algorithms are trained to predict what will keep each user active for the longest time, continually adapting to individual behaviour. That’s why outrage, misinforma­tion and sensationalism spread faster; they keep users scrolling(Geers et al. 2024). In this sense, the system is not malfunctioning, but doing precisely what it was de­signed to do, even if the broader social consequences are toxic. In the world of work, algorithmic management systems can be tuned to extract maximum productivity. Gig plat­forms such as Deliveroo or Uber use real-time data and optimisation algorithms to allocate jobs, monitor perfor­mance and adjust pricing, all automatically(Zhu et al. 2024). These mechanisms decide who gets the next task, how long they have to complete it and whether their account will remain active. For many workers, the model effectively functions as both supervisor and eval­uator, a digital boss that enforces rules without negotia­tion. Drivers rarely know how their»rating« is calculated or how fares are set, because the decision logic is em­bedded deep within proprietary systems. The algorithm acts as both boss and judge. Similar logics appear in more traditional workplaces. Warehouse workers can find their performance broken down into seconds per task, with dashboards that rank them against colleagues and trigger automated warn­ings when they fall behind(Claburn 2025). Call-centre agents may be scored on call length, script adherence and sales conversion in real time, with little scope to challenge how these metrics are defined. In such cases, the optimisation goal – maximising output per unit of time – has been hard-wired into the system, while con­cerns such as health and safety, autonomy or the right to organise are treated, at best, as secondary con­straints. These systems can also be strategically opaque (Langer and König 2023). When an algorithm decides who gets a shift or bonus, or who is first in line for redundancy, the company may claim that the system is too complex to explain. This obscures accountability and makes contes­tation nearly impossible even where EU law formally grants workers’ rights to information and explanation. In other cases, exclusion is built in by design. For exam­ple, predictive policing tools(Ziosi and Pruss 2024) such as PredPol rely on historical arrest data, datasets that overrepresent minority communities because of existing systemic racial bias in past policing. The algorithm then directs more patrols to those areas, reinforcing the same bias it inherited. The result is a feedback loop: biased data produces biased predictions, which in turn generate more biased data. Employment and welfare systems can create similar loops(Considine et al. 2022). A risk-scoring model that classifies certain neighbourhoods or demographic 6 Friedrich-Ebert-Stiftung e.V. groups as»high risk« for fraud or»low employability« will channel more inspections or fewer resources to­wards them, making it harder for individuals in those groups to improve their situation and confirming, over time, the system’s original assumptions. In each case, the algorithm does not act alone; it opera­tionalises and amplifies the intentions, incentives, ine­qualities and power structures of the humans who built and deployed it. Recognising this is crucial for trade un­ions and regulators: changing outcomes is not just a matter of»fixing the code«, but of questioning the objec­tives, business models and governance arrangements that those systems are designed to serve. Part IV: auditing the black box As awareness grows, so does demand for algorithmic audits , processes that evaluate whether systems are fair, transparent and accountable. In theory, auditing an algorithm is like inspecting a fi­nancial statement: an independent party reviews inputs, processes and outputs. In practice, it’s far trickier. Most large-scale models are proprietary and protected under trade secrecy or intellectual property laws. Companies often argue that disclosing their internal mechanisms would compromise competitiveness or security. There are three main types of audit(Metaxa et al. 2021): (i) Code audits – reviewing the source code and model parameters. These are rare because few companies grant access. (ii) Outcome audits – analysing how decisions affect different groups(for example, gender, race, region); (iii) Process audits – evaluating governance, documen­tation and design decisions. There is growing recognition that meaningful auditing requires strong, enforceable rights to information and explanation(Kerckhofs 2025). Transparency cannot be reduced to a generic notice that»an algorithm is in use«: affected individuals need to understand what the system does in practice(for example, whether it allo­cates tasks, scores performance, monitors productivity or supports disciplinary decisions), which data it col­lects and processes, who can access those data(includ­ing external providers), and which key factors drive its decisions. Where adverse outcomes occur – such as dis­missal, demotion, pay reduction, or loss of work oppor­tunities – clear, reasoned explanations must be provid­ed and made open to challenge. The same standard should apply in recruitment contexts, particularly where high-risk AI systems determine who is shortlisted or in­vited to interview.