In April of this year, Meta rolled out an internal project called the Model Capability Initiative (MCI) on company computers used by its U.S. employees. It recorded keystrokes, mouse movements, page clicks, and screen content, then fed these daily work traces into training AI models. In other words, how employees wrote code, used internal tools, and communicated with colleagues on company computers could all become material for models learning real knowledge work. After running for approximately two months, Meta announced a suspension in late June. The story was first broken by WIRED, followed by coverage from The Guardian, BBC, and Business Insider. After the project launched, over 1,600 Meta employees signed a petition in protest.
Regarding the project’s suspension, outsiders might easily arrive at two speculations: forceful intervention by regulators, or Mark Zuckerberg concluding that the training failed to meet expectations. However, based on evidence publicly disclosed by various sources so far, neither speculation matches the facts.
For the suspension, Meta’s official reason was cooperation with an internal review. Meta spokesperson Tracy Clayton explained to WIRED that the company had designed privacy safeguards into the project, had found no evidence of improper data access to date, but decided to proactively pause the project to cooperate with the review. On the external front, regulatory bodies such as the Federal Trade Commission (FTC), the Equal Employment Opportunity Commission (EEOC), or the National Labor Relations Board (NLRB) have not issued any cease-and-desist orders against the MCI project. On the technical side, there is likewise no credible indication that the model’s training results fell short of expectations.
The brakes were applied to the project by two intertwining forces: a sudden internal data leak, and months of accumulated employee resistance.
First, internal access controls went unmanaged. Screen recordings that were supposed to be kept strictly confidential, along with employee daily communications and performance data, were exposed on the company intranet due to careless permission configurations, enabling employees to access data beyond the expected scope. According to tech outlet HR Brew, employees shared screenshots internally confirming that the sensitive data was at the time completely open. Business Insider likewise corroborated the leaked content, noting that it was intermingled with a large amount of private conversations and performance details. The monitoring program had already made employees uneasy, and the intranet security incident completely detonated a crisis of trust.
Second, intense employee resistance. The MCI project was implemented through a mandatory rollout model, offering employees no option to refuse or opt out. BBC cited an anonymous employee stating that the system was imposed on employees, with voluntariness entirely out of the question. As discontent spread, over 1,600 employees launched a petition, accusing the project of privacy violations, disregard for the right to informed consent, and serious harm to fundamental workplace trust.
Beneath this fierce resistance lies a deeper layer of professional anxiety and conflict of interest. For many knowledge workers, the unease caused by MCI goes beyond privacy to a crisis of deskilling: their daily accumulated professional experience and problem-solving thought processes are being efficiently converted into the company’s proprietary AI capabilities through continuous recording and monitoring. In this process, the role of the laborer shifts — they must not only use tools to create value, but also unwittingly become training material that feeds a potential replacement system. This real tension of using today’s labor to accelerate the obsolescence of tomorrow’s self has bred a defensive mindset among employees when facing surveillance, and has put the foundation of trust between labor and management under unprecedented strain.
To understand the rationale behind the MCI project, one must begin with the iterative trajectory of AI training data. The first generation of nourishment for large language models came from static text on the public internet, such as Wikipedia, public forums, and open-source code. After those public web resources were nearly exhausted, the industry stepped into the second generation: using AI-generated synthetic data for self-training. MCI represents a third, more ambitious direction: directly converting internal corporate work traces into training material.
According to The Guardian, Zuckerberg internally remarked that AI models should learn skills by observing how smart people do things. This line of thinking does hold water from an engineering standpoint. If the goal is for AI to troubleshoot system vulnerabilities like a top-tier engineer, or to write specification documents like an excellent product manager, then directly recording their actual problem-solving trajectories is clearly far more effective than merely extracting final documents. The problem, however, often lies not in the grand vision, but in the concrete execution plan.
First, the right to informed consent existed in name only. The essence of the MCI project was to convert daily work traces into a training dataset, not to conduct routine workplace compliance audits. These two application scenarios carry fundamentally different risks regarding data retention, reuse, and leakage. Yet Meta did not provide employees with any option to refuse or opt out of the project. This is not a minor oversight in details, but a case of the system’s architectural design phase shutting the door on autonomous choice entirely.
Second, data collection boundaries were excessively expanded. While metrics such as keystroke frequency and mouse trajectories do not contain much private information, continuous screen transcription indiscriminately captured employees’ private venting on Slack, performance review conversations, and even personal bank account pages. MCI clearly did not follow the principle of data minimization; the capture scope was drawn too broadly while security filtering and sensitive information classification measures lagged far behind.
Third, internal access controls completely collapsed. This was also the fatal blow that ultimately broke the project. Those deeply sensitive work behavior records, which had undergone neither de-identification nor scope filtering, found themselves in a near-public state on the company intranet. This incident was not the result of an external hacker attack, nor an accidental slip by a senior administrator — rather, the system itself suffered from a critical deficiency in intranet security policy, with no basic perimeter defenses established.
Although the MCI project’s suspension was directly triggered by employee resistance and a security permission incident, the governance challenges it reveals are by no means unique to Meta. When tech giants are no longer content with the vast public web and turn their gaze to behavioral traces within corporate firewalls, the computers in employees’ hands undergo a transformation in nature — from production tools to data mines. This profound shift is bound to play out in every enterprise that seeks to deeply integrate AI; Meta merely stepped on the landmine first.
In the current climate of public opinion and policy, what the public, academia, and regulators repeatedly discuss largely revolves around individual privacy protection, copyright ownership, or the safety boundaries of large models. Yet when attention turns back inside the corporate firewall — under what conditions work records can be converted into AI training material, and what ethical and compliance procedures should be followed — the industry still lacks established standards. The gap in rules between harvesting public web text and monitoring corporate internal behavior is an undeniable fact.
Looking back at MCI’s practice, what truly made the project untenable was not a court ruling, but the collapse of internal corporate trust. When employees were shocked to discover that their computer screens were effectively open to a large number of colleagues, the project’s foundation had already crumbled, no matter how compelling its technical vision might be. This serves as a warning to all enterprises planning to mine internal work data: before external laws and regulations catch up, the three barriers of employees’ data rights, the right to informed choice, and access segregation will inevitably become the hard constraints that internal training-data projects collide with first.
Going forward, any enterprise that attempts to empower internal knowledge circulation through AI will have to confront similar governance questions: How to provide clear disclosure before data collection? How to implement segregation and minimization? Who controls access rights? When data usage changes, should employee authorization be sought again? How to introduce a credible audit mechanism? Meta’s pressing of the pause button on MCI marks the first clear footprint left by the entire industry as it explores these boundaries.