Two research teams have documented how algorithmic pay finds the minimum a worker will accept and the US is the market where this has gone furthest
Len Sherman spent months analysing decade-long trip histories from three veteran Uber drivers in Florida and Texas who had collectively completed 50,000 rides. What he found, published in June 2026, is a precise account of how algorithmic pricing finds the floor of what a worker will accept.
In Uber's early years, those drivers kept around 80 to 85 cents of every dollar a passenger paid. After Uber cut driver pay ahead of its 2019 IPO, that fell to roughly 60 cents. After the company introduced upfront algorithmic pricing from mid-2022, pay stagnated while passenger fares continued to rise. By 2026, in many US cities, Uber's average take rate has passed 50%. The drivers are getting less than half of what passengers pay.
A decade of take rate changes
Uber's effective take rate, 2014–2026
From 15% in the early years to 50%-plus today. The 2019 jump came when Uber cut driver pay ahead of its IPO. The post-2022 surge followed the introduction of algorithmic upfront pricing.
Important: US line based on Sherman (Columbia Business School, June 2026) longitudinal case study of three veteran US drivers (50,000 trips). UK median data points from Oxford University peer-reviewed study (arXiv:2506.15278v1, ACM FAccT '25, June 2025). These are research findings, not Uber's own disclosed figures. Uber reports take rate differently, deducting insurance and third-party fees before calculation. Values are approximate; Sherman's study covers specific markets (Florida and Texas) and may not reflect all US cities.
Sherman, adjunct professor at Columbia Business School, is precise about the mechanism. He analysed 100 near-identical trips by a single driver on the same route - same vehicle, same driver, same 60-mile road, same service type. Uber's reported insurance and operating expenses on those trips varied from $13.75 to $50.00. A regression found that the risk profile of the trip - day of week, time of day, service type - was statistically insignificant in explaining that variance. What predicted it: the rider price and the driver's pay. Higher fares were associated with higher reported costs. Lower driver pay was also associated with higher costs. On trips where both applied simultaneously, the reported costs were highest.
The algorithm, in short, is not pricing risk. It is pricing workers.
Oxford confirmed it in the UK
Independently, researchers at Oxford University's Department of Computer Science spent two years analysing data from 258 UK Uber drivers covering 1.5 million trips between 2016 and 2024, obtained through data access requests under GDPR. Their peer-reviewed paper, lead by Associate Professor Reuben Binns and DPhil student Jake Stein, found the same structural pattern: the higher the passenger's fare, the less the driver earned per minute in absolute terms. After dynamic pricing was introduced in London in early 2023, driver pay fell from £22.20 to £19.06 in inflation-adjusted terms. Uber's median take rate rose from 25% to 29%. The surplus Uber extracted per driver working hour rose 38%.
Binns told the Financial Times: "To have that uncertainty hanging over you all the time... it turns work into a gamble, basically."
Veena Dubal, law professor at the University of California Irvine, has a name for what the Oxford and Columbia research documents: the "gamblification" of work - algorithmic decision-making that undermines economic stability by transforming the basic terms of how workers are paid. In this system, as Dubal has written, the house always wins.
The US context is specific
About 16% of Americans have worked for an app-based platform, according to Pew Research data cited by HCAMag's US coverage of the ILO negotiations. The regulatory picture remains fragmented - the legal treatment of gig workers rests on state-by-state rules where progress is uneven, and the US was one of the countries at the June 2026 ILO Geneva talks favouring a less prescriptive international framework.
The domestic picture is moving, if slowly. Senators Brian Schatz and Chris Murphy introduced the Empowering App-Based Workers Act in July 2025, which would require platform companies to disclose more information to workers about how pay is calculated. Massachusetts reached an agreement with Uber and Lyft requiring that rideshare drivers earn at least $32.50 per active hour - a regulation visible in the Sherman data, which found that pay distributions for Boston drivers were notably higher than in unregulated markets like Dallas and Tampa, where Uber's algorithms are free to probe for the lowest acceptable offer without constraint.
That contrast is instructive. In Dallas, roughly half of all trips offered pay below $1 per mile. In Boston, the mandated floor is visible in the distribution. What the algorithm does in the absence of a floor is exactly what the research documents: it finds the floor itself, dynamically, using behavioural data on what individual drivers have accepted before.
The question for HR leaders
The natural response is that this is a gig economy story, not an HR story. The workers are contractors. The company is a platform. The regulatory questions belong to legislatures, not people directors.
That response is getting harder to sustain.
The Oxford paper's description of Uber's CEO explaining the company's algorithmic direction to investors is worth reading carefully: "you've gone from just flat time and distance to now kind of point estimates for every single trip based on the driver... targeting of different trips to different drivers based on their preferences or based on behavioral patterns that they're showing us." Replace "trips" with "tasks" and "driver" with "employee" and you have a description of where workforce management software is heading in industries well beyond ride-hailing.
AI-driven scheduling, task allocation and performance management are already reshaping how work is distributed in warehousing, logistics, retail and healthcare. The same data infrastructure that enables Uber's algorithmic pricing - behavioural data on individual worker responses to varying offers, accumulated at scale, used to calibrate future allocations - is being built into workforce management systems that apply to employees, not contractors.
The Oxford researchers found that their predictive models, trained on pre-2023 Uber trip data, could not predict post-2023 driver earnings at all. The relationship between work done and pay received had been structurally severed. If you run a compensation system where employees increasingly cannot predict what their performance will yield, you have imported the gamblification problem into your own workforce, regardless of whether anyone calls it dynamic pricing.
Three things the research tells HR leaders
Transparency is a design choice, not a default. The Oxford study found that even with 1.5 million trips of backend data from Uber's own systems, calculating the take rate required inferring connections between database tables because Uber does not provide a trip ID. The opacity is not technical necessity. It is a decision. Any compensation or performance system that accumulates behavioural data on workers while withholding from those workers the information they would need to understand their own pay is making the same decision, for the same reasons.
The information asymmetry compounds over time. The Oxford team found that drivers with the lowest earnings post-dynamic pricing were those with wider distributions in their take rates and lower acceptance rates. The algorithm learns from individual behaviour. Workers who push back on low offers get fewer offers, or lower ones. Workers who accept readily train the system to offer them less. The asymmetry is self-reinforcing. The same dynamic is available to any sufficiently granular workforce management system - and the data to power it is accumulating in HR systems across every sector.
Regulation is coming, and the US is behind. ILO Convention No. 193 on platform economy standards was adopted in June 2026 with 406 votes in favour - the US voted against it. As Ogletree Deakins noted in its analysis for US employers, US-headquartered companies will still feel the convention's effect wherever they retain platform workers in countries that ratify it. Domestically, Massachusetts is the clearest example of what mandatory pay floor legislation does to algorithmic pricing: it creates a visible floor in the distribution where there was none before.
What to ask about your own systems
The Oxford and Columbia research does not argue that algorithmic workforce management is inherently wrong. The same infrastructure that enables exploitation could theoretically enable better, more predictable work allocation. The architecture is neutral. The incentives determine the outcome.
For HR leaders, the research suggests three questions worth asking about any algorithmic or AI-assisted compensation or scheduling system your organisation is deploying or considering.
Can the workers affected by it understand how their pay is determined? Not in general terms - in specific terms. Can a warehouse picker, a home care worker, an on-demand customer service agent reconstruct how a given offer was calculated? If not, you have the opacity problem.
Can they predict what a given amount of work will yield? The Oxford study found that after dynamic pricing, drivers' accumulated knowledge about when and where to work was no longer useful. Predictability is not a luxury. It is what allows people to plan a financial life around their work.
And is the system accumulating data on individual workers' acceptance behaviour? If it is - and most modern workforce management systems are - the question is what that data is used for. If the answer includes calibrating future offers to individual workers based on their past responses, you are running an experiment in finding each worker's floor. That is what Uber built. It took them a decade to get to 50%. The infrastructure to do it faster already exists.
