Meta Deploys Behavior Tracking Software on Employee Computers to Build and Train Automated AI Agents
According to reports from several English tech media outlets citing internal memos, Meta is deploying behavior tracking software on employee computers to record mouse movements, click paths, and operational processes for building and training automated AI agents. This data is used to reconstruct the decision-making paths and operational logic of employees in their actual work.
Related reports indicate that this type of data will be used to optimize AI's autonomous execution capabilities in scenarios such as code development, content review, and the use of internal tools. Essentially, Meta is transforming "human operation trajectories" into learnable datasets to shorten the gap between understanding instructions and executing tasks.
Similar approaches have been seen in the industry. Institutions like OpenAI and Google DeepMind have previously enhanced model capabilities through human feedback reinforcement learning (RLHF) and operational demonstration data, but the practice of directly collecting "real-time workflow behavior data" signifies a further refinement of data granularity to the micro-operation level.
Source: Public Information
ABAB AI Insight
This type of data collection is not primarily about monitoring employee efficiency, but rather about acquiring "tacit knowledge." Traditional training data mainly consists of text, code, or labeled results, while what truly determines productivity are the sequences of operations, judgment paths, and methods for handling exceptions—elements that have long been difficult to record in a structured manner. Meta's approach essentially extracts "expert behavior functions," allowing AI to learn not just the outcomes, but the processes.
This reflects a shift in AI competition from "model capabilities" to "data forms." The marginal returns of parameter scale and general corpora are declining, while high-quality, strongly structured human behavior data is becoming a new scarce resource. Such data is characterized by being non-public, non-scrapable, and strongly context-dependent, meaning that platforms mastering internal workflows will have new training barriers.
From an industrial structure perspective, this trend will promote the "datafication of work processes." Knowledge-based positions in software development, design, and operations are having their operational behaviors deconstructed into learnable units. This not only changes the boundaries of AI capabilities but also reshapes the definition of labor value: value is no longer solely reflected in final outputs but in replicable decision paths.
Furthermore, this represents a reconstruction of the "human-machine relationship" within companies. Employees are no longer just executors but also providers of training data. In the long run, this may form a new production function: humans are responsible for exploring and generating new paths, while AI is responsible for scaling replication. Once paths are fully learned, the marginal value of humans in that segment will quickly decline, pushing job structures towards greater uncertainty and creativity.