Meta has officially paused a controversial internal program that logged its own employees’ moment-to-moment computer behavior to generate AI training data.

Known as the Model Capability Initiative (MCI), the program was temporarily halted after a high-priority internal security incident (a “SEV” report) revealed that highly sensitive employee data had accidentally been exposed across the company.

The sudden freeze marks the culmination of weeks of intensifying internal friction between Meta’s workforce and leadership over workplace surveillance and data security.

1. What Was the Model Capability Initiative?

Launched in late April 2026 by Meta’s SuperIntelligence Labs, MCI was not built as a traditional remote-work productivity monitor. Instead, it was an aggressive data-harvesting tool designed to build next-generation autonomous AI work agents.

  • The Logging Blueprint: The software tracked and recorded mouse movements, mouse clicks, keyboard keystrokes, and took occasional screen snapshots on the work computers of U.S.-based employees.
  • The Training Goal: Meta used this granular behavioral dataset to teach its AI models how humans perform everyday digital work—specifically mastering complex computer tasks like navigating dropdown menus, utilizing keyboard shortcuts, and interacting with enterprise software UIs.

2. The Internal Leak That Triggered the Freeze

The program hit a wall when employees discovered that the data collected by the tool was being stored insecurely and was accessible to a broad audience of other Meta staffers.

An internal investigation was launched after an employee filed a high-priority security report. The documentation reviewed by media outlets showed that the exposed, internally visible information included:

  • Full AI prompts and text transcriptions.
  • Private employee chat conversations and performance review data.
  • Internal sensitivity data classifications (DSS levels 1–4).
  • Unencrypted personal data, including the private medical and tax information of “many thousands” of workers who had accessed personal accounts on their work machines.

Meta spokesperson Tracy Clayton confirmed the temporary halt, stating that while the company built the initiative with strict privacy safeguards and has found no evidence that data was improperly accessed, it is pausing the program globally while a thorough data security review is conducted. Because of the scale of the rollout, Meta noted it would take time to fully deactivate the recording software across all employee laptops.

3. A History of Internal Backlash

Even before the security leak forced a complete stop, the MCI program had triggered massive internal unrest across the tech giant, with employees heavily criticizing the company as a “data extraction factory.”

April 2026 ──────► MCI rolls out; records clicks, keystrokes, and screenshots.
May 2026   ──────► Backlash grows; 1,600+ staff sign protest petition over battery drain and privacy.
Early June 2026 ─► Meta compromises; introduces a "30-minute pause" button and limited exemptions.
Late June 2026 ──► SEV filed over internal data exposure; program officially halted for investigation.

Beyond core privacy violations, workers complained that the continuous, background recording software heavily degraded laptop battery life and significantly spiked home internet data usage—a major point of contention for Meta’s remote workforce. Resistance became highly public inside the company’s offices, where staff circulated protest flyers and an internal essay by a protesting engineer went viral, garnering over 20,000 views from colleagues.

4. The Broader AI Data Bottleneck

The friction at Meta highlights a massive structural hurdle facing the entire tech sector. As frontier AI labs race to build agentic, human-level digital assistants, they have run into a severe “data wall.”

Public web data is largely exhausted, and scraping outside platforms carries massive legal and copyright liabilities. Turning inward to harvest the data of their own highly skilled engineering and corporate workforces has become a tempting pipeline for tech giants—but as Meta’s current lockdown proves, transforming employees into non-consensual AI training models carries immense operational and security risks.