What this page is — and is not
The displacement question is usually asked as a count: how many jobs does AI take? That matters, and Labor and Economic Displacement covers it. This page asks a different question: what happens to the jobs that remain? A worker can keep a paycheck while losing control of the schedule, the pace, the judgment calls, and the dignity that made the work worth doing.
The mechanism described here — algorithmic management — is observed, documented, and real. The projection of today’s gig-economy and warehouse findings onto the whole economy is an argument, not a measurement. This page labels which is which.
The mechanism: algorithmic management
Kellogg, Valentine, and Christin’s widely cited review defines algorithmic management as software performing managerial functions: assigning tasks, monitoring performance, evaluating output, and disciplining or rewarding workers (Kellogg, Valentine & Christin, “Algorithms at Work: The New Contested Terrain of Control,” Academy of Management Annals, 2020). They organize the toolkit into six mechanisms — recommending, recording, rating, replacing, rewarding, and disciplining — and note that the contested terrain is not new. Employers have always tried to extract more effort for less pay; what changed is the granularity, speed, and asymmetry of the system doing it.
In practice this means: schedules set by forecast demand rather than a manager who knows you have a child in school; productivity scores computed from keystrokes, scans, or seconds-per-task; customer star ratings treated as performance reviews; and termination triggered automatically when a metric drops below a threshold, sometimes with no human in the loop at all. The worker experiences this as a boss that is always watching, never negotiable, and impossible to appeal to, because there is no one there.
The direction of travel
As AI absorbs routine cognitive work, the jobs left to humans sort into two buckets. The first is work machines cannot yet do: skilled trades in unpredictable physical environments, complex human negotiation, genuine creative judgment. The second is work humans must be kept for — because a law, a liability rule, or customer preference requires a person somewhere in the chain, even when the system can do almost everything else.
It is the second bucket where quality degrades fastest. If a person is present only to satisfy a requirement, the employer’s incentive is to make that presence as cheap and controlled as possible: scripted discretion, monitored compliance, no slack. The worker becomes a human seal of approval stamped on a machine’s output — responsible for the outcome, stripped of the authority to shape it. A nurse required to “validate” AI triage in twelve seconds per patient is not augmented; she is a liability buffer with a stethoscope.
What the evidence shows — and what it doesn’t
The strongest evidence comes from settings where algorithmic management is already total. In gig platforms, researchers documented how ratings systems and opaque deactivation rules shift risk onto workers while denying them information about who decides and why (Rosenblat & Stark, 2016). In warehouses, monitoring granularity down to seconds-per-task is associated with elevated injury rates and attrition in journalistic and advocacy reporting, though peer-reviewed causal estimates remain thin. The International Labour Organization’s global analysis of generative AI is explicit that exposure to automation is not the same as job loss, and that quality dimensions — task intensity, autonomy, and opportunity for skill use — move separately from headcounts (ILO, 2023).
Honesty requires the boundary: these studies measure specific, heavily monitored workplaces. Extending them economy-wide is an argument by analogy, not a measurement. Some remaining jobs will get better; some monitoring will fail or backfire; some workers prefer app-mediated flexibility to the alternative, which is often nothing. The defensible claim is narrower and still serious: wherever humans are kept in the loop against an employer’s preference, the default economic pressure runs toward control, not autonomy.
Three signals indicate which way a workplace is moving: whether workers can see and contest the data used to evaluate them, whether any human with authority can override the system, and whether productivity targets ratchet automatically as the software learns what the best-performed shift looked like. The third is the tell. A system that keeps raising the bar as workers adapt is not a tool for organizing work; it is an extraction ratchet, and its ceiling is set by physiology, not by judgment.
The complementarity question
Whether AI augments a remaining role or degrades it is not decided by the technology. It is decided by who chooses the design. The same triage model that lets a nurse spend her hour with the patients who need her can instead be used to set her quota. The same coding assistant that handles boilerplate so a junior developer can learn architecture can instead be used to justify supervising four juniors instead of hiring two seniors.
| Design choice | Augmenting outcome | Degrading outcome |
|---|---|---|
| Who sets pace | Worker controls sequence | System sets quota, worker keeps up |
| Who sees the data | Worker gets performance feedback | Employer surveils, worker can’t inspect |
| Who owns judgment | AI advises, human decides with reasons | Human rubber-stamps, liability stays human |
| What happens on error | Review and retraining | Automatic downgrade or termination |
The determinants are institutional, not technical: bargaining power, enforceable standards, whether the worker can inspect and contest the data, and whether anyone above the algorithm has both the authority and the incentive to override it.
One asymmetry deserves emphasis. When augmentation is promised, the promised beneficiary is the worker; when degradation occurs, the beneficiary is the cost line. That mismatch means the burden of proof should sit with the deploying organization: if a system genuinely augments, its designers can show preserved discretion, inspectable data, and stable targets. “The human is always in the loop” is not evidence of augmentation when the human’s only permitted action is to continue.
Policy and bargaining responses
Job quality is a labor-standards question, not merely a transition-assistance question. Retraining programs do nothing for a worker whose new job is algorithmically paced, rated, and terminated. The relevant instruments are older and blunter: transparency rights over the data used to evaluate you, notice and human review before adverse action, limits on off-duty monitoring, and sectoral bargaining that sets floors on scheduling and pace. The ILO’s conventions on working conditions and occupational safety provide the established backdrop — the framework for treating intensity and surveillance as standards issues, not perks to negotiate individually (ILO, safety and health at work).
None of this requires treating AI as uniquely evil. It requires refusing the framing that the only choice is between job counts and nothing. Workers facing degraded quality face the same question as workers facing displacement in Distributional Economics and Power Concentration: whether gains flow to the people the system watches, or only to the people who own it. Individuals can hedge at the household level through Career and Income Diversification; the standards themselves are a collective fight, and worth having before the second bucket of jobs becomes the norm.