Declining pay for food-delivery riders driven by opaque algorithms — require transparency, audits and data access

“I am making half the money I was making four years ago, for the same amount of hours. It makes no sense.”

That sentence from David, an Edinburgh food‑delivery rider, signals a wider problem riders are describing: falling pay and a sense that opaque, automated systems at Deliveroo, Uber Eats and Just Eat have made the economics of the job worse. Workers organised through the Workers’ Observatory have been tracking offers on the pavement, running small experiments and logging pay. Their claim is simple: if you can’t see how the machine decides who gets which order and at what price, you can’t verify whether the system is fair.

Quick executive snapshot, what leaders should do now

Require publishable algorithmic impact assessments before major pricing or matching changes.
Provide workers machine‑readable access to the data that affects offers and pay (offers, timestamps, payouts, cancellations).
Mandate independent pre/post audits for pricing changes, and include unpaid waiting time in guaranteed‑pay calculations.

What riders are reporting, numbers, examples and limits

Multiple riders interviewed describe steady declines in per‑order fees over the last three years while their orders‑per‑hour rate stayed roughly constant. One Deliveroo courier who tracks his income, using the name Dylan, provided monthly averages for fee per order: £3.67 in 2023, £3.63 in 2024, £3.51 in 2025 and £3.42 in the first half of 2026. Dylan reports his orders‑per‑hour has hovered between 3.6 and 3.8 across that period.

Those figures are self‑reported and represent a single rider’s time series (n=1). From 2023 to H1 2026 Dylan’s per‑order fee fell from £3.67 to £3.42, a decline of roughly 6.8%. That is meaningful at the margins for an individual. Extrapolating to the whole workforce requires larger, independently verified datasets.

“There has been a dramatic change in the last three years, ”

, Xabier Villares, Workers’ Observatory lead organiser, who says what once covered rent and bills no longer does.

“I like to keep a close eye on everything, normally just to make paying my taxes easier. I started noticing a couple of years ago there was actually a big drop‑off in my earnings. If you are doing the same amount of work, that equates to getting paid quite a lot less for the same work.”

, Dylan (Deliveroo rider)

Which features riders blame, matching, stacking, dynamic pricing

Riders point to three interlocking platform features that change the economics of delivery work:

Matching algorithms, software decides which courier gets which order. Deliveroo has described its matching system, what the company calls Frank, as “super smart” and said it uses machine‑learning to predict timings and match deliveries “in the most efficient and reliable way.”
Stacking, the app bundles two or three nearby deliveries into one offer. Riders say stacked offers often pay less per stop than the same jobs taken separately; one rider put it bluntly: “Some of the offers they give you view you as a clown.”
Dynamic pricing, real‑time price adjustments intended to balance supply and demand. Critics argue such systems can redistribute earnings away from workers even as customer fares rise.

All three are technical and behavioural levers that can change how value is divided between platform, customer and worker. Which of them, alone or together, is responsible for observed pay declines is an empirical question. Riders’ logs and on‑the‑ground experiments generate important clues, but large, independent datasets are needed to isolate cause from correlation.

The strongest independent evidence so far (ride‑hailing, not food delivery)

The clearest large‑scale empirical analysis so far examines ride‑hailing rather than food delivery. A University of Oxford study (preprint on arXiv: arXiv:2506.15278, Oxford press release 23 June 2025) used logs from 258 UK Uber drivers covering more than 1.5 million trips from 2016-2024. The paper finds that Uber’s 2023 switch to a dynamic pricing system coincided with passenger fares rising while drivers’ pre‑cost hourly earnings fell, from just over £22 an hour to about £19 an hour. The platform’s commission rose from roughly 25% to 29% and unpaid waiting time increased.

Methodologically this study is large and longitudinal, and it is suitable for detecting systemic shifts. It was presented at ACM FAccT and is available as an arXiv preprint. Important caveat, ride‑hailing trips differ from food deliveries in trip lengths, routing, stacking behaviour and fee models. The Oxford findings are therefore suggestive for the broader gig sector but not definitive proof that identical dynamics caused the pay patterns reported by food‑delivery riders.

Law, litigation and organising, rising pressure

Legal and collective responses are accelerating. According to reporting in The Guardian and summaries from AI Work Insights, a pan‑European class action against Uber was filed in Amsterdam on 2 September 2026. The claim alleges opaque, personalised algorithmic pay and potential breaches of data‑protection law. Plaintiffs argue systems assess what each driver will accept and adjust offers accordingly, producing anxiety and unstable earnings.

Drivers say they live in “constant fear” of the “soulless” algorithm.

Class actions and regulatory complaints frame these issues not only as labour disputes but as data‑protection problems. Under EU GDPR (and similar UK rules) people have rights about automated decision‑making (see Article 22 on automated individual decision‑making and profiling, and Article 15 on access to personal data). Plaintiffs will need access to technical evidence, models, logs and audits, to quantify alleged harms. At filing, reporting indicates detailed independent audits were not yet published. That evidentiary gap will be central to the legal process.

How platforms respond

Platforms routinely say algorithms are overseen by humans and point to guarantees. The company statements used in these disputes matter, but they do not by themselves resolve whether outcomes are equitable in practice.

“Riders are guaranteed to earn a minimum hourly fee while on an order, which … rose by 3.8% this year, an above inflation increase. The vast majority of riders are satisfied working with us, and we continue to partner with the GMB Union to raise the bar on security and wellbeing standards. Information on our algorithm and fees are set out on our rider website. We have a dedicated team in place to support riders, and any significant decision over a rider account is not automated, but reviewed by our team.”

, Deliveroo spokesperson

“Our matching tools follow a process that balances a number of different factors, such as time and distance, to provide the best possible experience for everyone using the app.”

, Uber spokesperson

“As a delivery platform we use technology, overseen by a human team, to match deliveries with available couriers.”

, Just Eat spokesperson

Those statements matter, but they raise follow‑up questions. How is the “minimum hourly fee while on an order” calculated in practice? Does it include unpaid waiting time? What constitutes “significant” account decisions that receive human review? Independent, machine‑readable worker data and third‑party audits are the only reliable ways to verify the practical impacts.

Opacity matters, three concrete harms

Inability to verify fairness. Without access to the data and decision logic, riders can’t distinguish between market effects, platform commission changes, or personalised profiling that reduces pay.
Limits on enforcement. GDPR gives rights around automated decisions and access to data, but those rights are hard to exercise if platforms refuse to publish or escrow the necessary logs and models.
Breakdown of operational trust. Arbitrary‑feeling deactivations or unexplained penalties push workers toward coordinated resistance, which can trigger unpredictable algorithmic responses and further instability.

Practical, prescriptive transparency steps

Transparency needn’t mean open‑sourcing core IP. It means measurable, auditable practices that allow workers, regulators and independent researchers to check outcomes.

Algorithmic impact assessments (AIAs). Require publishable AIAs before major changes. A useful checklist includes the objective of the change, the data inputs used, the model class (rule‑based versus ML), the geographic and temporal scope, and the key outcomes to monitor such as average payout per order, distribution by rider decile, unpaid waiting time and take rate. The AIA should also specify mitigation measures if adverse distributional effects appear.
Machine‑readable worker data. Provide exports (JSON/CSV) containing at minimum: offer_id, timestamp, pickup_lat/lng, dropoff_lat/lng (or hashed equivalents for privacy), quoted_fee, acceptance_flag, payout, cancellation_reason, and account_status_changes. A weekly dump covering the last 90 days is a pragmatic starting point. An independent escrow mechanism can support audits.
Independent audits for pricing changes. Mandate accredited third‑party pre/post audits when platforms roll out major dynamic‑pricing or matching changes. Reports should summarise pre/post earnings distributions, platform take rates, changes in unpaid waiting time, and deactivation patterns by cohort.
Include waiting time in guarantees. If guaranteed pay excludes waiting, effective hourly income can shrink significantly. Guarantees should reflect total engaged time, not only in‑order time.
Fast‑track dispute resolution. Deactivations and biometric failures should be routed to rapid human review with evidence logged and made available to the worker.

Short, high‑impact advice for riders

Track the essentials. Log offers, timestamps, distances (if possible), accept/decline, and final payout. Consistent records from many riders create the datasets researchers need.
Share and aggregate. Join groups (unions, Workers’ Observatory, Worker Info Exchange) that can pool anonymised records and collaborate with academic researchers for more robust analysis.
Escalate safety or biometric failures immediately. If you encounter deactivation or facial‑recognition issues after injury or illness, preserve timestamps and communications and request a manual review in writing.

Key takeaways, questions you’d want answered

Are riders’ pay declines proven to be caused by algorithms?

There is credible evidence that algorithmic pricing and matching can shift earnings, notably a University of Oxford preprint (arXiv:2506.15278; press release 23 June 2025) showing Uber drivers’ hourly earnings fell after a 2023 dynamic‑pricing change, but causal proof specific to food delivery platforms is not yet established. Riders’ logs and small experiments are suggestive; large‑scale independent audits are needed to isolate cause from correlation.
Do platforms deny responsibility?

Platforms state that algorithms are overseen by humans and point to guaranteed pay while on an order (Deliveroo has said its minimum hourly fee while on an order “rose by 3.8% this year”). Those statements are company claims; they do not on their own resolve whether systemic distributional harms exist or whether guarantees fully compensate for unpaid waiting time.
What legal tools are in play?

A pan‑European class action was filed in Amsterdam on 2 September 2026 (reported by The Guardian and AI Work Insights) alleging personalised algorithmic pay and data‑protection breaches. Plaintiffs will invoke GDPR provisions on automated decision‑making and access to personal data (see Articles 22 and 15). Courts will need technical evidence, model logs, matching rules and audits, to evaluate the claims.
What can regulators or companies do quickly to reduce harm?

Publish algorithmic impact assessments, provide machine‑readable worker data (offers, timestamps, payouts, cancellations), include waiting time in pay guarantees, and require independent audits for major pricing changes. These steps are practical, protect worker rights, and preserve legitimate IP through summarised reporting rather than raw source disclosure.
What should executives monitor right now?

Track these KPIs: average payout per order (and its trend), platform take rate (commission share), unpaid waiting time per hour, orders per hour adjusted for distance, and deactivation/error rates by cohort. Watch for divergent trends across cohorts (e.g., new vs experienced riders), unequal impacts often reveal hidden algorithmic effects.

Riders on the street report shrinking pay pockets. A large Oxford study of ride‑hailing documents a clear association between a 2023 pricing change and lower driver earnings. A major class action has taken these questions into the courts. For executives and policymakers the pragmatic question is clear: how to run algorithmic systems that improve service and efficiency without shifting the burden invisibly onto the people who deliver it. Transparency, auditable data and independent checks are not just regulatory niceties, they are the building blocks of trust and of a sustainable platform business model.

“There is so much infrastructure of knowledge and data that’s concealed…we’re working in the dark.”

, Cailean Gallagher, director, Workers’ Observatory