In June 2019, union members at IBM Japan opened their summer-bonus assessments and found a number attached to their labor that they had not chosen and could not contest. Their individual performance rate averaged 63.6 percent. Everyone else's averaged 100. Many union members were scored at zero. The software that produced those numbers ran on IBM's own Watson, was marketed as a "compensation advisor," and stands as one of the earliest documented cases of an algorithm setting a human being's pay in an ordinary job — not a gig, not a side hustle, a job with a contract and a union and a break room (Dubal, Columbia Law Review, 2023). Seven years later, that experiment is no longer an anomaly. It is the operating system.
The First Thing AI Did to Workers Was Not Fire Them
We were promised the robots would take the jobs. The more interesting — and better documented — story is that they took the raises instead. A 2026 whitepaper analyzing the Anthropic Economic Index against Bureau of Labor Statistics wage data across 321 occupations found that workers in the most AI-exposed jobs saw real wage growth fall roughly 6.7 percentage points behind their low-exposure peers after 2023, with no detectable effect on employment ("The Impact of AI on the U.S. Labor Market," 2026). Read that twice. Nobody got laid off. The productivity gains were real. They simply did not arrive in anyone's paycheck. Economists have a tidy word for this — wage compression — and it sorts itself cruelly by income: the bottom wage quartile fell 10.7 percent behind, and service workers fell roughly a quarter behind everyone else. A worker's AI exposure, it turns out, predicts not that a machine will replace them, but that their employer will keep the surplus their new efficiency produced. The people with the least were charged the most for the privilege of being made more productive.
Surveillance Pay Has a Name Now
The legal scholar Veena Dubal gave the practice its name. Algorithmic wage discrimination, she writes, pays "individual workers ... different hourly wages — calculated with ever-changing formulas using granular data on location, individual behavior, demand, supply, or other factors — for broadly similar work" (Dubal, 2023). The Washington Center for Equitable Growth calls the workplace version surveillance pay, and in August 2025 it published the first systematic audit of the vendors selling it: 500 AI labor-management products, 20 at high risk of setting wages this way, and 16 of those wired directly into payroll or HR systems (Monroe et al., 2025). This is algorithmic management — the use of automated systems to monitor, evaluate, and direct workers — graduating from the rideshare dashboard into the hospital ward, the warehouse aisle, and the call-center headset. Most of the audited vendors, the researchers found, give workers no visibility at all into the data or the logic behind their own pay.
The Watching Is the Point
How closely are American workers watched? Closely enough that surveillance is now the default setting. Algorithmic monitoring tools have a 90 percent adoption rate among U.S. firms; 72 percent use software to track the speed of work, 88 percent to track working time, and 55 percent monitor the content and tone of conversations, voice calls, or emails (OECD, 2025). Sit with that last figure. Fifty-five percent in the United States, against 6 percent in Europe and 8 percent in Japan. This is not a story about what technology makes possible. The same tools exist everywhere. It is a story about what one particular labor market has decided it is permitted to do.
Objectivity Is the Alibi
The vendors have an answer, and it sounds reasonable. In the OECD's survey, 60 percent of managers said the tools improved their own decision-making, and more reported rising than falling job satisfaction, largely from shedding repetitive drudgery (OECD, 2025). Data-driven pay, the pitch goes, is consistent, objective, and free of human prejudice. But American managers are far likelier than European or Japanese counterparts to believe the software reduces bias — a belief the European Parliament's own researchers dispute, noting that "seemingly objective" algorithms tend to reinforce the bias already baked into them (EPRS, 2025). And objectivity for the firm reads as opacity for the worker.
An objective number you are forbidden to inspect is not objectivity. It is authority wearing a lab coat.
Where This Goes Without a Fight
The trajectory is not hypothetical. In the EU, workers' exposure to algorithmic management is projected to climb from 42.3 percent to 55.5 percent in the medium term, and 27 percent of managers already report inadequate protection of workers' physical and mental wellbeing where these tools run (EPRS, 2025). The Equitable Growth researchers are blunt about the stakes: without policy intervention, they "fear that these practices will become normalized, thus growing income uncertainty, entrenching bias, and eroding wage-setting transparency" (Monroe et al., 2025). Normalization is the entire game. A wage cut you can see is a grievance you can organize around. A wage cut delivered as a personalized formula, recalculated by the hour from data you generated simply by doing your job, is just Tuesday.
The Question You Now Have to Carry
Here is the discomfort I cannot spare you. Every keystroke, every logged minute, every measured pause that an algorithm converts into a slightly lower number was produced by the worker's own hands. You are, in the most literal sense, manufacturing the evidence used to pay you less. The estimated cost of this arrangement is already a conservative $28 billion a year, borne by the 5.8 million Americans in high-exposure occupations ("The Impact of AI on the U.S. Labor Market," 2026). The mirror this technology holds up does not show a future of leisure. It shows us who we were willing to watch, and how cheaply we were willing to price them. The tools are neutral; the people who bought them, aimed them, and quietly kept the difference are not. So the question is not whether the machine is fair. The question is who gets to read the formula — and so far, it isn't you.
