Meta CTO explains why keystroke tracking program was paused

Meta's controversial AI training experiment hit two roadblocks

Meta's Chief Technology Officer Andrew Bosworth has provided new details about why the company suspended its Model Capability Initiative, a program that tracked employee keystrokes and mouse movements to train artificial intelligence models.

In an interview with The Atlantic CEO Nicholas Thompson released Wednesday, Bosworth disclosed that the program encountered both a data handling mishap and a fundamental design flaw. The interview was recorded in late June.

A researcher moved data to the wrong location

Bosworth explained that while the keystroke data was "quite secure" with limited access, a Meta researcher working with the information inadvertently placed it somewhere it shouldn't have been within the company's internal systems.

"One of the researchers who was working downstream with that data—and there was no breach here—but had put it in a place it wasn't supposed to go," Bosworth told Thompson. The employee data, which had been transformed from its original state, ended up accessible to the broader company rather than remaining restricted.

Bosworth emphasized that Meta did not suspect malicious intent and characterized the incident as an error rather than foul play. The company decided to lock down the entire program while investigating.

The program also generated redundant data

Beyond the data handling issue, Bosworth revealed a second problem: the system was producing too much similar information rather than the diverse data needed for effective AI training.

"Variance is far more important than a high volume of the same thing that gets collapsed into one example, basically," he explained. This lack of variety undermined the program's core purpose of improving Meta's AI models.

In response to both issues, Meta expanded opt-out options several weeks after the initial launch. Employees could now pause their participation indefinitely—a significant shift from the original mandatory enrollment that sparked internal backlash.

Why it matters

The Model Capability Initiative, launched in April, represented one of the tech industry's most invasive approaches to gathering AI training data from employees. The program's troubles highlight the tension companies face between advancing AI capabilities and respecting workforce privacy. Bosworth himself acknowledged during an internal meeting that employee morale at Meta was "probably one of the worst it's ever been" in the company's two-decade history—a stark admission from a senior executive at one of the world's largest technology firms.

The incident also underscores the operational challenges of handling sensitive data at scale, even within organizations with sophisticated security infrastructure. When a single researcher's error can expose employee information company-wide, it raises questions about data governance practices in AI development.

Meta paused the program in June after the data became accessible across the company. A Meta spokesperson told Business Insider at the time that the company had "carefully designed this program with privacy safeguards" and found no indication that data was improperly accessed, though the investigation continued.

Meta representatives declined to provide additional comment to Business Insider beyond Bosworth's interview remarks.