Meta Platforms has suspended an internal program that collected employee computer activity to train artificial intelligence models after a security issue raised concerns about potential access to sensitive information, according to a report by Wired.
The initiative, known internally as the Model Capability Initiative (MCI), was launched in late April as part of Meta's broader effort to improve AI systems that can interact with computers in ways that more closely resemble human users. The program recorded mouse movements, keystrokes and screen activity from participating employees as they performed everyday tasks.
The suspension marks a setback for Meta's push to develop more capable AI agents at a time when major technology companies are racing to build systems that can navigate software, complete tasks and operate digital tools with minimal human intervention.
"We have carefully designed this program with privacy safeguards and while we have no indication at this time that any data was improperly accessed by Meta employees, we're pausing it while we investigate," Meta spokesperson Tracy Clayton told Wired.
Internal documents reviewed by the publication suggest Meta viewed employee activity as a valuable source of training data. A staff AI research scientist described the effort as a way to improve areas where AI systems continue to struggle.
"This is where all Meta employees can help our models get better simply by doing their daily work," the scientist wrote in a company memo.
The program quickly became controversial inside the company. Employees circulated flyers across Meta's U.S. offices urging colleagues to oppose the initiative and sign a petition questioning its privacy protections.
"Don't want to work at the Employee Data Extraction Factory?" one flyer stated, according to a report published in May.
The petition argued that management had failed to provide adequate transparency regarding privacy reviews. It claimed that "when employees asked what privacy reviews were conducted, including any 'people data reviews' (which are required for processing employee data), no completed privacy reviews were provided."
The document further stated that "The outlined privacy mitigations were vague, and leadership's confidence in them appeared limited - evidenced by the selective opt-out afforded to executives."
The dispute reflects a broader debate unfolding across the technology industry as companies seek larger volumes of real-world data to train increasingly sophisticated AI models. Meta executives have openly discussed ambitions to build AI systems capable of performing significant portions of workplace tasks.
In a separate memo, Meta Chief Technology Officer Andrew Bosworth wrote, "The vision we are building towards is one where our agents primarily do the work and our role is to direct, review and help them improve."
Meta spokesperson Andy Stone defended the need for the data collection effort, saying the company's AI systems require "real example" of how people use computers in daily life. He cited "Things like mouse movements, clicking buttons, and navigating dropdown menus" as examples of behaviors useful for training models.
According to Wired, a former employee who opposed the project described the security lapse as a "mess" and said concerns had been repeatedly raised before the incident occurred. The former employee added, "When workers raised concerns, leadership doubled down and failed to acknowledge the risks workers raised about the safety and privacy of worker and customer data."
Stephane Kasriel, a Meta vice president overseeing AI research efforts, later informed employees that the issue had been identified and addressed within hours. Although an initial fix reportedly failed, Kasriel said the company would resume the initiative only "when we are confident in the effectiveness of our data protection control."
She added that Meta has already "gathered sufficient data to assess the long-term value of the tool," suggesting the company may use lessons from the paused program to shape future AI training efforts.
