Power appears where someone can say enough. In dignified work in the automatic age, supervision was decorative because responsibility and authority were separated.
A person may be listed as the supervisor of a hiring, scheduling, or performance system yet lack access to its records, time to investigate, or permission to stop a consequence. That human presence does not amount to effective control. It merely adds a name to a process whose power remains lodged in the automated rule, the vendor, or a management structure that never receives the complaints.
Decent work requires authority to be located before harm appears. This does not mean asking people to remake every decision manually. It means ensuring that a competent person can challenge a datum, hear the affected worker, suspend a measure, and order a remedy. Three institutional cases show why that capacity matters: Uber drivers before the United Kingdom Supreme Court, Amazon France Logistique employees before the French data authority, and iTutorGroup applicants before the United States employment equality authority.
Supervision means being able to change the outcome#
Automation distributes power through apparently small decisions. It defines which job arrives first, which break counts as unproductive, which application deserves a reading, and which score opens access to a shift. Each operation may be described as a recommendation even though its repetition organizes income, time, health, and continued employment. To identify who is in command, an organizational chart is not enough. The inquiry must ask who sets the objective, who can inspect the evidence, and who has the authority to say enough.
Automation distributes power through apparently small decisions. It defines which job arrives first, which break counts as unproductive, which application deserves a reading, and which score opens access to a shift.
The difference between responsibility and authority can be recognized through a simple test. If a person must answer a complaint but cannot see how the result was produced, responsibility is decorative. If that person can see it but needs the vendor’s permission to correct it, control remains insufficient. If an individual case can be changed but a rule that repeats the same harm cannot, there is partial remedy rather than governance. Meaningful control combines access, competence, independence, time, and power to intervene.
Timing matters as well. A later review may recognize that a dismissal, loss of shifts, or rejection was wrong after the opportunity has disappeared. For high-impact decisions, an authority must be able to suspend the consequence while a plausible disagreement is checked. The speed of the system cannot become a reason to make its error irreversible. The harder a consequence is to repair, the earlier control must operate.
Dignity does not mean immunity from evaluation. An employer may organize work, measure results, and address failures. It should not confuse partial signals with a complete history, hide the rule, or force the person bearing the consequence to prove a defect without access to the information. A dignified employment relationship allows the criterion to be understood, context to be offered, and an appeal to be made without retaliation.
Case one: Uber and the control behind the application#
In February 2021, the United Kingdom Supreme Court decided Uber BV v Aslam and confirmed that the claimant drivers were workers for the purposes of British employment legislation. The judgment did not stop at the wording of the contracts. It examined the practical relationship: Uber set the fare, imposed contractual terms, controlled information about each journey, used passenger ratings, and could restrict access to the application.
The case helps distinguish formal freedom from material power. A driver could choose when to log in, but once available operated inside a framework configured by the platform. Acceptance of trips, ratings, and the risk of disconnection affected the driver’s room for decision. The algorithm was not a neutral coordinator placed between equal parties. It executed conditions that the company designed and could modify.
The issue of authority is especially clear in management by reputation. A rating aggregates different experiences and may reflect bias, events beyond the driver’s control, or incompatible expectations. If a score triggers warnings or loss of access, review cannot merely repeat the average. It must examine particular journeys, unusual patterns, context, and representations. The reviewer must also be able to prevent disconnection before it removes a source of income.
The lesson is not to transfer a British legal category mechanically to every country. It is to observe the reality of control. An organization retains responsibility when it determines prices, incentives, thresholds, or consequences even when everyday contact occurs through an interface. Calling an automated decision independent does not make the affected person independent either. Analysis must follow effective powers, not only the labels selected by the system’s designer.
Case two: Amazon France and surveillance turned into discipline#
In January 2024, the Commission nationale de l’informatique et des libertés announced a fine of 32 million euros against Amazon France Logistique for processing connected with warehouse employees’ activity. Scanners recorded events that fed indicators about quality, productivity, and inactivity. The CNIL considered certain indicators excessively intrusive and challenged the detailed retention and use of data for some purposes.
An interval between scans looks precise, but it does not explain its cause. It may represent movement, waiting, a necessary conversation, a faulty device, reorganization, or a required break. When an indicator turns the absence of a scan into suspicion, the person must defend every fragment of time. Measurement stops describing the work and begins disciplining it.
The CNIL acknowledged that planning, training, or safety could be legitimate purposes, but found that some objectives could be met with less detailed information. That proportionality has a direct effect on power. The finer the granularity, the easier it is to place the burden of explaining every deviation on a worker. The organization should justify why each datum is needed, who will see it, how long it will remain, and which decision it may activate.
Real supervision would require that no isolated threshold automatically produce a significant employment consequence. Review would compare equipment incidents, task changes, safety risks, local conditions, and the person’s account. It would also search for collective patterns. If many intervals appear in the same area, poor organization may be a more reasonable hypothesis than individual misconduct. The supervisor needs authority to correct both the record and the indicator.
Case three: iTutorGroup and perfectly consistent discrimination#
In August 2023, the United States Equal Employment Opportunity Commission announced a 365,000-dollar settlement of its lawsuit against iTutorGroup. According to the EEOC, the company’s recruiting software automatically rejected women aged 55 or older and men aged 60 or older. More than two hundred qualified people were allegedly screened out because of age. The settlement included compensation, training obligations, policy review, and monitoring.
In this case the system did not fail to execute the instruction. It applied it consistently. That was precisely why automation could expand the harm. Technical stability did not demonstrate neutrality because the embedded criterion was discriminatory. An overall accuracy figure would not have answered the relevant question either: who was excluded, on which attribute, and with what opportunity for reconsideration before the vacancy closed.
Responsibility does not disappear because a vendor writes the code or a tool filters thousands of applications. The organization choosing to use it must know the relevant fields, test outcomes across groups, review thresholds, and retain the ability to stop it. Human review placed at the end is of little value if it sees only the applications that the filter allowed to advance. Control must include samples of rejections, tests with equivalent profiles, and analysis of changes after every update.
Remedy does not end when the criterion is removed. Rejected people lost the opportunity to compete while the position was available. Effective authority therefore combines prevention, suspension, and restitution. When an error is discovered, affected decisions must be located, the problem communicated, opportunities reopened where feasible, and harm remedied under the applicable framework. Fixing the model without addressing its consequences preserves institutional efficiency at the expense of those who bore the damage.
Designing authority that is not ceremonial#
The three cases support an architecture of control. First, each system needs a limited purpose and a named responsible person inside the user organization. Second, that person requires access to understandable records rather than only a final score. Third, the person must be able to stop a consequence, request additional evidence, and change the outcome. Fourth, corrections must feed a review of the rule. Fifth, an independent level of review must exist for serious or repeated disputes.
Separation of functions protects both the workforce and the reviewer. The team pursuing productivity should not be the sole authority over health, discrimination, or wage complaints. Participation from occupational safety, human resources, data protection, worker representation, and legal advisers may be needed according to the risk. Cooperation does not dilute accountability when the final decision is clearly assigned. It strengthens responsibility by gathering knowledge that the indicator does not contain.
The appeal channel should be visible, free, and usable during working time. It must explain which decision is disputed, which information influenced it, what the person may provide, who will decide, and within what period. It should not require technical language or automatically return the complaint to the same system. It must also protect against retaliation and offer an alternative route for anyone unable to use the application or requiring an accommodation.
Authority is tested through exercises, not by a checkbox in a policy. Before deployment, an organization can rehearse a network failure, an incomplete record, a biased score, a safety break, and a non-standard application. The responsible person should demonstrate an ability to find the evidence, pause the effect, hear the context, and document the decision. If permissions must be improvised during the exercise, the system is not ready.
Measuring power, harm, and remedy#
Average accuracy is an insufficient indicator. A responsible evaluation includes false positives and negatives, differences among groups, appeal frequency, the proportion of corrections, response time, suspended consequences, and restored opportunities. It also observes recurrence. If the same kind of complaint returns, the organization has not learned even when it resolves isolated cases.
Data must be accompanied by accounts and participation. A break may look identical in a table while arising from completely different causes. People performing the task know dangerous shortcuts, equipment failures, invisible burdens, and moments when the indicator rewards unsafe conduct. Collective representation can bring dispersed experiences together and reveal that an alleged exceptional case is a structural condition.
Information proportionate to the impact should be published: purpose, data categories, general logic, consequences, controls, evaluation results, and appeal routes. Transparency does not require the disclosure of trade secrets or mechanisms that enable fraud. It requires enough information for a person to recognize how the rule affects them and to challenge an error. An explanation that merely lists technologies does not perform that function.
Remedy should be defined before an incident occurs. It may include correcting the record, restoring shifts or income, reopening an application, withdrawing a warning, compensating harm, and changing the system. Each measure needs an owner and a deadline. Advance preparation stops the organization from treating the first complaint as a surprise and reduces the temptation to defend the tool in order to protect an investment already made.
Institutional sources and a decision standard#
The analysis rests on four principal institutional sources. The United Kingdom Supreme Court case page and judgment in Uber BV v Aslam document the reality of control examined by the court. The CNIL publication concerning Amazon France Logistique explains the indicators, penalty, and proportionality assessment. The EEOC release about iTutorGroup records the allegations and settlement measures. The International Labour Organization study of algorithmic management locates these mechanisms in platforms and conventional workplaces.
The sources perform different functions. Judgments and agency releases describe particular facts and legal frameworks, while the ILO study supplies a comparative view of work organization and job quality. They do not make every use of automation unlawful or allow a holding to be moved unchanged into another jurisdiction. They do support a shared conclusion: when a tool organizes opportunities, conduct, or income, its governance must examine effective power rather than stop at technical appearance.
Before approving a system, management should answer six questions in writing: which decision it assists, which data it uses, which harm it can cause, who can pause it, how an appeal works, and how remedy will be delivered. If a responsible person is named but lacks permission, budget, or independence, the answer is incomplete. If the vendor alone understands the records, the organization has outsourced capability, not responsibility.
Power appears where someone can say enough and make that instruction effective. In work that does not surrender dignity, supervision is not a figure seated beside a machine. It is an institutional capacity to understand, interrupt, listen, correct, and learn. Automation can coordinate tasks and expand knowledge; it deserves trust when it also accepts limits. Where nobody can stop an unjust consequence, there is no human supervision. There is an automated decision accompanied by a witness.
