Meta Halts Its AI After Internal Data Leak

Meta Interrupts Its AI Program Following Data Leak

Andrew Bosworth, Meta's Chief Technology Officer, recently revealed details about a data leak that affected the company's artificial intelligence training program. This leak has led to a temporary suspension of the program since last June.

In an interview with Nicholas Thompson, CEO of The Atlantic, Bosworth explained that the incident occurred when an internal researcher accidentally moved data to an inappropriate location. This interview, although published recently, was recorded at the end of June.

Bosworth clarified that the data in question, derived from the training program, was initially well-protected. Only a few individuals had access to this information. However, human error led to its unauthorized relocation by a member of Meta's research team.

• "A researcher working with this data placed it in an inappropriate location, but there was no security breach," he stated to Thompson.

Employee data, although transformed, was placed in an internal location where it should not have been. Bosworth emphasized that Meta suspected no malicious intent behind this incident.

In response to this event, the company decided to "lock down the entire system" to better understand the circumstances surrounding the leak, he added.

The Model Capability Initiative and Its Repercussions

Launched in April, the Model Capability Initiative aimed to install software on the computers of the majority of Meta employees in the United States. This software was intended to track their keystrokes and mouse movements to improve the company's AI models. However, this initiative faced strong opposition from employees, particularly due to the prohibition of opting out of the program.

Bosworth acknowledged in an internal meeting that employee morale was particularly low, an unprecedented situation in Meta's history.

The program was suspended in June after a leak exposed sensitive employee data to the entire company, as shown by screenshots obtained by Business Insider.

• "We designed this program with privacy safeguards, and although we have no evidence of inappropriate access by Meta employees, we are suspending it to investigate," a Meta spokesperson told Business Insider.

Data Variability Issues

During the interview, Bosworth also mentioned another challenge faced by the program. It generated a large amount of similar data, whereas the goal was to obtain diverse data to effectively train the AI.

• "Data diversity is far more valuable than a large volume of similar data," he explained to Thompson.

To address this, Meta has expanded opt-out options for employees who do not wish to participate, allowing for a pause at any time.

Meta representatives chose not to comment further on the matter when Business Insider reached out to them.