Meta eases Zuck’s employee keylogging plan after staff backlash

Meta is not abandoning its workplace monitoring push. It is, however, giving employees a 30-minute break from it. According to an internal memo first reported by Reuters and described by The Register, Meta still plans to capture employee keystrokes as previously understood, but staff will be able to switch off the monitoring for half-hour periods and request a total exemption. The memo was sent Tuesday by Stephane Kasriel, vice president at Meta’s Superintelligence Labs AI division. In other words: the data grab stays, but the company is handing out short privacy windows after employees complained loudly enough to make the original plan look expensive, invasive, and clunky.

The reason Meta is doing this is not subtle. The software sits inside what Meta calls the Model Capability Initiative, and its job is to collect workers’ keystrokes, mouse movements, and screenshots at different points so Meta can build AI agents that better understand how humans use computers. That makes Meta’s own employees the training set. The irony is obvious: one of the internet’s most aggressive data collectors is now collecting data from the people who work for it. Staff criticism has centered not just on privacy, but on the practical drag. Some employees said the initiative drained their devices’ batteries, while remote workers said it put undue strain on home internet usage. Meta did not respond to The Register’s request for comment, and it also did not respond six weeks earlier when asked about the scheme.

That employee backlash matters because it exposes the hidden cost of building AI systems that learn from real-world behavior. Meta is not just trying to make models bigger. It is trying to make them more useful by teaching them how people actually move through software, click by click and keystroke by keystroke. That is the logic behind collecting granular interaction data from employees rather than relying entirely on outside contractors. In a leaked audio recording from an internal Meta meeting on April 30, CEO Mark Zuckerberg argued that the company wanted to feed “a very large amount of content into the AI model” so it could learn how “smart people use computers to accomplish tasks.” He added, “I think that this is going to be a very big advantage if we can do it.”

Zuckerberg’s framing gives the plan a blunt strategic edge. Meta is trying to accelerate advanced AI model development faster than competitors, and internal usage data is a shortcut to understanding workflow behavior at scale. In the same six-minute monologue, Zuckerberg repeatedly called Meta staff “smart people.” That may have been meant to soften the blow, signal confidence, or both. He also said Meta chose to capture data from its own people rather than outside contractors because they were smarter than the workers the company could bring in on a temporary basis. The point, from Meta’s perspective, is efficiency and quality: if you want AI agents that understand how capable people use computers, you train them on capable people. The point, from employees’ perspective, is a lot less charming. Being asked to help build the company’s AI future is one thing. Having your activity slurped up in the process is another.

Meta’s partial retreat also fits a broader pattern in workplace AI deployment. Companies increasingly want more behavioral data, but the operational and cultural costs are becoming harder to ignore. Monitoring tools can create friction on devices, raise questions about what is being captured and why, and intensify distrust when staff already feel squeezed by layoffs or internal restructuring. The source notes that Zuckerberg confirmed Meta had no intention of using the data captured by the monitoring software to surveil employees’ activity or productivity, though he did not commit to saying the data would be anonymized. That distinction matters. Collecting data to train AI is not the same as policing performance, but the technical capability to observe behavior is still there, and employees know it. Once a company normalizes recording keystrokes, mouse movements, and screenshots for one purpose, the boundary between model training and workplace oversight gets harder to explain in plain English.

For founders, operators, and board members, this is the part to watch. Meta’s move shows that AI training ambitions can collide with employee expectations faster than policy teams can draft memos. The backlash forced a concession, even if a limited one, which tells every executive team two things. First, internal data advantage is valuable enough that companies will keep chasing it. Second, rollout design matters almost as much as the technology itself. If the data collection burns battery, eats bandwidth, or feels opaque, employees will push back, regulators may take a closer look, and the original business case can get dragged into a very public argument about trust. Meta is still moving ahead, but the 30-minute privacy break is a small signal with a large meaning: even inside Big Tech, surveillance has a user experience problem.