The App Knows Before You Do
For millions of gig workers, the boss is an algorithm — one that scores, routes, and silently drops you with no appeal and no explanation.
Marcus Webb has been driving for Amazon Flex in Memphis since 2021. He drives his own car — a 2019 Hyundai Sonata with 94,000 miles on it — and he picks up three or four four-hour blocks a week, mostly early morning, delivering to the kind of neighborhoods where people are already at work by the time the packages arrive. He makes about $900 on a good week. On a bad week, which usually means the algorithm didn't offer him blocks, he might make $300.
Last November, Marcus was deactivated. The notification came through the app: his account had been suspended pending review due to "delivery quality concerns." No specifics. No appeal form. A phone number led to a recorded message. He was reactivated nineteen days later with no explanation, during which time he had made $0, fallen behind on his car payment, and paid a $37 overdraft fee. Nobody at Amazon ever spoke to him. The entire event, from deactivation to reactivation, was handled by automated systems. He still doesn't know what triggered it.
No appeal. No explanation. No one.
The phenomenon of Amazon Flex drivers deactivated without explanation has been documented repeatedly over the past several years, and the pattern is consistent: the action comes from an automated quality score, the driver is given no specifics, the appeals process (where one exists) is slow and opaque, and the drivers who get their accounts back often do so because the score eventually corrected itself rather than because anyone reviewed their case. It is a labor relationship in which one party has all the information and the other has none.
What Amazon has built — and what Instacart, DoorDash, Uber, and Lyft have built in parallel — is a management system of substantial sophistication. The algorithms track on-time rates, delivery photo compliance, customer ratings, route efficiency, app usage patterns. They surface scores and route assignments and deactivation decisions at a scale no human HR operation could match. They are, in a narrow sense, more consistent than human managers: they don't play favorites, don't have bad days, don't hold grudges. They are also incapable of context. They cannot know that Marcus's on-time rate dropped last October because his mother had a stroke and he was distracted for a week. They don't care. The score is the score.
The specific logic of these algorithms is proprietary. Drivers know their "delivery quality" rating in rough terms but rarely the weight assigned to any individual input or the threshold that triggers review. This information asymmetry is not accidental. It is the central feature of the system.
When a human manager can terminate you, you can at least ask why. The answer may be pretextual, the process may be unfair, but there is a person across a table. You can push back. You can find a witness, file a grievance, escalate. When an algorithm deactivates you, the pushback goes into a support ticket queue and surfaces, weeks later, as a form response. The worker has been moved from a labor relationship into something closer to a customer service relationship — one where the "customer" has very little bargaining power.
I found myself reading through Flex's Terms of Service while reporting this, looking for the language around termination. What I found was a clause that effectively reserved Amazon's right to end the relationship "at any time, for any reason, or for no reason." Standard contractor boilerplate, technically. But when you read it alongside the deactivation stories, it stops looking like legal caution and starts looking like a design philosophy.
Labor lawyers have started calling this "ghost management": the company maintains that no employment relationship exists (these are "independent contractors") while exercising detailed, real-time control over how the work is performed. The legal structure says one thing; the algorithm does another. The courts are still working through what this means.
Marcus told me he has learned to work around the algorithm as best he can. He photographs every delivery, even when it isn't required. He never leaves a package without a confirmation photo. He keeps his acceptance rate high because he's heard — though he can't verify it — that drivers who decline too many blocks get fewer offers. He checks the app constantly, because blocks disappear fast.
This is a specific kind of cognitive labor that doesn't show up in any efficiency metric: the mental overhead of managing your relationship with a system that is simultaneously your employer, your performance reviewer, and your judge, and that will not explain itself to you or hear your explanation in return. It falls hardest on the people with the fewest choices — those who can't easily fall back on other income, who are driving because the job doesn't require a credential or a fixed schedule — and it costs them something that doesn't appear in any quarterly report.
What is striking about algorithmic management, if you spend time with people subject to it, is how thoroughly it has solved the problem of accountability for the company while creating a new one for the worker. There is no one to be angry at. The manager who fired you could, in theory, be shamed. The algorithm cannot be shamed. It doesn't read the editorial you wrote about it.
This is not a side effect. It is, on some level, the point. A system that makes consequential decisions without a human face is a system that is much harder to hold responsible. The decision came from a model. The model is proprietary. The company cannot comment on individual cases. The worker is welcome to reapply.
Marcus is still driving. He said he can't afford to stop, and that the work, when it's steady, is actually okay — he likes being alone in the car, likes the early-morning city before it fills up. What he doesn't like is the uncertainty. He checks the app before he goes to bed and again when he wakes up. The algorithm knows his schedule better than his family does. He just can't know what it thinks of him.