The driver the app deactivated

Consider a driver who has spent the last two years earning a living through an app. The work has a shape by now: the early shifts that pay best, the neighborhoods that tip, the ratings kept carefully above the line that keeps the account in good standing. Then one morning the app will not start a shift. In its place is a notice. The account has been deactivated. The reason given is a category — a policy heading, a phrase about "suspicious activity" or a rating that has "fallen below our community standards" — and beneath it a button to acknowledge and, sometimes, a form to appeal. There is no name attached, no manager to call, no office to visit. Somewhere a model produced a score, the score crossed a threshold, and the income stopped. The whole event took less than a second and involved no human decision at all.

What has happened here is worth naming precisely, because the ordinary words do not quite fit. The driver has not been fired, exactly; no one fired them. They have not been disciplined, exactly; discipline implies a disciplinarian. They have been deactivated — a word borrowed from equipment, and revealing for it. It is the vocabulary of turning something off. And it captures something true about the new situation of work: that the manager, the party who used to owe the worker a reason and a hearing, has been replaced by a system that owes nothing because it is not anyone.

The manager that is not a person

Algorithmic management is not a metaphor. It is a real and now-common arrangement in which the functions a human supervisor used to perform — assigning work, measuring performance, enforcing standards, and removing people who fall short — are carried out by software acting on data. Much of this can be defended. A dispatch model that routes the nearest car is not a moral problem; it is logistics. The problem arrives at the sharp end, where the same apparatus that assigns a ride also revokes the ability to earn. A deactivation is the most consequential act an employer can take against a worker, and in this arrangement it is performed by a model whose output is a category and a threshold.

The notice the driver receives is the tell. It categorizes the event without explaining it. "Suspicious activity" is a bucket, not a finding; it names the drawer the decision was filed in, not the thing that was actually observed. And the driver is not a party who can shrug this off. They are economically dependent on the platform — often it is most or all of their income — which means the deactivation is not merely a lost opportunity but a coercive act. Dependence is what turns an automated output into a punishment. A recommendation you can walk away from is advice. A determination that ends your livelihood and answers to no one you can reach is power.

The reason that isn't one

The deepest defect is not that the system is wrong — it may often be right — but that its reason is not a reason the worker can engage. A flag is not a finding. When a fraud model marks a pattern of trips as anomalous, or a rating aggregate dips under a line, what the worker receives is the conclusion of a process whose inputs are invisible to them. They cannot see the customer complaint they are being judged by, cannot examine the trips that were scored as fraudulent, cannot know whether the anomaly was a stolen device or a bad week or a false positive that fell out of the math. You cannot rebut what you cannot see. The appeal form asks the driver to contest a determination whose actual basis has been withheld from the person contesting it.

The app did not fire the driver in anger or by mistake; it fired them the way a thermostat turns off a room — and left them the same amount of recourse.

That comparison is the crux. A thermostat is not cruel; it is simply indifferent, and it cannot hear you. There is a special harm in being removed from your work by something that has no capacity to receive an appeal — not a hostile manager who might at least be persuaded, not a bureaucracy that might at least be exhausted, but a process to which the very concept of "but wait, let me explain" does not apply. The worker is left arguing with a closed system whose only interface is a button that says they understand. They do not understand. They have been told a category and asked to accept it.

What the account owes the worker

The remedy is not that platforms must stop using models. It is that a decision this consequential must carry the accountability a human manager would have owed, because moving the decision into software did not dissolve the obligation — it only hid it. Three things follow. The worker is owed the actual basis of the adverse determination in a form they can contest — not the policy heading but the specific conduct or signal that triggered it, rendered concretely enough to answer. They are owed a route to a human with authority to reinstate — a person, not a form, who can look at the underlying facts and reverse the outcome, because a right of appeal to something that cannot change the result is not a right at all. And the whole event is owed a preserved record — the inputs the model saw, the rules in force at the time, the score and the threshold it crossed — so that the determination can be examined after the fact rather than taken on faith. This is what a Decision Receipt is for: not to explain the model's inner workings, but to fix what was decided, on what basis, under which rules, so the worker has standing to challenge it and an auditor has something real to check.

Honesty requires a caveat, because it cuts against the platform's strongest argument. Some fraud signals genuinely need partial confidentiality; publishing the exact features that trip a fraud model teaches the next fraudster how to evade it, and a worker's right to contest does not extend to a blueprint for gaming the system. But that concern is narrower than it is usually made to seem. What must stay secret is the detailed mechanism of detection. What need not — what cannot, if the worker is owed anything at all — is the fact that an adverse determination was made, its existence as a contestable event, its route to a human who can reverse it, and its auditability by someone with authority to review. You can protect the fraud model and still owe the worker a real hearing. The two are separable, and keeping them fused is a choice, not a necessity.

Livelihood decisions are the ones we have always held to the highest standard of accountability, precisely because they are the ones that can ruin a life quietly and quickly. Automating them does not lower that standard; it raises the stakes of meeting it, because the automated version is faster, cheaper to issue, and easier to leave unexamined. A worker disciplined by a model is still a worker being disciplined. The account that manages them owes what any employer has always owed — a stated reason, a chance to answer, and someone who can say yes.

The scenario above is illustrative — a composite drawn to show a pattern, not an account of any real person, company, or event.

— Dispatches · Summit Cognitive