AI and the hidden wage theft of automation training

Reuters recently reported that Meta plans to install tracking software on U.S.-based employees’ computers to capture mouse movements, clicks, and keystrokes for AI training. Meta says the data will not be used for performance evaluation and will include safeguards. Most revealingly, employees would help train these systems by doing their ordinary work.

One of the most degrading workplace experiences is being required to train the people who may replace you. This is familiar in outsourcing, restructuring, and redundancy processes: the employer asks workers to transfer knowledge, document routines, explain exceptions, and hand over the tacit understanding that makes a job possible.

Employers have always learned from how workers do their jobs. New workers learn by watching experienced colleagues; managers redesign work protocols from existing practice; automation has long depended on studying human tasks. AI changes the scale, visibility, and quality of that extraction. Replacement training can now be absorbed into keystrokes, screen activity, call transcripts, chat logs, emails, and the many traces workers produce while doing their jobs. The daily performance of the job can itself become the handover, without the need to sit with a successor.

The worker produces the output for which they are paid: code, edited text, customer interactions, medical notes, or administrative files. At the same time, that work generates material that helps an employer build systems capable of performing, monitoring, deskilling, or replacing it. As TechCrunch put it, Meta had found a new source of training data for its AI models: its own employees.

Other examples point in the same direction. The Guardian recently covered the story of an academic editor who was asked to correct the work of new “assistant editors”, only later discovering that these assistants were an AI system. After the AI was introduced, the editor’s fee was reduced, even though she said that correcting the AI’s errors could take longer than editing from scratch. Skilled editing became the correction of machine output, at lower pay and with less control.

This helps clarify the difference between training people and training machines. The workers did not appear opposed to helping or training other humans. The problem arose when their work was used to train AI systems, especially where this felt like automating or degrading their work. Training another worker is relational. Hidden capture for AI training makes the worker much more of a tool in the training process.

The Communications Workers of America has documented similar concerns among U.S.-based AI data workers: 52 per cent of surveyed workers said they were training AI to replace other workers’ jobs, and 36 per cent said they were training AI to replace their own jobs.

These examples point to a broader transformation. Workers already produce records, logs, messages, documents, metrics, and corrections as part of modern employment. Employers have also long learned from monitoring workers’ activities. But no human manager can realistically follow every worker around every minute of every day. Digital systems enable continuous observation while making the process largely inscrutable to workers. One’s computing alter ego can be trained in the background, with limited ability to know, contest, or shape what is being extracted.

Current employment law offers only limited ways to resist this development in the United States and Canada.

In Canada, constructive dismissal may be the most promising tool. It recognizes that an employee can be treated as dismissed where the employer unilaterally imposes a substantial change to the employment relationship. If the claim succeeds, the employee may be entitled to reasonable notice, pay in lieu, and sometimes severance or damages.

This provides an opening, because compulsory participation in automation training can change the nature of a job. It also shows the doctrine’s limits. In many AI-training situations, the employer can say that formal conditions have not changed: the same title, salary, hours, and nominal tasks. The more important change is harder to see: the employee is producing today’s output while also generating information that may be used to redesign, fragment, automate, or replace that labor.

In the United States, the weakness is sharper because of at-will employment. For an individual worker, refusing to participate in AI-training surveillance is risky. Unless the refusal connects to an established legal protection, the employer may present it as insubordination or non-cooperation.

The law often assumes that employers may organize work and require co-operation with training, transition, and business continuity. Sharing knowledge with a colleague differs from systematically capturing a worker’s actions, decisions, and corrections for a system that may later reduce the need for that worker’s labor. Compulsory automation training should be recognized as a substantial change when it adds a new, unremunerated function: producing the material that enables the job to be automated or deskilled.

With automated capture, the employer can obtain both today’s work and tomorrow’s training material. The working day and wage do not expand, but the employer’s productive yield from the worker’s time multiplies. In practice, this is stealth wage theft: a hidden double shift within the same working hour, paid as one job while producing the material that may automate tomorrow’s job.

There is also a dignity dimension. Many workers are not hired to train AI systems. Their expertise is situated and often tacit: judgment, handling exceptions, communicating with others, and navigating institutional contexts. Capturing that expertise through surveillance and repurposing it for automation treats workers as instruments for extracting their own know-how, especially when they have no meaningful say over these systems.

Legislation and collective agreements should respond directly. Employers should not have an unrestricted managerial prerogative to require workers to train systems intended or likely to automate, deskill, or replace their work. Data collected for monitoring, quality control, security, or ordinary administration should not be repurposed for AI training without robust safeguards and worker power over the decision. Where workers assist automation, that contribution should be recognized as work and compensated accordingly.

Collective bargaining is especially important because these decisions are rarely isolated individual choices. Capturing work processes for AI training changes the organization of work, the allocation of skills, and the distribution of power within the enterprise. It should trigger duties to bargain before the system is introduced and while it is being developed. Individual refusal rights matter but cannot substitute for collective power where the employer can redesign work at scale.

Ultimately, this is about the boundaries of managerial power. Employers have always claimed broad authority to organize production and introduce new technologies. AI extends that authority by making the labor process more observable, extractable, and reproducible.

This transformation should not be treated as a routine incident of work. When a worker is hired to perform a task, the law should not presume that they have also agreed, for the same wage, to produce the data infrastructure through which that task may be automated. It requires legal limits, collective bargaining, and a renewed understanding of doctrines such as constructive dismissal. Training a replacement has been folded into the ordinary performance of the job, where it is harder to see, harder to refuse, and easier to deny. This demands urgent action at the bargaining table, in courts, and in legislatures.