Meta Tracked Keystrokes, Clicks And Screens: Did Its AI Training Project Cross The Line?

Imagine logging into your office laptop, replying to emails, switching between spreadsheets, copying text from one window to another and using keyboard shortcuts to finish a task faster. Now imagine every click, keystroke and mouse movement becoming part of a massive dataset designed to train artificial intelligence.

This scenario reportedly took place inside Meta, which had announced in April its ambitious Model Capability Initiative (MCI), an AI training programme intended to teach AI agents how humans work. But the Mark Zuckerberg’s company is now pausing the programme after facing backlash from its employees over privacy.

A report by Business Insider showed that tens of thousands of tables containing employee activity, their private conversations,

performance data and transcriptions were accessible across the company. This incident was classified as a SEV 2 on a scale of 0 to 5, with 0 being the most severe.

The episode has increased concerns over workplace surveillance, privacy and the growing role of employee data in building the next generation of AI.

As businesses worldwide race to develop AI assistants that can perform office tasks independently, the question is becoming increasingly urgent: How much employee data is too much?

What Is Meta’s Model Capability Initiative?

Launched in April 2026, the MCI was a mandatory AI training programme for staff. Unlike traditional AI models that learn from publicly available text and images, the programme focuses on observing how employees interact with software during their daily work. It records digital behaviour such as keystrokes, mouse movements, clicks, scrolling patterns and on-screen actions on company-issued devices.

The objective is to help AI systems understand how people complete everyday office tasks.

For example, instead of simply knowing what a spreadsheet is, an AI agent could learn how an employee opens multiple applications, copies data between documents, uses keyboard shortcuts, searches internal tools and navigates complex workflows. The long-term goal is to build AI assistants capable of carrying out many of these tasks autonomously.

As AI companies compete to create digital agents that can function like human co-workers, behavioural data has become an increasingly valuable resource.

Why Did Meta Pause The Programme?

The programme came under scrutiny after an internal security lapse reportedly exposed a massive volume of employee activity data.

More than 45,000 tables of internal information were allegedly made accessible to workers across the company. The exposed material reportedly included private workplace conversations, performance assessments, AI prompts and transcripts that employees believed would remain restricted.

Although there is no indication that the information was accessed maliciously or leaked publicly, the incident immediately raised questions about how such sensitive data was being collected, stored and protected.

Meta subsequently paused the MCI while reviewing its privacy safeguards and internal security measures.

The frustration among employees inside Meta has also grown after the company cut thousands of jobs, focusing more on AI projects. The company will spend $135 billion this year on AI programmes, which is roughly equal to the amount Meta has spent on AI in the previous three years combined, a BBC report said.

Why Employees Are Calling It Workplace Surveillance

The backlash did not begin with the data exposure. Many employees had already expressed concerns that the programme blurred the line between AI research and constant workplace monitoring.

In an initial response to worker frustration, which was displayed in part through a petition signed by nearly 2,000 Meta workers demanding that the programme be cancelled, Meta said it would allow workers to not be tracked for up to 30 minutes at a time, the BBC report mentioned.

Their concerns extend beyond simple productivity tracking.

Traditional workplace monitoring generally focuses on attendance, login times or software usage. But the Model Capability Initiative reportedly captures far more detailed behavioural patterns, including how employees navigate applications, communicate with colleagues and interact with AI tools throughout the workday.

Critics argue that such comprehensive monitoring transforms ordinary work into a continuous source of AI training data, often without employees fully understanding the extent of information being collected.

The recent data exposure has strengthened those concerns by demonstrating how even internal systems can become vulnerable.

Why Companies Want This Data

Early AI models relied primarily on internet content, books and publicly available datasets. The next generation of AI agents, however, aims to perform real-world tasks rather than simply answer questions.

To achieve that, developers need examples of human behaviour inside digital workplaces. Every click, drag-and-drop action, keyboard shortcut or sequence of software interactions teaches AI systems how people solve problems, switch between applications and complete assignments efficiently.

The richer the behavioural data, the more capable these AI assistants become. For technology companies investing billions of dollars in enterprise AI, employee workflows represent one of the most valuable training resources available.

Microsoft gained advantage over its competitors when it invested in OpenAI in 2019, giving it a direct pipeline to the latest AI research. Microsoft quickly embedded AI into its ubiquitous productivity apps. Microsoft Copilot agents are now an integral part of the daily routine of enterprise end-users.

Google’s Vertex AI is a leader in Gartner’s Magic Quadrant for AI application development platforms and conversational AI platforms. Google has successfully integrated Gemini into its broad portfolio that includes Google Search, YouTube, and Gmail. And enterprises are increasingly turning to Google Cloud as the place to develop and run AI workloads

Could Indian Companies Face Similar Questions?

India is home to one of the world’s largest technology workforces and a rapidly expanding ecosystem of Global Capability Centres (GCCs).

Indian IT firms and multinational companies are aggressively adopting generative AI across software development, customer support, finance, human resources and enterprise operations.

According to ADP Research’s People at Work 2026 report, one in five workers say they use AI almost daily, highlighting how quickly the technology has evolved into a routine workplace tool.

India stands out as a global leader in this shift. A striking 41% of employees in India report using AI nearly every day, and 4 out of 5 (80%) employees use it at least multiple times a week – the highest levels recorded across all surveyed markets.

Many organisations already track AI adoption metrics and employee interactions to improve workflows and evaluate productivity.

Companies utilize internal dashboards — similar to the models employed by consulting firms like KPMG — to monitor how frequently employees interact with tools like Microsoft Copilot, ChatGPT, and custom LLMs.

As businesses build increasingly sophisticated AI assistants, they may also seek larger datasets based on real employee behaviour. That creates a new challenge for employers and policymakers alike.

Unlike traditional workplace monitoring, AI training can require collecting detailed behavioural information that reveals not just what employees do, but how they think, navigate software and solve problems.

The Meta controversy highlights the need for clearer safeguards around consent, data collection, privacy and security before such practices become commonplace.