Why Does an Algorithmic Decision Not Automatically Have Authority?

A company’s rostering algorithm rejects an employee’s request for leave next week. The system may combine customer forecasts, colleagues’ hours and role qualifications. It may calculate faster than a manager and may even be more accurate most of the time. But why should the employee accept its result? “It calculates well” answers an epistemic question; “it has the right to decide” answers a normative one. The two cannot simply be merged.

Philosophy commonly distinguishes theoretical from practical authority. A doctor may be an expert whose medical judgement deserves belief, but that does not give the doctor command over every part of a patient’s life. Practical authority claims that its directives themselves give reasons for action. Joseph Raz’s service conception holds that one important route to legitimate authority is that subjects are more likely to comply with the reasons already applying to them by following its directives than by reconsidering every reason independently. This is an account of how authority can serve agents, not a licence that automatically turns any efficient tool into a ruler.

Applied to algorithms, the distinction reveals three levels. An algorithm may merely provide a prediction, such as next week’s customer numbers. It may implement rules already settled by an institution. Or it may effectively select rules, trade interests and make the final decision. The first level mainly demands reliable evidence; the second also requires fidelity between rule and implementation; the third enters the territory of power and legitimacy. No accuracy score can by itself decide who may set the objective, whose interests may be sacrificed or who is responsible for error.

My view is that an algorithm can carry delegated functions, but technical performance alone cannot make it an authority. For a rostering decision to bind, it should at least be traceable to a legitimate institutional mandate, explain the relevant standards, permit correction and leave an accountable person or institution answerable to appeal. A human signature does not automatically make a decision legitimate either. What matters is whether its reasons, powers and responsibilities can be examined.

When told “that is what the system calculated”, the right response is therefore not immediate obedience or rejection. It is to ask whether the system is supplying information, applying a rule or exercising a judgement that the institution remains obliged to justify.

https://plato.stanford.edu/entries/authority/
https://plato.stanford.edu/entries/legal-obligation/