Meta pauses AI training programme after internal data leak

Meta temporarily suspended an internal artificial intelligence (AI) training programme after an employee researcher mistakenly stored sensitive data in an unauthorised internal location, according to Chief Technology Officer Andrew Bosworth.

Speaking during an interview with The Atlantic CEO Nicholas Thompson, Bosworth said there was no external security breach or evidence of malicious activity. However, the company decided to pause the initiative while it investigates how the data was handled, as per Business Insider.

Researcher mistakenly moved employee data

Bosworth said the data collected through Meta's Model Capability Initiative was securely stored and accessible to only a limited number of employees.

However, he explained that a researcher working with the data inadvertently placed it in an internal location where it should not have been stored.

The executive stressed that the information remained within Meta's systems and described the incident as an internal handling error rather than a cyberattack or data breach.

According to Bosworth, the company is keeping the programme paused until it fully understands what happened and strengthens its safeguards.

Programme tracked employee activity for AI training

Meta launched the Model Capability Initiative in April to collect employee behaviour data, including keystrokes and mouse movements, to help improve its AI models.

The programme sparked criticism among employees because participation was initially mandatory for most US staff, with no option to opt out.

The initiative was suspended in June after an internal data exposure reportedly made sensitive employee information accessible more broadly within the company.

Meta says privacy protections remain in place

Following the suspension, Meta said it had built the programme with privacy safeguards and had found no evidence that employee information had been improperly accessed.

The company described the pause as a precautionary step while it investigates the internal data handling issue.

Repetitive data also limited AI training

Bosworth said the data leak was not the programme's only challenge.

He explained that the initiative generated large amounts of similar behavioural data, which was less valuable for AI development than a broader variety of examples.

To improve the quality of the training data, Meta later expanded employee opt-out options, allowing staff to pause their participation whenever they chose.

The company has not announced whether or when the AI training programme will resume.