On April 21, 2026, Reuters published a report that triggered a wave of outrage in the tech industry: Meta, the company behind Facebook, Instagram, and WhatsApp, has begun collecting data on how its own employees use computers. The program records mouse movements, keystrokes, clicks, and occasionally takes screenshots of the screen. The goal is to train AI agents to perform typical office tasks.
What exactly Meta collects and why
According to Reuters, Meta sent employees an internal memorandum through the Meta Superintelligence Labs team channel. It concerns a new initiative called Model Capability Initiative — a monitoring tool installed on work computers that tracks interaction with certain apps and websites.
Meta’s official position looks like this: the company is building AI agents capable of helping people perform ordinary computer tasks. For such agents to work realistically, they need real examples of human behavior — not synthetic data.
A Meta spokesperson commented on the situation as follows: “If we are building agents to help people with everyday computer tasks, our models need real examples of how people actually use computers — mouse movements, button clicks, navigation through dropdown menus. For this, we are launching an internal tool that will capture such interactions in certain applications. There are safeguards to protect sensitive content, and the data is not used for any other purposes.”
At the same time, the memorandum seen by Reuters told employees that the data would not be used to evaluate productivity. Whether that should be believed is another question.
Context: where else tech companies are looking for AI data
This news is not isolated. It reflects a broader alarming trend: the artificial intelligence industry is exhausting traditional data sources and is looking for new ones, often in the most unexpected places.
In just the past few months, several such initiatives have become known:
- OpenAI, through contractor Handshake AI, asked freelancers to upload real work materials from previous jobs — PowerPoint presentations, Excel spreadsheets — after formally “removing confidential data”
- Zombie startups: according to Forbes, old shuttered startups are being deliberately acquired for their corporate archives — Slack conversations, Jira tickets, internal email chains
- Scale AI: last year Meta acquired a 49% stake in this data-labeling company for more than $14 billion. Former Scale AI CEO Alexandr Wang now leads Meta Superintelligence Labs — the division directly responsible for the new monitoring initiative
Taken together, these facts paint a picture: corporate communications that were still internal yesterday are turning into fuel for machine learning. The boundaries between “workplace” and “training dataset” are rapidly blurring.
What privacy experts think about this
The problem of consent and anonymization
Even if Meta is sincere in its promise not to use the data for productivity evaluation, that alone does not solve the problem. To train AI on real patterns of human-computer interaction, those patterns need to be captured in their authentic form. Any filtering or anonymization potentially reduces the value of the training data — and here a fundamental contradiction emerges between usefulness and privacy.
The legal difference between the US and Europe
The legal landscape differs significantly depending on the jurisdiction. In the European Union, the General Data Protection Regulation (GDPR) requires explicit employee consent and data minimization. California’s CCPA and new privacy laws in other states also call into question the legitimacy of large-scale data collection for AI training. Reuters specifically notes that the initiative applies to U.S. employees — not to the company’s entire global workforce, which may be a deliberate decision in light of stricter regulation in the EU.
The bigger goal: AI agents instead of office workers
To understand why Meta is taking such a step, it is worth looking at the strategic picture. The company has announced AI investments totaling $600 billion by 2028. At the same time, it is known that Meta is developing AI agents capable of performing typical office tasks — including a special AI agent for the personal use of CEO Mark Zuckerberg and a chatbot in the form of Zuckerberg for communicating with employees.
The task that the Model Capability Initiative is solving is to teach these agents to appear “human”: to move a mouse the way a human does, to click with the same behavioral logic, to use hotkeys, and to navigate complex menus. Without real examples of human behavior, synthetic data does not provide enough realism.
The Meta memorandum directly encouraged rank-and-file employees to “contribute” simply by doing their daily work. In other words — to train their own replacement without noticing it.
Reaction online and among employees themselves
The public reaction was predictably sharp. Critics point to the obvious: Meta is asking people to train systems that could potentially replace them, offering no additional compensation and requiring them to give up any sense of privacy in the workplace.
Trade press comments have appeared along the lines of: “Facebook employees used to train the surveillance system. Then the advertising algorithms. Now — their replacement.” Gizmodo described the initiative bluntly: “a surveillance tool — there is no point pretending it is anything else.”
Meta, for its part, insists that the data is protected, not used for performance evaluation, and not shared with third parties. But trust between employers and employees on AI data collection is already at a low point — and this news is unlikely to improve it.
What this means for ordinary users
Directly, this situation concerns only Meta’s U.S. employees. But in a broader context, it concerns everyone who uses tools with artificial intelligence elements in the workplace.
If your company is introducing AI assistants — ask: what data is being collected to train them? Where is this data stored? Do you have the right to opt out? What guarantees are there that corporate communications will not become training data without your knowledge?
These questions are already relevant today for large corporations. Tomorrow they will become standard for any office that uses AI tools.
In brief: the key points about Meta’s initiative
- What is happening: Meta is installing the Model Capability Initiative program on the computers of its U.S. employees
- What is recorded: mouse movements, keystrokes, clicks, screenshots
- Why: training AI agents in the realism of human interaction with computers
- Promises: data is not for productivity evaluation, protection of sensitive content
- Context: Meta is investing $600 billion in AI by 2028; the direction is led by Meta Superintelligence Labs
- Trend: OpenAI, zombie startups, Scale AI — corporate data is becoming “fuel” for AI across the industry
Article prepared by the TechVisor team — practical IT media for people.
