Workers rights in the age of algorithmic management

Cameras, wearable sensors, voice-analysis software that claims to read fear, boredom, or distraction — all of it, treated as a category of AI too dangerous to deploy against employees, with narrow exceptions reserved for medical and safety contexts.

Twelve days later, on February 14, the Acting General Counsel of the National Labor Relations Board issued Memorandum GC 25-05, rescinding the framework his predecessor had built to challenge precisely this kind of surveillance. Memorandum GC 23-02 had proposed treating intrusive electronic monitoring and algorithmic management as presumptive violations of the National Labor Relations Act. That presumption is now gone.

The distance between those two dates is the distance between two political economies. One is choosing to govern the technology. The other is choosing to be governed by it.

The Great Regulatory Divergence

We are watching the world's two largest regulatory blocs move in opposite directions at speed, and the material consequences for workers are already accumulating.

In Brussels, the legislative machinery is tightening. The Platform Work Directive, approved in April 2024, must be transposed into national law across the European Union by December 2026. The AI Act, finalized in 2024, layers a separate but interlocking set of obligations on the same employers. From August 2, 2026, every AI system deployed for recruitment, task allocation, performance evaluation, or access to self-employment will be legally classified as "high-risk" — triggering mandatory risk assessments, audit logs, transparency disclosures to workers, and meaningful human oversight before any consequential decision is taken.

In Washington, the machinery is loosening. The NLRB's withdrawal from the position articulated in GC 23-02 is not a technical adjustment. It is the abandonment of an attempt to read the existing National Labor Relations Act as a tool capable of constraining algorithmic control. That attempt — modest, internally contested, and never a final agency rule — was the closest the United States had come to a federal position that surveillance-driven management could be unlawful.

The asymmetry is structural. The EU is constructing a perimeter: certain practices are prohibited outright, others are permitted only under tightly controlled conditions, and the burden of compliance sits with the deployer. The US is dissolving the perimeter: the question of whether algorithmic management is illegal now returns to the slow, case-by-case adjudication of unfair labor practices, with no presumptive protection for the worker and no doctrinal hook for prosecutors.

For workers, the practical effect is that the same piece of software — the dispatch algorithm, the productivity score, the keystroke logger — can be illegal in Madrid and entirely unregulated in Memphis. What one jurisdiction requires as basic workplace governance, the other treats as optional. The worker, in both, is the same.

The EU AI Act: What "High-Risk" Actually Means

The temptation is to treat the AI Act as a piece of consumer-protection theater — a transparency layer for cameras, a disclosure regime for chatbots. It is not. For workers, it is the most consequential employment regulation of the decade, because it redefines who carries the burden of proof when an algorithm makes a decision that affects a livelihood.

Recruitment and selection tools. Task-allocation systems in warehouses and logistics. Performance evaluation engines that score workers against metrics only the software can see. Access-to-work platforms that gatekeep self-employment. Each of these is classified as a high-risk AI system under the Act. That classification triggers four concrete obligations: documented risk assessments before deployment, persistent logging of system activity for audit purposes, effective human oversight built into the design and operation of the system, and transparency disclosures to affected workers about how the system functions and what data it consumes.

The penalties are calibrated to the balance sheets of the firms most likely to deploy this technology. Fines for prohibited practices — including the emotion-recognition ban that took effect February 2, 2025 — can reach €35 million or 7% of annual global turnover. That figure is not symbolic. It is designed to make compliance cheaper than non-compliance.

The EU has made a simple structural bet: if the algorithm cannot explain itself to a human being, it cannot be allowed to fire one.

The deeper intervention is the human-oversight requirement. The Act does not merely regulate outputs; it regulates the capacity of the system to act without review. High-risk systems must be designed and deployed so that the people assigned to oversee them can properly understand how the system works, monitor it in real time, correctly interpret its outputs, and intervene — including by reversing or disregarding the system's recommendation — when the situation warrants. An employer cannot deploy an opaque model to make consequential decisions about workers and then shrug at questions about its logic. This is not a transparency garnish. It is a structural redesign of the employment relationship — a refusal to allow the firm to launder its discipline through software that no one inside the company is equipped to defend.

What "oversight" means in practice is closer to an engineering specification than a moral appeal. It means the system produces logs a human can read. It means the model's logic is documented in language a supervisor can interpret. It means a designated reviewer has the authority — and the training — to halt, modify, or reverse the system's output before a worker loses a shift, a wage, or a contract. In jurisdictions where this obligation is enforced, the question of whether a particular firing was lawful becomes an investigable question. The answer can be sought, contested, and litigated. That is the practical content of "meaningful human oversight" — not a slogan, but an evidentiary regime.

The Platform Work Directive and the End of Unaccountable Dismissal

The AI Act governs the systems. The Platform Work Directive governs the employment relationship those systems are increasingly used to administer, and it does so with two mechanisms that dismantle the legal architecture of the gig economy as we have known it.

First, a rebuttable presumption of employment: when a platform exerts control over the worker's pay, performance, scheduling, or conduct, the law presumes an employment relationship exists, and the platform must prove otherwise. Second, a prohibition on algorithmic terminations without human review — a worker cannot be deactivated, fired, or meaningfully sanctioned by an automated decision without a human assessing the case first.

The presumption of employment is the more consequential of the two. It attacks the foundational trick of the gig economy: the classification of workers as independent contractors in order to strip them of wage protections, social insurance, collective bargaining rights, and the legal standing to unionize. By reversing the burden of proof, the Directive effectively makes misclassification a cost the platform must litigate, not a protection the worker must litigate around.

For American labor organizers, this is the most instructive provision on the table anywhere in the world. The decades-long struggle over worker classification in the US has been a defensive battle, with workers forced into court to prove that they are employees. The EU model flips the script: prove they aren't. It shifts the leverage from the firm to the worker, and it does so by statute rather than by litigation.

We should be clear about the timeline. The Directive is not yet binding in every member state. National governments have until December 2026 to transpose it, and several are already pushing back against provisions that threaten the business models of platforms operating in their jurisdictions. But the framework is fixed, and the transposition process forces every member state to construct legal infrastructure capable of investigating how algorithms allocate, evaluate, and punish work. That infrastructure, once built, will not be disassembled easily.

There is a quieter implication worth naming. The Directive's review requirement does not specify which human must perform the review, or what seniority they must hold. But by demanding that an automated sanction be subject to human assessment, it forces firms to create a place in the workflow for human judgment — and with that place, a person whose reasoning can be questioned and whose decision can be challenged. The firm can no longer point at the algorithm and disclaim responsibility; the reviewer becomes answerable, even if the Directive does not impose a per-decision named-human requirement from the outset.

The US Rollback: From Presumption to Permission

The collapse of the federal position on algorithmic management is not the result of a Congressional vote or a Supreme Court ruling. It was accomplished by a single memorandum, signed on February 14, 2025.

Memorandum GC 25-05 rescinded Memorandum GC 23-02, which had proposed that the NLRB treat intrusive electronic monitoring and algorithmic management as presumptive violations of Section 8(a)(1) of the NLRA — the provision protecting workers' rights to concerted activity. Under the prior framework, an employer who deployed surveillance technology that tended to chill organizing could be presumed to have committed an unfair labor practice, shifting the burden of proof to the employer.

That framework was never a regulation. It was an enforcement theory — a legal interpretation of existing statute. But it was the most aggressive federal reading of the NLRA in the algorithmic era, and it would have given regional NLRB offices a doctrinal hook to pursue cases against Amazon, Uber, and the rest of the platform economy. Its rescission does not make algorithmic management legal. It makes algorithmic management ungoverned at the federal level — a vacuum where the states, the courts, and the unions now have to do all the work, and where the federal agency charged with protecting collective action has effectively withdrawn.

When the federal labor board withdraws from the field, the work does not disappear. It is transferred — to state legislatures, to workers themselves, and to the courts. The cost of that transfer is paid in years of litigation and in unrepresented workers disciplined by software they cannot read.

This is what rollback looks like in practice: not a prohibition, but an absence. The technology continues to spread. The legal infrastructure to challenge it retreats. The gap between deployment and redress widens, and the firms that benefit from that gap are not waiting for permission to widen it further.

There is a secondary consequence that should not be overlooked. GC 23-02 was not only an enforcement theory about surveillance; it was a signal to regional offices that algorithmic management belonged inside the NLRB's reading of the statute. Removing it changes what NLRB investigators will look for, what they will ask about, and what they will treat as cognizable harm. An investigator who previously might have asked whether a productivity score was being used to chill union activity now has no doctrinal encouragement to ask. The rescission is therefore not just the removal of a position; it is the slow erosion of the institutional muscle required to litigate the issue at all.

California and the Warehouse as Laboratory

The federal retreat has thrown the burden of resistance onto the states, and California has emerged as the most aggressive laboratory in the country.

Assembly Bill 701, effective January 1, 2022, was the first state law in the United States to regulate warehouse quotas directly. The threshold is specific: employers with 100 or more workers at a single distribution center, or 1,000 or more workers across multiple facilities, must provide written descriptions of any quota to which workers are subject. The law prohibits quotas that prevent workers from taking mandated meal or rest breaks, or from using the restroom. It also creates a private right of action: workers can sue, and the Labor Commissioner can cite.

In June 2024, the California Labor Commissioner cited Amazon nearly $6 million — specifically $5.9 million — for violations of AB 701 at two warehouse facilities in Moreno Valley and Redlands. The violation was not abstract. Amazon failed to provide the written quota disclosures that the statute requires. The case is significant for two reasons. It establishes that the quotas themselves are not illegal under AB 701 — what is illegal is the opacity. And it demonstrates that a single state agency, armed with a specific statute, can extract meaningful penalties from the most capitalized logistics company in the world.

The model is replicable, and it is being replicated. Other states have begun drafting parallel legislation, and union organizers have used AB 701 disclosure orders as a tool to surface the specific metrics that warehouse workers are expected to meet — the basis on which they are disciplined, terminated, or denied hours. Disclosure is not protection. But it is the precondition for protection: you cannot organize against a quota you cannot see, and you cannot grieve a metric you cannot name.

What AB 701 exposes, more than anything, is how thin the baseline of legal disclosure actually was in the United States. The law did not invent the right to know whether you are being timed; it declared, for the first time at the state level, that this right exists and that it is enforceable. The fact that this counts as a milestone illustrates the depth of the prior legal vacuum. The same dynamic will play out — and is already playing out — for algorithmic scheduling tools, for AI-driven hiring assessments, for productivity scores that dock pay in near-real time. California's law is a template, not a destination.

Human Oversight as a Material Condition

In December 2025, the European Parliament adopted a resolution calling on the European Commission to propose a dedicated directive on algorithmic management in the workplace. The resolution asks for explicit human oversight of all AI-supported decisions affecting workers, with protections extending to physical and mental health. The final legislative text and timeline are not yet fixed — but the political signal is unambiguous. The EU is not treating the AI Act and the Platform Work Directive as the final word. They are the foundation on which more specific obligations will be built.

The choice the EU has made is structural, not symbolic. It says that certain decisions in the employment relationship — who is hired, who is fired, who is disciplined, who is surveilled, who gets the next shift — cannot be fully delegated to systems whose reasoning is opaque to the humans whose lives they shape. This is not a question of technological capability. Modern AI systems can absolutely allocate shifts and score performance at scale. The question is whether we allow them to do so without oversight, or whether we build the legal architecture that ensures a qualified human can review, challenge, and where necessary override the system's output before it becomes a sanction against a worker.

The algorithm does not need to be sentient to be an instrument of power. It only needs to be opaque.

It is worth pausing on what an oversight regime makes possible in practice. Without it, a worker who is disciplined by software faces an evidentiary abyss: the system has flagged them, the system has a score, the system has produced a decision, but no one can explain in language the worker can challenge how the score was produced or what would have changed it. With an oversight regime, that same worker can demand to see the relevant logs, can require that a human reviewer explain the reasoning, and can contest the decision in a forum with the legal authority to overturn it. The difference is not in whether the algorithm makes decisions — both regimes permit that. The difference is in whether the decision can be reviewed, and by whom.

For the United States, the lesson is not that algorithmic management must be banned outright. It is that the absence of a federal framework is itself a policy choice — a choice to let employers write the rules in private, enforce them through software that workers cannot read and courts cannot easily interpret, and discipline anyone who objects. The complicity is not in the deployment. It is in the silence.

What we can build is concrete. A labor movement that treats algorithmic management as a collective bargaining issue — every contract negotiation a fight over what data the employer can collect, what decisions the algorithm can make unilaterally, and what human review must precede any sanction. A regulatory movement that demands disclosure as a precondition for deployment, the same way we demand disclosure for chemical exposures and safety hazards. A political movement that draws the line where the EU has drawn it: certain technologies, deployed under certain conditions, are incompatible with a free labor market and a functioning democracy.

The technology is not destiny. The legal architecture is. And right now, that architecture is being assembled — by legislators in Brussels and Sacramento, by labor boards in Washington choosing what to enforce and what to abandon, by union locals on warehouse floors demanding the numbers behind the discipline, and by workers who refuse to be managed by software that cannot explain itself, cannot be questioned, and cannot be held accountable.

The question is not whether we will be governed by algorithms. The question is whether we will be governed by them alone. The EU has answered that question in statute. The United States has answered it in silence. We owe ourselves a louder answer than that.