Meta Leverages Internal Data to Enhance Its AI Models

Meta has decided to tap into a new source of data to train its artificial intelligence models: the interactions of its own employees. By analyzing mouse movements and keyboard strokes of its staff, the company hopes to enhance the efficiency and capability of its AIs.

This initiative, initially revealed by Reuters, highlights the efforts of tech giants to diversify their training data sources, which are essential for developing effective AI models. This data allows programs to learn to perform tasks more efficiently and to better respond to user requests.

When asked by TechCrunch, a spokesperson for Meta explained: “If we are building agents to help people accomplish everyday tasks on computers, our models need real-world examples of how people actually use them — things like mouse movements, clicking on buttons, and navigating dropdown menus.” To achieve this, Meta has implemented an internal tool that captures these interactions within certain applications, while ensuring that protective measures are in place to safeguard sensitive content and that the data will not be used for other purposes.

However, this approach raises concerns regarding privacy. The AI industry seems increasingly to view internal communications as a valuable resource to fuel its models. Recently, it was reported that startups have been leveraging archives of their internal communications, such as Slack and Jira, to extract useful data for AI training.