A Reuters report published on April 21, 2026 confirmed that Meta will begin monitoring its employees' computers in the United States to train its next artificial intelligence models. The initiative, internally called Model Capability Initiative (MCI), will capture mouse movements, clicks, keystrokes, and occasional screen snapshots while employees use work-related applications and websites.
The stated goal is to build AI agents capable of autonomously executing computer tasks, such as navigating menus, filling forms, selecting dropdown options, and using keyboard shortcuts. Meta claims the data will not be used for performance evaluation, only for model training. Even so, the decision triggered immediate reactions from privacy communities, labor lawyers, and technology professionals.
In this article, we break down what MCI actually is, why Meta needs this data, how it fits into a broader Silicon Valley trend, the ethical implications, and most importantly, what would change if the same program ran in Brazil under LGPD. If you are a business owner, CTO, HR leader, or develop software, this text provides the technical and legal context you need to decide how to respond.
In this article
- What happened - the Reuters report and Meta's internal announcement
- What MCI is - technical scope of the Model Capability Initiative
- Why Meta needs this data - the autonomous agent bottleneck
- Meta's official position - statements from Andy Stone and Andrew Bosworth
- The larger trend - Microsoft Recall, GitHub Copilot and corporate data dogfooding
- Ethical implications - consent, compensation, and conflict of interest
- What if it happened in Brazil? - what LGPD and labor law say
- What changes for software companies
- Ethical and legal checklist for similar projects
- The future of work with AI agents
- Frequently asked questions
What Happened: The Reuters Report on Meta
On April 21, 2026, Reuters published an exclusive report based on internal Meta documents and conversations with company employees. According to the investigation, Meta's CTO, Andrew Bosworth, internally communicated that the company would begin installing monitoring software on employee computers in the United States.
The story was quickly replicated by global outlets such as Fortune, TechCrunch, and The Irish Times. In Brazil, portals like InfoMoney and Mundo Conectado translated and contextualized the news.
The central point is that the initiative does not involve external customers or users. The target is Meta's own employees. The company is already one of the world's largest data collectors via Facebook, Instagram, and WhatsApp, but this specifically operational data collection, within the corporate work environment, creates a new category of data capture for model training.
What Exactly is the Model Capability Initiative (MCI)
The Model Capability Initiative is an internal Meta program that installs a monitoring agent on employees' work machines. According to Reuters and outlets that obtained access to the internal announcement, the software collects:
- Mouse movements - cursor paths on screen, speed, and movement patterns
- Clicks - which UI elements are clicked, click duration, double-clicks
- Keystrokes - key sequences, shortcut usage (Ctrl+C, Cmd+Tab, etc.)
- Screen snapshots - occasional captures of what is displayed at the moment
- Application metadata - which app or site is in focus, window title, URL
The collection is neither continuous nor universal. According to Meta, MCI operates only in a pre-defined list of work-related applications and websites, avoiding capture of personal activities. There are also mechanisms to pause capture in sensitive contexts, although the company has not publicly specified which types of content are automatically excluded.
What MCI Does NOT Collect (According to Meta)
Meta stated that MCI does not capture:
- Personal employee accounts (personal email, own social networks)
- Activities outside the monitored corporate app list
- Data used to evaluate individual performance
- Information that individually identifies the collaborator in the final model training (the process includes anonymization)
The problem, as flagged by privacy experts, is that none of these guarantees are externally auditable. The employee cannot verify that capture really stops when they open their personal browser, nor how anonymization happens in the training pipeline.
Why Meta Needs This Data: The Autonomous Agent Bottleneck
To understand Meta's decision, you need to understand a real technical problem affecting the entire AI industry in 2026: large language models (LLMs) are excellent at conversing but poor at operating computers like humans. We cover this transition in more depth in our article about the impact of artificial intelligence in Brazil and the rise of autonomous agents.
A modern AI agent needs to perform tasks like:
- Open a corporate application (e.g., a CRM)
- Navigate menus until finding the right form
- Fill fields with contextual data
- Use keyboard shortcuts to speed up repetitive operations
- Interpret validation errors and correct them
- Switch between tabs, copy data, paste into another system
These actions seem trivial to a human but involve a layer of tacit operational knowledge that is not in any text from traditional LLM training. You don't find on a Wikipedia article the exact click sequence to fill a SAP report, nor the best way to copy data from a spreadsheet to a Tableau dashboard.
This bottleneck was documented in benchmarks like WebArena, OSWorld, and VisualAgentBench, which show that even top-tier models like GPT-4o, Claude Sonnet, and Gemini have success rates below 40% on desktop automation tasks in 2025. To overcome this limit, AI labs need real demonstrations of how humans use computers in professional contexts. That is exactly what Meta is collecting with MCI.
The Imitation Learning Concept
The technique behind this is known as imitation learning or, more recently, behavioral cloning. The model does not learn by trial and error in a simulated environment; it learns by observing thousands of hours of real human trajectories and trying to reproduce those patterns.
The issue is that human trajectory data on corporate computers is rare and expensive. There are small public datasets (like Mind2Web), but nothing near the scale Meta needs to train a competitive agent. The company calculated that collecting this data from its own employees is cheaper, more diverse, and better aligned with real corporate use cases than hiring external annotators or buying datasets.
Meta's Official Position: What the Company Declared
In response to Reuters' inquiries, Meta spokesperson Andy Stone issued an official statement stating:
- Collected data will be used only for AI training, never for individual performance evaluation
- There are technical safeguards to protect sensitive content, although the list of these safeguards was not disclosed
- Employees are informed in advance about the program and there is an internal communication process
- MCI is part of a larger initiative called AI for Work, led by CTO Andrew Bosworth
In the internal statement attributed to Bosworth, he reportedly said the company needs this data to build useful agents: "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 logic presented by Meta is consistent with the company's strategic direction: after investing billions in AI infrastructure, the company needs its models to outperform competitors in pragmatic use cases. Without operational data, it is unlikely to achieve that goal.
The Larger Trend: Corporate Data Dogfooding
Meta's initiative is not an outlier. It fits a trend that industry employees call "data dogfooding" - when a company uses its own employees as data sources to train products that will later be sold to third parties.
GitHub Copilot and Internal Collection
In parallel with the Meta news, GitHub announced that, starting April 24, 2026, it will begin using user and employee interaction data to train GitHub Copilot. The format is opt-out, meaning all users are automatically enrolled and must manually leave if they don't want to participate. Collection includes prompts, generated code, suggestion acceptance or rejection, and editor usage patterns.
Microsoft Recall and Continuous Screen Capture
Microsoft launched in 2024 and refined in 2025 the Recall feature, which takes continuous screenshots of the user's PC to enable subsequent contextual search. After security controversies, Microsoft made the feature opt-in with local encryption, but the precedent of continuous screen capture was established on a scale of tens of millions of devices.
OpenAI, Anthropic, and Trajectory Datasets
Both OpenAI (with Operator) and Anthropic (with Computer Use) launched in 2024 and 2025 agent products that interact with computers. Both companies massively hired annotators to generate usage trajectories in virtualized environments. The difference from MCI is that Meta is using internal workers in real environments, producing superior data quality but raising questions that outsourcing to professional annotators did not raise.
The big picture is clear: all big techs understood that the next leap in AI depends on real operational work data. Meta just took a more explicit step, capturing this data directly from employees, without intermediaries.
Ethical Implications: Three Core Issues
Three ethical questions appear repeatedly in journalistic coverage and in statements from specialists analyzing MCI.
1. Consent Is Not Free When You Depend on the Job
Even if Meta claims to inform employees, consent in labor relations is legally fragile. A worker who depends on salary to pay rent is not in a symmetric position to negotiate with the employer. In data privacy, this is the classic problem of "coerced consent".
The European Union, via GDPR, has already recognized this problem: employee consent rarely serves as a valid legal basis for data processing because the freedom requirement is absent. The basis must be something else (legitimate interest, contract performance, legal obligation). The same principle tends to apply in Brazil under LGPD.
2. Compensation: You're Training the AI That Will Replace You
The structural irony of MCI is that captured data helps train agents designed to automate tasks currently done by humans. Meta employees are, effectively, teaching the system that can reduce the need for their own work in the future.
The discussion about compensation for training data has existed in other sectors for years (artists vs. image models, authors vs. LLMs), but in the case of corporate employees it gains a new angle: payment for observed work does not include the transfer of operational data. Would an additional be fair? Participation in the results of the generated product? The answer is still open.
3. Scope Creep and Accidental Capture of Sensitive Data
Meta promises to capture only "work-related applications and sites". But the boundary is fuzzy:
- An employee who uses the corporate browser to log into a bank during work hours had banking data captured?
- A chat with a colleague about a medical problem in a corporate Slack enters the dataset?
- An engineer who types an API key in a monitored IDE saw their key captured in a snapshot?
Even with legitimate intent, the effective scope of what is captured tends to leak beyond what was planned. This is the type of risk that the LGPD addresses explicitly in article 46, requiring technical and administrative measures to protect personal data.
What If It Happened in Brazil? LGPD, Labor Law, and Employee Monitoring
The Reuters report covers the program in the United States, where federal workplace privacy law is state-fragmented and relatively permissive. In Brazil, the legal landscape is completely different. If a company wanted to implement a program equivalent to MCI with Brazilian workers, it would face three regulatory layers simultaneously.
Layer 1: LGPD (Law 13.709/2018)
The National Data Protection Authority (ANPD) and Brazilian courts have already made it clear that data captured in the workplace is personal data when linkable to an identifiable person. That means:
- Needs legal basis (art. 7) - pure consent probably doesn't hold due to labor relationship asymmetry. The company would have to invoke legitimate interest with a documented balancing test.
- Needs principles (art. 6) - especially purpose, necessity, adequacy, and non-discrimination.
- Needs Data Protection Impact Assessment (DPIA) - mandatory for high-risk processing, which is clearly the case.
- Needs transparency - informing exactly what is collected, for how long, with whom it is shared, and how to exercise rights.
Layer 2: Labor Law and Employer Management Power
Brazilian labor law recognizes the employer's management and oversight power over corporate resources. TST (Superior Labor Court) jurisprudence has consolidated that the employer can monitor corporate email, company-provided equipment usage, and work-hour browsing, provided there is prior and unequivocal employee awareness.
The problem with MCI in Brazil is not monitoring itself, but secondary data use. Overseeing productivity is one thing; using an employee's mouse movements to train a commercial product sold to third parties is another. This secondary use extrapolates management power and enters intellectual property and economic participation territory that labor law does not directly address.
Layer 3: Union and Collective Actions
In Brazil, large-scale monitoring programs tend to be subject to collective bargaining via unions. Unilateral implementation of a Brazilian MCI would face challenges through public civil action (Labor Prosecutor's Office), conduct adjustment terms, and eventual collective action. Companies that tried to implement aggressive monitoring systems in Brazil in recent years (e.g., cases involving call center platforms) reached unfavorable judicial outcomes.
Practical conclusion: Meta's MCI, as designed for the US, would not pass LGPD and labor law filters without deep modifications. Any Brazilian company attempting to replicate this practice needs to rethink consent, legal bases, DPIA, and collective bargaining before moving a line of code.
What It Means For Companies That Develop Software
If you lead a company that develops software or applications, there are three practical moves to consider in 2026.
1. Review Your Development Contracts and NDAs
Projects involving AI tools (Copilot, Cursor, Codium, Windsurf, Claude Code) are generating developer interaction data. That data may be going to train vendor models. Corporate client contracts need to make clear:
- Which AI tools are used in the development process
- Whether prompts and source code are sent to third parties
- Whether there is training opt-out (most enterprise plans offer it)
- Who is responsible for IP leaks
In our experience developing more than 30 corporate apps, this clause became mandatory in fintech, healthcare, and government projects in 2026. It's a technical discussion that quickly became legal.
2. Plan AI Usage Considering Data Origin
If your company plans to integrate AI into the software development workflow or in end products, treat training data origin as a vendor selection criterion. Models trained on ethically questionable data can create reputational and legal liability for those using them in the final product.
3. Don't Replicate MCI Without Legal Shielding
It's tempting to want to train internal AI agents using your own employees' data. If you have a similar idea for your company, stop and consult the DPO and labor lawyer before implementing. The apparent economy of collecting data from your own employees vanishes when compared to the risk of a collective labor action and ANPD fines.
Ethical and Legal Checklist For AI Projects With Employee Data
If your company is evaluating any capture of operational data from employees to train models, run this checklist before executing.
| Item | Question | Where to answer |
|---|---|---|
| Legal basis | What LGPD legal basis justifies the processing? | DPO opinion |
| Purpose | Is the purpose documented, specific, and aligned with the minimization principle? | Internal privacy policy |
| Scope | Which applications, windows, URLs, and content types are captured? And excluded? | Technical document + DPIA |
| Anonymization | Are data anonymized before training? How is it audited? | Technical pipeline + external audit |
| Retention | How long is raw data stored? | Retention policy |
| Transparency | Do employees receive clear, plain-language communication about the program? | Onboarding document |
| Consent | Is there real opt-out without (implicit or explicit) penalty? | HR policy + exercise channel |
| Union | Has the union been consulted or is there a clause in collective agreement? | Bargaining record |
| Secondary use | Will data be used for evaluation, dismissal, or shared with third parties? | Data use terms |
| Data subject rights | Is there a channel for the employee to access, correct, delete their data? | Data subject portal (LGPD art. 18) |
If any line is blank or incomplete, the project is not production-ready. Well-intentioned AI projects die in court due to gaps here. For a structured development approach, see our guide on strategic mobile app development in 2026.
The Future of Work With AI Agents
MCI is a clear indication of a transformation already underway: autonomous agents will execute tasks that today occupy much of knowledge workers' days. 2026 reports from McKinsey, Gartner, and IDC converge on the estimate that between 30% and 45% of office tasks can be automated by agents by 2030.
To understand how this already affects customer service with chatbots and autonomous agents, software development, and back-office operations, it's worth observing three movements:
- Vertical products with specialized agents - startups are building agents for niches (legal, health, accounting) rather than generalist agents. See also the seven AI features successful apps will have in 2026.
- Operational data value grows - companies that document processes in structured format will have a competitive advantage when training their own agents. Well-described process becomes a technological asset.
- Worker role shifts - from execution to supervision. Workers will spend more time reviewing what the agent did and less time manually executing.
Meta's decision accelerates this cycle. If it works, other big techs will follow. If it fails (legally or publicly), it will be a case study on how far AI training can go before hitting fundamental rights.
Frequently Asked Questions
What is Meta's Model Capability Initiative (MCI)?
MCI is an internal Meta program that installs software on employees' computers in the US to capture mouse movements, clicks, keystrokes, and screen snapshots during corporate app usage. The data will be used to train AI agents capable of autonomously executing tasks on computers.
Does Meta monitor all employees, all the time?
No, according to Meta itself. MCI operates only on a pre-defined list of work-related apps and sites, and only in the US in the initial phase. Activities on personal accounts and outside the corporate scope would not be captured. Even so, experts point out that these guarantees are not externally auditable.
Would MCI be legal in Brazil under LGPD?
Not in the form designed for the US. LGPD requires valid legal basis, DPIA, transparency, minimization principle, and rights exercise channel. In addition, Brazilian labor law and jurisprudence limit secondary use of data collected in the workplace. Implementing a Brazilian MCI would require deep modifications and union bargaining.
Will my data as a Facebook or Instagram user be used in MCI?
No. MCI is specific to Meta employees in the US and does not involve data from external platform users. Meta already uses user data to train AI in other programs, but that is a separate topic regulated by distinct privacy policies.
What should my company do if it wants to train internal AI agents?
Before implementing any collaborator data capture, consult the DPO, labor lawyer, and category union. Document legal basis, purpose, scope, anonymization, retention, and rights exercise channel. Running the checklist from this article is a good starting point, but production operation requires specific legal opinion.
Which technologies compete with MCI to train AI agents?
OpenAI (Operator), Anthropic (Computer Use), Google (Gemini Agent), and several academic labs use alternative approaches: simulated environments, contracted annotators, opt-in voluntary user capture, and public datasets like Mind2Web and WebArena. MCI's differential is data volume and realism, at the cost of ethical controversy.
How does MCI relate to Microsoft Recall and GitHub Copilot?
All three are part of a trend called "data dogfooding": using their own employees or users as data sources to train products. Recall captures continuous snapshots on the user's PC. Copilot collects prompts and suggestion acceptance patterns. MCI does the equivalent within Meta. All three raise similar questions about consent, anonymization, and secondary use.
Can this practice affect software development contracts in Brazil?
Yes. In 2026, B2B contracts started including clauses about AI tools used in the development process, training opt-out, and liability for leaks. Companies that hire external development should audit the AI stack used by the vendor, especially in projects involving sensitive data or critical intellectual property.
Next Step
Meta's decision marks a turning point in how big tech companies collect data to train AI. The line between end-user and training data supplier has become thinner, and the legal and ethical debate will only intensify in the coming months.
If your company is planning projects involving AI, autonomous agents, process automation, or corporate data analysis, the right time to design the legal and technical architecture is before the first line of code, not after the first lawsuit.
FWC Tecnologia has been developing apps and web systems with integrated AI since 2022, with specific attention to LGPD compliance, secure data architecture, and ethical model usage practices. To understand how this applies to your project, request a custom quote or run a quick cost estimate to talk to our team.