£3.67 to £3.42 an Order: Inside Edinburgh Riders' Fight to Open the Algorithmic Black Box

A quiet pay cut, measured one order at a time

On a weekday afternoon in Bristo Square, central Edinburgh, a group of food delivery riders gather beside their bikes and insulated bags before the dinnertime rush. They compare notes — not just about restaurants and routes, but about a number that has been sliding for three years: the average fee per order.

David, a rider for seven years who withheld his surname, told The Guardian: “I am making half the money I was making four years ago, for the same amount of hours. It makes no sense.”

Xabier Villares, riding for eight years and now lead organiser of the Workers’ Observatory, put it more sharply: “There has been a dramatic change in the last three years. I used to work four or five days a week, especially evenings and some lunchtimes, and was able to pay my rent and bills and make a decent living with that. But that’s not an option any more.”

The riders are not imagining the decline. Dylan (name changed), who has ridden for Deliveroo in Scotland for more than five years, has kept meticulous personal records since mid-2023 — income, orders, average fee. Orders per hour have held steady between 3.6 and 3.8. The average fee per order has not: £3.67 in 2023, £3.63 in 2024, £3.51 in 2025, and £3.42 in the first half of 2026. Same workload, roughly 7% less pay per order — compounding year after year.

Year	Avg fee per order (Scotland, one rider’s records)
2023	£3.67
2024	£3.63
2025	£3.51
2026 H1	£3.42

The black box nobody can see into

Platforms describe their allocation and pricing systems in confident terms. Deliveroo’s algorithm, named Frank, has been described by the company as “super smart” and able to decide “which rider to offer which order” — machine-learning technology that “predicts the timings of every order” so food arrives “in the most efficient and reliable way.”

Riders see a different system. Rider Graeme Frances believes the platforms “rely on [the work] being opaque”: “You don’t know what is in the black box of how they work out what to pay to whom.”

What fills that black box is a family of techniques the industry calls dynamic pricing — offer levels that move in real time with supply and demand, and, unions allege, with what an individual worker has shown willingness to accept. Trades unions are campaigning to ban the practice outright, arguing it leaves workers with no stable basis to plan their earnings. The academic evidence leans their way: research from the University of Oxford and Columbia Business School found that the introduction of a dynamic pricing algorithm in 2023 resulted in Uber drivers earning “substantially less” per hour.

That research now has a legal counterpart. This week, drivers from the UK, the Netherlands and other countries launched a landmark class action against Uber in Amsterdam — a case that could run into billions of dollars — alleging the AI-powered system breaches data protection law and “pushes down their earnings based on what each driver is willing to take.” Drivers say they live in “constant fear” of a “soulless” algorithm. Uber denies adjusting trip prices based on an individual driver’s behaviour, attributing discrepancies to other system features such as GPS.

Turning workers into researchers

The Edinburgh group organises through the Workers’ Observatory, a charity founded by gig economy workers alongside academics at St Andrews and Edinburgh universities. Its premise is simple: if platforms won’t explain their systems, workers can measure them.

The observatory has just secured research funding for the next decade, and its method is already producing results. In one experiment in Dunfermline, Fife, a group of riders logged on to a delivery platform simultaneously, with some rejecting job offers below a certain rate. The outcome: pay briefly rose for some workers, while others were penalised — one was deactivated from the platform shortly afterwards, for reasons that were never made clear. Observatory director Cailean Gallagher, a lecturer at St Andrews business school, said the response “seemed very arbitrary.” Platforms insist riders are not deactivated for rejecting offers.

Gallagher frames the observatory’s mission as getting “on the ground floor in terms of understanding how the whole apparatus works,” because “there is so much infrastructure of knowledge and data that’s concealed” — leaving riders “working in the dark.”

This is algorithmic management meeting its most systematic countervailing force: coordinated workers with spreadsheets, structured logging protocols, and academic partners. It resembles nothing so much as citizen science turned industrial.

When automation has no fallback: the facial recognition problem

Pay is only half the story. The other half is access to work itself.

In August 2025, rider Graeme Frances got a black eye in an accident outside work. Deliveroo’s facial recognition system then refused him log-on, repeatedly, and no human support was reachable. He was locked out of earning for the duration of his visible injury; he says his request for a manual review went unanswered until he next logged in — leading him to believe he simply could not work until he had healed. Deliveroo acknowledged he failed the checks multiple times and attributed it to a blurry picture, saying he would have been able to work after the manual review.

The pattern is familiar to anyone who has studied automated decision systems: the automated gate fails precisely when reality deviates from training data (a bruised face), and the human fallback is thin to invisible. Under the EU’s Platform Work Directive — fully applicable since August 2026 — platforms must inform workers about automated monitoring and decision systems and, in many cases, obtain human review. The UK, post-Brexit, has no equivalent blanket rule, which is precisely why the Amsterdam class action leans on GDPR’s data protection hooks instead.

The industry response

Deliveroo, Uber Eats and Just Eat all say riders earn more than the national living wage while on an order — a framing that excludes waiting time between orders, which for many riders is a large share of working hours. Deliveroo noted its minimum hourly fee while on an order rose 3.8% this year (above inflation) and says it partners with the GMB union; it also said “any significant decision over a rider account is not automated, but reviewed by our team.” An Uber spokesperson said its matching tools “balance a number of different factors, such as time and distance.” Just Eat said it uses technology “overseen by a human team.”

These are carefully qualified statements. They don’t address the core demand: publishing the logic that maps rider history, location, and acceptance behaviour to offer levels. Until they do, the observatory’s decade of funded research will keep pulling in the opposite direction.

Why this matters beyond delivery riders

The Edinburgh riders are early participants in a fight that will define AI-era labour markets: who gets to audit the algorithms that set pay?

Three forces are converging:

1. Regulation is arriving. The EU Platform Work Directive (applicable August 2026) creates transparency and human-review rights around automated decision-making for millions of platform workers. Early enforcement patterns — like Italy’s competition authority requiring Glovo to disclose its AI delivery-assignment logic — show regulators are willing to force the box open.
2. Litigation is scaling. The Amsterdam Uber class action tests whether GDPR’s protections against solely-automated decisions can reach into pay-setting algorithms, with damages potentially in the billions.
3. Worker science is professionalising. The Workers’ Observatory’s decade of secured funding means systematic, peer-reviewed measurement of algorithmic pay is coming — not anecdotes, but datasets.

For AI practitioners, the lesson is uncomfortable. The same optimization objective that makes Frank “efficient” — minimizing delivery cost under demand constraints — has a distributional consequence: the surplus squeezed out of each order has to come from somewhere, and riders’ logs suggest where it is coming from. Efficiency gains that are invisible in aggregate metrics show up as £0.25 declines in per-order fees in individual riders’ tax spreadsheets.

The measurable decade ahead

The observatory’s funding horizon — ten years — is itself a statement. Algorithmic management will only deepen: stacked orders, facial recognition gates, dynamic offers tuned to individual willingness-to-accept. Against that, riders are building the counter-infrastructure: logging protocols, coordinated experiments, legal strategies, and now the patience of a funded research institution.

The question the next decade answers is whether transparency can be won measurement by measurement — or whether the black box holds until a court or regulator opens it wholesale. Either way, the riders of Bristo Square have decided they will not wait to be told what their time is worth.