googleLinkedIn and Wikipedia are among the hundreds of websites and apps. Meta The company plans to capture employee keystrokes and mouse clicks as part of a project to train artificial intelligence models, according to internal messages seen by CNBC.
A new employee tracking tool called Model Capabilities Initiative (MCI) will allow Meta to observe and collect data on what employees do on work computers, Reuters first reported on Tuesday. List of sites being tracked (also included) microsoft’s GitHub, sales force Slack and atlassianwhich had not been previously reported.
Meta properties like Threads and Manus are also included in the list, which is still in flux and initially included AI apps like OpenAI’s ChatGPT and Anthropic’s Claude.
The list of third-party sites and services tracked by the MCI tool was widely circulated internally and discussed on chat boards after members of the Meta Superintelligence Lab (MSL) sent out a memo aimed at allaying employee surveillance and privacy concerns. CNBC has viewed the memo.
This data collection project is a collaboration between Meta CEO Mark Zuckerberg, OpenAI, Anthropic, google. To close the gap, Zuckerberg began spending heavily last summer, bringing in Alexandr Wang of Scale AI to build a team and develop a new underlying model.
Earlier this month, Meta unveiled its first major AI model since hiring Mr. Wang at a high cost. The model, named Muse Spark, marked the debut of the new Muse series developed by MSL, the AI unit overseen by Wang.
Like other tech giants, Meta is focused on developing AI agents that can perform a variety of office and coding-related tasks typically performed by white-collar workers.
A Meta spokesperson confirmed the project but would not comment on the list of sites being tracked.
“If we’re building an agent to help people use their computers to complete everyday tasks, our models need real-world examples of how people would actually use the agent, such as moving a mouse, clicking a button, or navigating a drop-down menu,” the spokesperson said. “To assist with this, we are launching internal tools to capture this type of input in certain applications to help train models. Safeguards are in place to protect sensitive content and the data will not be used for any other purpose.”
In internal messages seen by CNBC, multiple Meta employees characterized the data tracking project as “dystopian.” Others expressed concern that MCI could widely release sensitive data, including user passwords, new product development details, and personal information about workers’ immigration status, health status, and family members.
In a memo, MSL staff said Meta said that in order to “teach models to use computers,” they need “large, unbiased” datasets that reflect how employees work and perform tasks on company devices.
“On-screen content must be captured as the context of an action or interaction,” the memo says.
An MSL representative cited “several guarantees” and noted that the new tool can only see “what’s on screen” exactly as employees see it, and “cannot read files or attachments.”
“Ancillary personal information in corporate emails that may be captured from the screen will not be learned by the model due to the mitigations described above,” the note states.
Meta employees who still have concerns about data-tracking tools can “control what they see on their screens by not doing any personal work on their work computers,” the memo said.
Note: AI demand metrics are broken and only human beings are realistic.
