A traditional manager has limitations the platforms found unacceptable. A human manager can only watch so many workers at once. A human manager builds relationships, which interferes with their willingness to enforce arbitrary rules. A human manager is expensive, slow, and prone to being argued with. A human manager has, in the long view, been the largest single source of friction inside any large workforce, and the entire economic logic of the gig economy is that this particular kind of friction can be removed by replacing the manager with an app.
What replaces the manager is not, technically, a manager. It is a control loop. The control loop sets a target — pickups per hour, units scanned per minute, delivery on-time rate, customer star rating — and then adjusts the conditions of the worker's environment in real time to keep the worker close to the target. The conditions it can adjust include pay per task, queue position, geographic dispatch, time between assignments, and the worker's continued ability to log in at all. The worker experiences this as a series of small inexplicable rewards and small inexplicable punishments. The system experiences it as control.
§ 2.1The Skinner box, productionised
The behaviourist tradition called this operant conditioning: shape behaviour by attaching consequences to it, schedule the consequences so the subject can't predict exactly when they'll arrive, and the subject will continue pressing the lever well past the point where they could articulate why. Skinner's pigeons would peck a panel for hours on a variable-ratio schedule. The schedule mattered more than the reward size. Unpredictability was the trick.
Rideshare dispatch is a variable-ratio schedule. You don't know when the good ping comes — the airport run, the surge zone, the long trip — but you know it might. The driver who keeps the app open for the extra hour is the driver who catches it. The driver who logs off is the driver who didn't. Over millions of drivers and billions of hours, this produces exactly the distribution of behaviour the dispatch system was designed to produce: drivers staying online longer than they otherwise would, in worse weather than they otherwise would, for marginal pay they otherwise wouldn't accept. The system isn't even hiding this. It's the documented design.
Amazon warehouse "rate" is the same loop in a different shape. The picker's handheld scanner displays an idle-time counter. Cross the threshold and a manager (or, increasingly, the system itself) issues a warning. Accumulate warnings and you're out. The picker has internalised the metric within their first week. By month three, they no longer notice the scanner is watching them; they notice their body, which has reorganised itself around hitting the rate. This is conditioning at the level of musculature.
§ 2.2Wages as the dial
The most ideologically inconvenient part of algorithmic management is that the wage rate itself is now a real-time control variable. Earlier labour systems treated the wage as something negotiated, posted, and stable across a shift. Rideshare and delivery platforms treat the wage as a tunable input — raise it when supply is short, lower it when supply is dense, target it precisely to whatever fraction of a driver's reservation wage the optimiser thinks will move them. There is no posted rate. There is the rate the algorithm offers you, individually, in this moment.
Internal Uber documents and subsequent academic reconstruction have shown that this targeting is not symmetric across drivers. Drivers identified as more elastic — newer, in tighter financial situations, with shorter recent earnings histories on the platform — get offered different rates than less elastic drivers on the same trip. The label of art for this is "personalised pricing." The label of art for what it does to a labour market is something else.
When the wage becomes a slider on the operator's dashboard, the labour market is no longer a market. It is a control system with the worker inside it.
"I drove the same Tuesday morning route for three years. Same airport pickup, same return. Last year the rate started moving every week. I asked support. They said it was 'dynamic.' I asked another driver and his rate was different than mine for the same trip at the same time."
"You stop trusting the number. Then you start chasing the number."
§ 2.3The rating as collar
The five-star rating system was the public-facing innovation of the early gig platforms. It was marketed as accountability — passengers could rate drivers, customers could rate Tasks. Almost nobody outside the platforms knew, in the early years, that the rating was not a piece of feedback for the worker. It was the leash.
Drop below 4.6 on Uber and you are at risk of deactivation. The line is opaque, the appeals process is automated, and the decision is, in practical terms, final. Workers do not know which passenger left the bad rating, or why, or whether it was the rating itself or some aggregate score that triggered the action. They know the consequence: the income stream is gone. The behavioural pressure this creates is enormous, because the punishment is severe and the trigger is illegible. The worker becomes hypervigilant about every interaction, every route choice, every comment. This is the classic shape of conditioning under uncertainty: the subject overcompensates because they cannot identify the rule.
This is also, not incidentally, a remarkably efficient way to extract emotional labour. Smile more. Open the door. Don't dispute the route. Don't mention you've already been driving for ten hours. The five-star rating made every rider into a part-time supervisor with the power of capricious termination, and made the company itself look like a neutral platform passing along feedback. It was never neutral. It was the discipline.
§ 2.4Why "the algorithm" is the wrong word
The word algorithm has done a great deal of rhetorical work for the platforms, because it sounds technical and neutral. It is neither. The systems described above are management decisions encoded in software. Someone, in a room, decided what the rate-cut threshold would be. Someone decided that drivers identified as more financially desperate would be quoted lower rates. Someone decided the deactivation appeal flow would route to an automated decision rather than a person. These are not properties of mathematics. They are properties of a firm.
The conditioning works because the firm wants it to work. When the firm does not want a conditioning effect — when, for example, the rate cuts are producing enough driver attrition to threaten supply — the firm adjusts the parameters and the effect changes. The lever exists. The hand on the lever exists. The opacity exists not because the system is too complex to explain, but because the explanation is bad public relations.
§ 2.5The result, by the operator's own metric
By the metrics the platforms set for themselves — driver supply elasticity, order throughput per labour hour, customer-facing on-time rate, churn, lifetime value — the systems work. They work so well that they are now the template for next-wave deployments in trucking, in-home care, retail floor staffing, and call-centre work. The proof of concept is closed. The remaining question is scope.
What none of this delivers is the thing it was supposed to deliver in the original marketing: flexibility, autonomy, "be your own boss." Workers on these systems are not their own bosses. Their boss is a control loop, and the loop is better at being a boss — by the firm's standards — than any human ever was. It does not get tired. It does not feel guilt. It does not develop loyalties. It does not need to be paid.
This is the second of three case studies. The conditioning patterns documented here did not originate in labour. They were imported from consumer platforms, where they were first proven at scale. They are now being exported again, into the public sector. That is Case 03.