Meta has paused an internal programme that recorded employees’ computer activity for AI training after discovering that some of the information collected may have been accessible to other workers across the company.

The company-wide Model Capability Initiative, known internally as MCI, was introduced around two months ago to gather data showing how people interact with computers. The programme recorded activity including mouse clicks and keystrokes, with the resulting information intended to support the development of artificial intelligence models.

Meta halted the initiative on Monday while it investigates how the collected information was stored and protected.

A company spokesperson told the BBC that the MCI was “on pause for now” following the discovery that some data had potentially been left accessible to anyone inside Meta.

“We have no indication at this time that any data was improperly accessed by Meta employees,” the spokesperson added.

The pause follows weeks of internal opposition to the initiative, with employees raising concerns about the extent of the monitoring and questioning where the collected information would be stored, who would be able to access it and how it would be protected.

Nearly 2,000 Meta employees reportedly signed a petition calling for the programme to be cancelled. In an earlier attempt to address the backlash, the company said workers would be allowed to opt out of tracking for periods of up to 30 minutes.

However, the discovery that some of the information may have been broadly accessible inside the company has added to concerns over how the programme was designed and implemented.

MCI was intended to provide Meta with detailed information about how employees use computers, potentially helping the company train AI systems capable of understanding and carrying out digital tasks.

The programme’s suspension comes as Meta reorganises large parts of its business around AI development and prepares to spend up to $145 billion (£109bn) on the technology this year.

The company has also carried out extensive layoffs and moved employees and resources towards AI-focused initiatives. Internal tensions have also surfaced publicly, with Wired reporting that an employee interrupted a company livestream with an expletive-filled outburst directed at Meta’s AI leadership.

AI expansion reshapes Meta

The employee-tracking programme forms part of a broader push by Meta to increase the role of AI throughout its products and internal operations.

The company is also changing how it moderates content across Facebook and Instagram, with plans to reduce its reliance on third-party human moderators and allow AI systems to handle a larger proportion of content review and enforcement.

Meta has historically depended on tens of thousands of contract workers employed by outside companies in countries including the US, India, Spain and the Philippines.

Recommended reading

• Tech Layoffs Jump 40% in Q1’26 as AI Fever Spikes
• Oracle Announces “Significant” Layoffs As AI Pressures Squeeze
• Amazon Plans Largest Corporate Layoff Round in Decades

These workers review content flagged for potentially breaching Meta’s policies on areas including hate speech, violence, harassment, child exploitation, misinformation and fraud.

Under the new strategy, AI systems will increasingly carry out routine and repetitive moderation tasks. Meta’s internal testing reportedly found that the technology could detect some categories of explicit content and fraudulent activity more quickly and at greater scale than contracted moderation teams.

The company expects automated systems to identify and act against accounts displaying suspicious behaviour, such as unusual login activity or repeated policy violations. AI tools will also be used to reduce the workload facing Meta’s internal review teams.

Human specialists will remain involved in reviewing complex cases that require a detailed understanding of context, as well as overseeing, training and evaluating the models.

However, a huge issue with this approach is that automated moderation tools may struggle to distinguish between prohibited content and material such as satire, political commentary or borderline speech. There are also legitimte concerns about transparency and accountability where users receive automated enforcement decisions without a clear explanation or straightforward route of appeal.