Meta is rolling out a new internal tracking system called the Model Capability Initiative (MCI) that will log U.S. employees’ mouse movements, clicks, keystrokes and periodic screenshots on work computers, in a bid to generate the real‑world interaction data its next generation of AI agents need to function more like human coworkers. The move, disclosed in internal memos and confirmed by the company, has already sparked a backlash inside Meta and fresh scrutiny from privacy advocates, who see the program as a test case for how far Big Tech is willing to go to feed its AI models.
What MCI actually does
According to internal memos seen by Reuters and Business Insider, MCI is a software agent that runs on U.S.‑based employees’ work computers and:
- Captures mouse movements, click locations and keystrokes.
- Takes occasional screenshots of what’s on‑screen “for context.”
- Operates only on a pre‑approved list of work‑related apps and URLs, such as Gmail, Google Chat, Metamate (Meta’s internal tool) and VS Code.
Employees see a pop‑up instructing them to enable the tool, which is being rolled out to U.S. full‑time staff and contingent workers, the internal announcement says. A Meta spokesperson confirmed to Reuters that the data collected will be “one of the inputs” used to train its AI models.
Meta stresses that MCI does not run on personal phones and is restricted to corporate laptops and desktops. The company says sensitive content will be filtered out and that existing security and privacy rules for internal data will apply to MCI logs.
The AI strategy behind the tracking
Internally, MCI sits inside a broader push called the Agent Transformation Accelerator (ATA) and is being spearheaded by Meta SuperIntelligence Labs, the model‑building group behind the new “Muse Spark” large language model.
The goal is to train workplace AI agents that can:
- Navigate menus, dialog boxes and dropdowns in real software.
- Use keyboard shortcuts efficiently.
- Recover from mis‑clicks and error messages the way humans do.
An internal note quoted by Business Insider reads: “For agents to understand how people actually complete everyday tasks using computers, we need to train our models on real examples. This is where all Meta employees can help our models get better simply by doing their daily work.”
Meta’s CTO Andrew Bosworth has told employees the company would “ramp up internal data collection” as part of its agent strategy, arguing that synthetic benchmarks and web‑scraped logs aren’t enough to capture the nuance of enterprise workflows. Intellectia.ai, summarizing Meta investor materials, says the company sees MCI as key to boosting its models’ ability to simulate human‑computer interactions, one of the hardest gaps in current AI systems.
The company has already invested heavily in AI content‑moderation systems like Few‑Shot Learner and general‑purpose multimodal models, but MCI marks a shift toward fine‑tuning on first‑party interaction data to power autonomous “coworker” agents.
“Not for performance reviews,” Meta insists
One of Meta’s main talking points is that MCI is for AI training, not employee surveillance.
The internal announcement stresses that:
- MCI data “will not be used in performance reviews” or routine HR evaluations.
- The scope is limited to work apps and sites, not personal browsing.
- Meta employees’ work devices have long been monitored for security and compliance, and MCI is “more an extension of existing rules than a brand‑new policy shift.”
“While employees can’t opt out of the tracking software program, staff are informed when they onboard that activity on work devices may be monitored,” one person familiar with the program told Business Insider.
The company says it is implementing safeguards to exclude clearly sensitive categories, such as password fields or certain HR systems, from training datasets, though details of those filters aren’t public.
Employee backlash and privacy concerns
Despite those assurances, internal chat logs obtained by Business Insider and other outlets show unease and anger among staff.
Concerns voiced include:
- A sense that “everything we do is now going into a black box,” as one employee put it.
- Fear that MCI could be repurposed in the future for productivity scoring or disciplinary action, whatever current promises say.
- Worries about accidental capture of confidential information, including early‑stage product plans, legal docs, or private conversations.
Privacy and labor advocates outside Meta argue that MCI normalizes AI‑enhanced workplace surveillance under the banner of innovation. They note that even if today’s stated purpose is model training, the existence of high‑fidelity logs of everyone’s digital behavior creates both temptation and risk: for management to use it in other ways, or for attackers to target it.
The program also arrives just weeks after a “sev‑1” security incident at Meta in which an autonomous AI agent posted incorrect technical advice internally, raising questions about guardrails around AI tools inside the company. Critics say that incident underscores the stakes of feeding more sensitive internal behavior into powerful models.
How MCI fits Meta’s wider AI ambitions
MCI is launching as Meta moves to position itself as a top‑tier AI lab, not just a social network company.
The company recently unveiled Muse Spark, a new flagship language model that internal tests say is competitive with Google’s Gemini and some OpenAI systems and has signaled it may spend up to $135 billion this year on data centers and GPUs. SuperIntelligence Labs, the unit overseeing both Muse Spark and MCI, is central to that bet.
Meta’s long‑term vision is to embed AI agents across its products and internal workflows: writing code, assisting HR, triaging support tickets, and automating routine tasks for employees and, eventually, for enterprise customers. To do that, it needs models that understand not just language and images, but the friction points of real software, the difference between a mis‑click and a deliberate choice, the way humans scan menus, the shortcuts they use when in a hurry.
In that sense, MCI is less about spying and more about instrumenting the workplace as a training ground. But the line between those two perspectives is exactly where the current debate sits.
Legal and ethical questions ahead
From a compliance standpoint, Meta argues that:
- Employees consent to device monitoring through employment agreements.
- Data collection is limited to work contexts on company hardware.
- AI‑training use falls within existing internal‑data policies.
Yet regulators in the U.S. and EU are already probing how tech giants use employee and user data to train AI, and MCI is likely to draw their attention. Key questions include:
- Whether staff have meaningful choice if they cannot opt out.
- How long raw MCI logs are retained and who can query them.
- How Meta will document and audit the removal of sensitive content from training sets.
Ethically, the program sits at the intersection of two trends: the push for ever larger, richer AI training corpora, and the quiet expansion of digital Taylorism, where every interaction at work can be recorded, analyzed, and optimized.
For now, Meta insists that the payoff, AI agents that can help people with real tasks, not just chat, justifies the experiment. Whether employees, regulators and the wider public agree will go a long way toward determining how far Big Tech can go in turning the workplace itself into training data.