Meta, the tech giant behind Facebook and Instagram, is embarking on an ambitious, and potentially controversial, initiative to gather employee keystrokes and mouse clicks from a wide array of websites and applications. This data collection effort, aimed at training its burgeoning artificial intelligence models, spans platforms including Google, LinkedIn, and Wikipedia, according to internal documents obtained by CNBC. The project, codenamed Model Capability Initiative (MCI), allows Meta to observe and log user actions on work computers, a detail first reported by Reuters. The scope of tracked third-party sites is more extensive than previously disclosed, encompassing platforms like Microsoft’s GitHub, Salesforce’s Slack, and Atlassian, in addition to Meta’s own properties such as Threads and Manus. Early versions of the tracking list also included prominent AI applications like OpenAI’s ChatGPT and Anthropic’s Claude.
The MCI tool’s data acquisition strategy was disseminated internally and sparked discussions on company chat boards following a memo from a member of Meta Superintelligence Labs (MSL), an AI unit overseen by Alexandr Wang, formerly of Scale AI. The memo was intended to address employee concerns regarding surveillance and privacy. This push for data underscores Meta CEO Mark Zuckerberg’s intensified focus on closing the generative AI gap with competitors like OpenAI, Anthropic, and Google. Following a significant investment in AI talent, including Wang’s recruitment last summer to spearhead the development of new foundation models, Meta recently unveiled its Muse Spark, the first major AI model under the Muse series developed by MSL.
In alignment with the broader industry trend, Meta is heavily investing in AI agents capable of performing a range of office and coding-related tasks typically handled by white-collar professionals. A Meta spokesperson confirmed the existence of the MCI project, stating, “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. 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.”
Despite these assurances, the initiative has drawn significant internal criticism. Multiple Meta employees have described the project as “dystopian” in internal communications viewed by CNBC. Concerns have been raised about the potential for MCI to inadvertently expose sensitive information, including user passwords, details of unreleased product developments, and personal data related to immigration status, health, or family members.
The MSL staffer’s internal memo emphasized the need for a “big and unbiased” dataset to train AI models for computer interaction. “We need to capture on-screen content as the context of what was being manipulated or interacted with,” the memo stated. While assuring employees that the tool would only view “screen contents” and would “not read in files or attachments,” the memo also acknowledged that “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.” For employees remaining apprehensive, the memo offered a stark suggestion: “you can control what shows up on your screen by not doing personal work on your work computer.” This internal directive highlights the delicate balance Meta is attempting to strike between its aggressive AI development agenda and the potential erosion of employee trust and privacy.
Original article, Author: Tobias. If you wish to reprint this article, please indicate the source:https://aicnbc.com/20928.html
