Week 4: Algorithmic control theory

Two competing accounts of how software disciplines a workforce — surveillance and gamification — and where each one actually fits

A taxonomy, not a vibe#

Lee, Kusbit, Metsky and Dabbish split algorithmic management into three functions — directing (assigning the work), evaluating (rating it), and disciplining (acting on the rating) — and the split matters because most critiques collapse all three into a single word, “surveillance,” that ends up meaning everything and explaining nothing. A dispatch algorithm that routes you to a pickup is directing. A star rating is evaluating. A suspended account is disciplining. Naming which one a specific mechanism is doing is the first move in arguing about it credibly.

Ratings as discipline nobody signed up to deliver#

Rosenblat and Stark’s 2016 paper, “Algorithmic Labor and Information Asymmetries: A Case Study of Uber’s Drivers” (International Journal of Communication), is the canonical version of this argument and it generalises directly to delivery: the customer supplies the number, the platform enforces the consequence, and no single actor — not the customer, who thought they were just answering a prompt, and not the platform, which is “just following the data” — has to own the decision to discipline anyone. That diffusion of responsibility isn’t an accident of the design; it’s most of what the design is for. Rosenblat and Stark’s central contribution is naming the asymmetry underneath it: the platform can see every driver’s full history and every passenger’s rating pattern, while each individual worker can see almost none of the data that determines their own outcome. Delivery platforms reproduce that same asymmetry almost exactly — a rider can see their own score, but not the distribution it’s being compared against, nor which specific deliveries moved it.

Gamification: the harder case to argue against#

Streak bonuses, badge tiers, and “you’re so close to your goal” nudges work by a different mechanism entirely — recruiting a rider’s own motivation rather than monitoring their compliance. It’s harder to build a case against a bonus than against a penalty, because nobody is visibly being punished; they’re being invited, repeatedly, to want the outcome the platform wants. A streak bonus that resets to zero the moment a rider takes a day off is functionally a penalty wearing a reward’s framing — the missed income is identical either way, but only one framing shows up in a platform’s marketing as a benefit. Weeks 8–10 come back to exactly this asymmetry: how much harder incentive design is to regulate than a formal disciplinary mechanism, because a regulator can point to a suspension and ask a platform to justify it, but there’s no equivalent hearing for a bonus structure that quietly narrowed someone’s real choices.

Before the session#

Skim the assigned Rosenblat and Stark extract and pick one sentence you’d quote in an essay, and one you’d want more evidence before trusting.