Google, LinkedIn and Wikipedia are among hundreds of websites and apps where Meta plans to capture employee keystrokes and mouse clicks as part of a project to train its artificial intelligence models, according to internal messages viewed by CNBC.
A new employee tracking tool, dubbed Model Capability Initiative (MCI), allows Meta to observe and collect data from staffers’ actions on their work computers, Reuters first reported on Tuesday. The list of sites being tracked, which also includes Microsoft’s GitHub, Salesforce’s Slack and Atlassian, has not been previously reported.
Meta properties like Threads and Manus are also on the list, which is still in flux and originally included AI apps like OpenAI’s ChatGPT and Anthropic’s Claude.
The list of third-party sites and services the MCI tool is tracking was widely circulated internally and discussed on chat boards after a member of the Meta Superintelligence Labs, or MSL, sent a memo intended to assuage concerns about worker surveillance and privacy. CNBC viewed the memo.
The data gathering project is tied to Meta CEO Mark Zuckerberg’s ambitious effort to catch up in generative AI, where the company has lagged behind OpenAI, Anthropic and Google. To try and close the gap, Zuckerberg went on a spending spree starting last summer, bringing in Scale AI’s Alexandr Wang to build a team and develop new foundation models.
Earlier this month, Meta unveiled its first major AI model since the costly hiring of Wang. Dubbed Muse Spark, the model marked the debut of the new Muse series developed by MSL, the AI unit that Wang oversees.
Like other tech giants, Meta is pushing hard into AI agents that can perform various office and coding-related tasks that are typically accomplished by white-collar workers.
A Meta spokesperson confirmed the project but didn’t provide a comment on the list of sites being tracked.
“If we’re building agents to help people complete everyday tasks using computers, our models need real examples of how people actually use them — things like mouse movements, clicking buttons, and navigating dropdown menus,” the spokesperson said. “To help, we’re launching an internal tool that will capture these kinds of inputs on certain applications to help us train our models. There are safeguards in place to protect sensitive content, and the data is not used for any other purpose.”
Multiple Meta employees characterized the data-tracking project as “dystopian” in internal messages viewed by CNBC. Others expressed concerns that MCI could widely expose sensitive data, including user passwords, details about new product development, and personal information about workers’ immigration status, health or family members.
The MSL staffer said in the memo that in order to “teach our models to be able to use computers,” Meta requires a “big and unbiased” data set that reflects how employees work and do tasks on their corporate devices.
“We need to capture on-screen content as the context of what was being manipulated or interacted with,” the memo said.
In listing a “few assurances,” the MSL representative noted that the new tool would only be able to view employees’ “screen contents” as they see them, and would “not read in files or attachments.”
“Any incidental personal information in your corporate email that may get captured from the screen, will not be learned by the model, due to the mitigations above,” the memo said.
Meta employees who are still concerned about the data-tracking tool, “can control what shows up on your screen by not doing personal work on your work computer,” the memo said.
WATCH: AI demand metrics are broken and only Anthropic is being realistic.
