Meta has urgently suspended an AI training project called the "Model Capabilities Initiative (MCI)" on June 23. The project previously trained AI models by recording internal employees' keyboard typing data and mouse movement tracks. However, due to a recent serious internal data breach, Meta was forced to halt the program.

Meta, Metaverse, Facebook

It is reported that this incident caused a large amount of highly sensitive information collected through the program — including private employee conversations, work performance data, and voice-to-text records — to be accidentally made accessible to all Meta employees. Although the project had previously caused widespread internal dissatisfaction because it placed employees under near-constant monitoring and was suspected of violating privacy regulations, the direct trigger for Meta's emergency action was the collapse of its data security barriers.

A Meta spokesperson responded that the company had built privacy protection mechanisms when designing the program, and there is currently no evidence that any employee has violated the rules by accessing the leaked data. However, the company has decided to completely suspend the project and conduct a thorough investigation.

This incident once again brings the issues of data boundaries and privacy security in AI large model training by tech giants to the forefront. As companies accelerate the use of multimodal and behavioral data to train AI, finding a balance between improving model capabilities and ensuring data compliance and employee privacy has become a pressing challenge for the entire generative AI industry.