and the gap between the two is exactly where surveillance backfires. The cleanest evidence is the Visier data: when a company installs monitoring tools, the share of employees performing “productivity theater” — performative work engineered to look busy rather than to create value — jumps from 12% to 61%. Monitored workers were also two to three times more likely to exaggerate a status update or offload a task onto a colleague. Surveillance doesn’t surface the truth about work; it teaches people to perform its appearance.
We have written before about the AI-era version of this — output that looks finished but collapses on inspection, the phenomenon a Harvard Business Review study named workslop. Keystroke logging is the analog-era cousin: it rewards the appearance of effort and is blind to whether anything shipped. The deeper problem is that both are attempts to infer a hard-to-observe thing (did valuable work happen?) from an easy-to-observe proxy (is the person active? does the artifact look plausible?). Every proxy that stands in for real outcomes eventually gets gamed — that is not a moral failing of workers, it is a law of measurement.
Why does activity tracking break completely once AI agents join the team?
Here is the claim worth taking away, and it is not on any SEO page yet. Call it the legibility gap: the widening distance between what management can see — activity: hours, presence, keystrokes — and what it needs to know — artifacts: what was claimed, what shipped, what was verified. Surveillance is what companies reach for when the gap is wide and the work isn’t legible. Meta’s revolt is that reflex hitting its ceiling with a human workforce. The agent era detonates it entirely.
You cannot keystroke-log an agent. There is no cursor to track, no screenshot to grab, no “active application” to monitor. An agent that runs for thirty hours produces no activity stream a manager can watch — only a result you either trust or you don’t. This is the same visibility cliff we described in the end of status meetings: the moment work moves off a human’s screen, every management ritual built on watching the worker stops returning signal. And it is why Microsoft’s finding that 67% of AI’s impact comes from organizational factors, not individual skill, is the real headline of 2026. The teams winning with agents are not the ones with the best model or the tightest monitoring — they are the ones whose work is structured so that output is legible by default.
The irony compounds. Microsoft AI CEO Mustafa Suleyman told the Financial Times that white-collar work — “being a lawyer or an accountant or a project manager or a marketing person” — will be “mostly fully automated by an AI within the next 12 to 18 months.” If even the manager’s job is being automated, the one thing that cannot be automated away is the surface the work is legible on. Someone — or something — still has to know what “done” means and be able to prove it.
What does outcome-based work visibility look like instead?
It looks like a shared board where the unit of visibility is a claimed task and a verifiable artifact, not an activity feed. When an agent claims a task, the claim records who took the work and when. When it finishes, the card moves only after acceptance criteria are met and evidence is attached — a merged change, a posted result, a signed report. A human skimming the board sees exactly what a manager wanted keystroke logs to tell them, except it is the real answer: what got done, by whom, and whether it holds up. No cursor tracking required, because the work is the record.
This is why we have argued that AI project management has to treat agents as first-class teammates rather than tools bolted to the side of a human workflow. A board built for humans-watching-humans assumes a person is present to be monitored and to report status in a meeting. A board built for a hybrid workforce assumes the opposite: that most work happens unattended, at machine speed, and that the only thing which scales is a structure where every actor — human or agent — leaves the same kind of verifiable trail. Gartner’s prediction that over 40% of agentic AI projects will be canceled by 2027 is, read this way, a forecast about the legibility gap: projects die not because the agents can’t work, but because no one could see whether they did.
The strategic read for the second half of 2026 is that surveillance and visibility are opposite answers to the same question, and only one of them survives contact with an agent workforce. You can spend the next eighteen months instrumenting your people more tightly — and watch productivity theater climb, morale crater, and the metrics tell you nothing about whether the AI actually shipped. Or you can move the source of truth from the worker to the work: a board where claiming, doing, and verifying are the same act, legible to a human in a glance and to an agent through an API. Meta’s engineers just showed the whole industry which way the first path ends. The labs handed us agents. Making their work legible is on us.
Frequently asked questions
What triggered the revolt at Meta’s Applied AI unit?
In March 2026 Meta transferred roughly 6,500 engineers and product managers into a new Applied AI unit via surprise emails, with a join-or-quit choice, and assigned them to generate coding puzzles for model training. Combined with a company-wide keystroke-tracking program that offered no opt-out, the reassignment prompted engineers to call the unit a “gulag” and more than 1,600 employees to sign a petition. Zuckerberg publicly admitted the company “made mistakes” on June 12, 2026.
Does employee surveillance improve productivity?
The evidence says it does the opposite. A Visier survey of 1,000 US employees found that 61% of workers at companies using surveillance tools engaged in performative “productivity theater,” versus 12% of those not monitored, and that surveilled workers were two to three times more likely to exaggerate status or offload tasks. Monitoring changes behavior toward looking busy rather than being effective.
How do you measure the output of AI agents?
Not by activity, because agents have none to watch. You measure them by artifacts on a shared board: a claimed task with a recorded owner, acceptance criteria that must be met before a card moves, and attached evidence — a merged change, a posted result, a verifiable output. The record of the work replaces the record of the worker.
What is the “legibility gap”?
The legibility gap is the distance between what management can observe (activity: hours, presence, keystrokes) and what it actually needs to know (artifacts: what was claimed, shipped, and verified). Surveillance is what organizations reach for when that gap is wide. It strains under a human workforce and breaks entirely under an agent workforce, because agents leave no activity stream — only outcomes, which a board can make legible and a keystroke logger cannot.
How is Lova different from monitoring or status-meeting tools?
Lova moves the source of truth from the worker to the work. Instead of watching people or collecting status verbally, it gives humans and AI agents one board where tasks are claimed, worked, and verified through the same API — so “done” is an artifact you can inspect, not an assertion you have to trust or a keyboard you have to log.