EU AI transparency rules leave workers unable to challenge algorithmic decisions

Roughly four out of five companies in Europe's four largest economies now rely on algorithmic tools to assign shifts, track performance, set targets or inform hiring and disciplinary decisions. The figure, drawn from an OECD survey, makes clear that automated management is no longer an experiment confined to gig-economy platforms. It is routine. Yet the rules Europe is putting in place to govern this shift address only part of the problem: they tell workers when a machine is involved. They do not give those workers any reliable means of contesting what the machine has decided.

What the AI Act actually requires

The EU AI Act, which entered into force in August 2024 with a phased implementation timeline running to 2027, classifies certain AI systems by risk level. Employment-related applications, including those used for recruitment, task allocation, promotion and termination, fall into the high-risk category. That classification brings obligations: deployers must ensure human oversight, maintain documentation and inform workers when an AI system contributes to a decision affecting them.

Transparency, in other words, is the regulatory baseline. Employers must disclose that AI is being used. What they are not required to provide is a structured mechanism through which an affected employee can identify the specific inputs that shaped a decision, challenge an error, or obtain a genuinely independent review. The disclosure obligation is real. The accountability gap is equally real.

Why knowing is not the same as being able to act

Consider the practical difference. A warehouse worker in Lombardy who learns that an algorithm helped determine her shift pattern has gained something, but not much. She now knows a machine was involved. She does not necessarily know which data points the system weighed, whether it made a recommendation or effectively determined the outcome, whether a human reviewed the result, or who that human was. She certainly has no standardised route to challenge an error without legal expertise, and she may reasonably fear retaliation for raising the issue with the same manager who relied on the tool in the first place.

Gleb Tsipursky, a behavioural scientist and author of a forthcoming book on AI adoption in the workplace, puts the point directly: telling an employee that AI was involved does little if that employee must still guess how to raise an error, fears retaliation for objecting, or waits weeks while the disputed decision remains in force. Disclosure without a practical right to challenge, he argues, is little more than a notice pinned to a locked door.

The Deliveroo ruling and its wider significance

Europe already has a concrete illustration of what happens when algorithmic management operates without adequate safeguards. In 2020, the Bologna Labour Court examined Deliveroo's rider-ranking system, which allocated access to the most desirable delivery slots based on a scoring model. Riders who cancelled booked sessions too late saw their rankings fall, reducing their future earning capacity. The court found the system discriminatory because it penalised cancellations regardless of reason, including absences connected to illness or strike action, both of which enjoy legal protection under Italian labour law.

The ruling mattered beyond one platform. It demonstrated that automated decision-making systems can embed discriminatory logic that is difficult for individual workers to detect or prove, particularly when the scoring criteria are opaque. It also showed that existing labour protections can be undermined not by employer intent but by system design, and that remedying the problem required a court order rather than any internal appeal mechanism Deliveroo had in place.

Three guarantees that transparency alone cannot deliver

The argument for going beyond disclosure rests on three interconnected guarantees. The first is meaningful transparency: not a generic sentence in a privacy policy, but plain-language information about what role the system played, whether it recommended or determined the outcome, and which categories of data it processed. Workers need to understand the mechanism, not merely acknowledge its existence.

The second is named human responsibility. Every AI-assisted employment process should have an identifiable person with the authority to examine the evidence, alter the outcome and, if a pattern of errors emerges, suspend the system. The phrase 'the algorithm decided' cannot function as a full stop. Someone must be answerable, and that someone must have the power to intervene.

The third is effective remedy. An appeal right that exposes a worker to income loss during a protracted review, or that routes complaints back through the same manager who deferred to the system, offers protection in name only. A meaningful process would shield the employee from serious harm while review takes place, keep a record of corrections so that recurring failures become visible, and involve works councils or unions in system design before problems become entrenched.

The cross-border inconsistency problem

The European Commission has emphasised that AI transparency rules should function consistently across the single market. Workplace appeal rights currently enjoy no such common standard. Protection depends on national enforcement capacity, the strength of existing labour institutions, and the willingness of individual employers to go beyond minimum legal requirements. A delivery driver in Amsterdam, a bank clerk in Frankfurt and a warehouse operative in Madrid may face similar algorithmic decisions but have vastly different options for challenging them.

This variation is not merely an equity concern. It creates distortions within the single market. Employers operating across borders face different obligations depending on where their workers are located, while workers performing similar tasks receive different levels of protection based on geography rather than the nature of the risk.

Employers might also benefit from appeal mechanisms

The case for stronger accountability is not solely about worker protection. Managers often receive AI-generated recommendations without sufficient context to judge their quality. A formal review process would generate better evidence about how systems perform in practice, expose weak or biased inputs, and give managers a legitimate basis for rejecting outputs rather than deferring to a tool that appears objective. Consultation with employees and their representatives can also improve adoption rates: workers are more likely to engage constructively with systems they understand and can question.

The current debate is frequently framed as a tension between innovation and regulation. That framing is too narrow. The relevant question is how the gains and costs of automation are distributed. A scheduling system that saves managers hours can simultaneously impose unpredictable working patterns on staff. A performance model that standardises evaluation can conceal a poor proxy for actual quality. A recruitment tool that accelerates screening can shift the burden of proving an error onto applicants who never learn why they were rejected. These are not technical problems. They are questions of institutional power.

A common floor, not a single procedure

Nobody is arguing for a single rigid procedure applied identically to a hospital, a warehouse, a bank and a delivery platform. The risks differ too much. What is being proposed is a common minimum: notice that explains the system's role in plain terms, a named human accountable for each decision, an accessible challenge mechanism, protection from harm during review, a record of corrections, and worker participation in system design and review. These six elements would apply regardless of sector or member state.

Without such a floor, transparency remains a label. With it, disclosure becomes a doorway: the starting point for a process that can actually correct errors, assign responsibility and prevent harm. The European Commission has committed to consistent transparency standards. Whether it will extend the same ambition to accountability mechanisms remains an open question, and one that will shape the lived reality of automated management for millions of European workers.