When researchers conducted an outcome audit of Twit­ter’s photo-cropping algorithm in 2021(Li et al. 2023) they discovered that it consistently favoured white fac­es in previews. Twitter responded by retiring the algo­rithm altogether. That’s an example of a successful outcome audit leading to reform. However, audits face real limits. Complex AI models such as GPT or Goog­le’s ranking systems are too vast, continuously updated and partly data-driven to be fully»opened up« or repli­cated. Even when companies release transparency re­ports, they often disclose only high-level summaries. The result is what scholars call »transparency thea­tre« (Cellard 2024), which merely gestures towards ac ­countability without enabling independent verification. To avoid such»transparency theatre«, these rights must be coupled with substantive review powers . In the workplace, mechanisms should be put in place to en­sure meaningful human review of decisions taken or supported by AI systems. Affected individuals should receive clear explanations and be able to request re­consideration by a competent human authority. Where violations occur, procedures must allow for rectification and, where appropriate, modification or suspension of the system to prevent recurring harm. One emerging practice is participatory auditing (Costanza-Chock, Raji and Buolamwini 2022), in which affected communities – workers, unions, con­sumers – are involved in testing and reviewing sys­tems. This expands the notion of expertise beyond data scientists, recognising that people impacted by automated decisions have contextual knowledge about harms and unintended effects. The AI Now In­stitute and European labour groups have proposed such frameworks, arguing that those most impacted must have a say in evaluation. Beyond individual audits, there is also a question of who is empowered to look under the hood. Given the increasing deployment of AI systems in the workplace, employers should be required to carry out regular fundamental rights impact assessments for all AI systems used there, not only narrow algorithmic man­agement tools. These assessments should look not just at isolated decisions, but at systemic patterns : for ex­ample, whether an attendance-scoring system systemat­ically penalises workers with care responsibilities, or whether a risk model consistently channels inspections towards particular groups. Where such assessments re­veal unlawful, discriminatory or otherwise unfair im­pacts, employers should be obliged to modify or discon­tinue the system, not simply to document its behaviour. A further step is to clarify what meaningful oversight looks like in practice. Persons charged with monitoring AI systems at work should receive specific training, have the authority to override automated decisions, and enjoy legal protection against dismissal or other retaliation when they exercise this role. They should be able to flag high risks of discrimination or fundamental-rights viola­tions and, where necessary, trigger the modification or suspension of the system. Without such protections, »human oversight« risks becoming a rubber stamp rather than a safeguard. Conclusion – towards algorithmic solidarity and critical literacy Algorithms are not destiny. They are human artefacts, conceived, built, tested and deployed through choices. Like all tools, they can liberate or constrain, democratise or dominate. The question is not whether algorithms will govern work, but how, for whom and under whose control. The challenge is not to reject technology, but to reshape it. That means demystifying the black box, insisting on enforceable rights to information and explanation, and turning audits and fundamental rights impact assess­ments into levers of accountability rather than box-tick­ing exercises. It means asserting collective rights over data and digital infrastructures, so that AI systems in the workplace are aligned with labour standards, equality law and occupational health and safety, not just with productivity metrics. In the end, understanding how algorithms are made is not just a technical exercise. It is a political act, one that reminds us that even in the age of automation, the most important variable in any equation remains human judge­ment and collective organisation. As scholar Shoshana Zuboff(2019) warned,»surveillance capitalism claims private human experience as free raw material«. But as unions have always known, what is tak­en can also be reclaimed. The task ahead is to reclaim the algorithm, not as a boss, but as a contested, negotiat­ed collaborator; to ensure that the next generation of sys­tems reflects solidarity, not subordination; to insist that the digital future is, above all, human . Conclusion – towards algorithmic solidarity and critical literacy 7