Gig platforms: efficient or unfair? Reading the Uber driver data

Between January 2015 and March 2017, each hour an Uber driver worked in the US cost riders $24.77 on average; after Uber's fees and vehicle costs, the driver kept $11.77 (Mishel, 2018, p. 20). If you sell through a delivery app, hire people through a platform or plan to build one, those two numbers ask a question: efficient, or unfair?

I read seven sources, and my answer is both, in two different places along the same flow of money. The figures below are American and dated; my library has no public source on platforms in Vietnam. The Uber driver economics case draws on the same sources.

Why are gig platforms seen as efficient?

Because they cut transaction costs, which Williamson called the economic counterpart of friction: the costs of planning, adapting and monitoring whether the work gets done (Williamson, 1981, p. 552–553). His theory rests on two assumptions: bounded rationality (nobody can write a contract covering every contingency) and opportunism (some people pursue self-interest with guile) (Williamson, 1981, p. 553–554). See the library's transaction cost economics framework.

Ferretti applies this to platforms. A cleaning firm once needed a management hierarchy and long contracts, because finding a new contractor for every job was too costly. A platform automates the monitoring, opens a far larger pool of workers and uses ratings to build trust between strangers (Ferretti, 2020, §1.2).

My inference: a car and a driver's hours are not tied to one rider, so asset specificity is low, and Williamson writes that unspecialised assets pose few hazards because buyers can turn to other sources (Williamson, 1981, p. 555). But his paper asks which way of organising economises most; who keeps the saving lies outside its main question (my reading).

How much of each dollar does the driver keep?

About 47.5 cents, after Uber's fees and vehicle costs, in US data for January 2015 to March 2017: Uber kept about 33.2 cents and vehicle costs after tax took about 19.3 cents (my calculation, 4.78 ÷ 24.77) (Mishel, 2018, p. 20). Lawrence Mishel (EPI, 15 May 2018) took Uber's data on fares, hours and trips and subtracted item by item (Mishel, 2018, p. 10).

Source: (Mishel, 2018, p. 10, 19); subtracting line by line may differ by a cent through rounding.

The $9.21 ("W-2 equivalent wage") sits beside $32.06 average hourly compensation in the private sector, benefits included (Mishel, 2018, p. 2), at about the 10th percentile of wages (Mishel, 2018, p. 13), and below the January 2018 minimum wage in 13 of 20 large markets (Mishel, 2018, p. 2).

Limits, by the report's own account: it compares an average with each place's minimum wage, with no distribution across drivers (Mishel, 2018, p. 14), and drivers are contractors, not entitled to the minimum wage (Mishel, 2018, p. 13). The benefits package fits a full-time driver who relies on the job (Mishel, 2018, p. 11). Excluding insurance costs, which Uber covers, lifts the equivalent wage by $0.82 to $10.03, though that cover has gaps (Mishel, 2018, p. 16). Mishel says in an endnote that he considers drivers employees (Mishel, 2018, p. 25); EPI is a policy research organisation.

Where do costs and risks go when transactions get cheaper?

Partly onto workers. Moradi and Levy ask whether all this monitoring and analysis makes workplaces more efficient. Their answer is "maybe", but not necessarily by removing waste: the technologies can move burdens from the firm onto its comparatively less powerful workers (Moradi and Levy, 2020, p. 278–279).

For Uber they write that drivers are paid only while carrying a passenger. Time spent cruising for a request, driving to the pickup or cleaning the car to protect a rating goes unpaid, and the costs fall on the driver (Moradi and Levy, 2020, p. 280–281). That is one of four mechanisms in the risk-shifting framework; the chapter applies one more to Uber (productivity incentives) and illustrates the other two with examples outside Uber.

Ferretti adds another form: dropping long-term contracts moves the risk of too few clients onto contractors (Ferretti, 2020, §2.2). In the table above, the two items drivers bear themselves are vehicle costs of $4.78 and the extra payroll tax plus assumed benefits of $2.56: $7.34 an hour, about 29.6 per cent of what riders pay.

Does app design lengthen driving hours?

According to The New York Times (Scheiber, 2 April 2017), yes for one feature at least: forward dispatch, which sends the next fare before the current one ends and which it says keeps drivers on the road substantially longer in busy periods, without giving Uber figures. It draws on interviews with several dozen current and former Uber officials, drivers and social scientists (Scheiber, 2017). One driver trying to log off saw a message that he was $10 short of $330 net earnings, with "keep driving" already highlighted. Uber officials say drivers complained they could not stop, so it added a pause button, which the article says must be pressed again for every new passenger (Scheiber, 2017).

Uber replied that a driver can stop "at the tap of a button" and that the decision is entirely theirs; its research director, Jonathan Hall, argued that exploiting psychological tics has at most a marginal effect, "icing on the cake" (Scheiber, 2017). I do not know whether these features still work as described in 2017. Moradi and Levy place such practices under productivity incentives, not a priori bad for workers in their view, though in many contexts they erode trust and autonomy (Moradi and Levy, 2020, p. 284–285).

When should you use the efficiency lens, and when the fairness lens?

Use the efficiency lens to choose how to organise one activity; use the fairness lens to set prices, classify contracts and design app defaults. The two conclusions do not exclude each other.

Williamson writes that autonomous trading works when effective competition can be presumed both before and after contracts are signed (Williamson, 1981, p. 554). Ferretti names two things that weaken this condition for large platforms: monopolistic tendencies, and reputation data that cannot move to another platform, so drivers find it hard to leave (Ferretti, 2020, §3.2).

One gap remains. Coyle writes that the value platforms create appears to go largely to investors and consumers, then calls this impressionistic (Coyle, 2017). Of the seven sources I read, none measures what riders save: the "efficient" half rests on theory and description, the "unfair" half on EPI's measurements plus one investigative report.

If you want a fix, which kind?

Ferretti, from a Rawlsian liberal-egalitarian standpoint, separates two strategies. Mitigating: tolerate any kind of organisation running platforms and intervene only through redistributive taxation, retraining and protections that do not depend on contract type (Ferretti, 2020, §2). Organisational: do all that and also change the kind of organisation in charge, by supporting cooperatives, breaking monopolies and requiring data sharing so that drivers' reputations can move (Ferretti, 2020, §3). The library's two-strategies framework summarises the two columns; platform cooperativism is the main tool of the second, and Schor argues that the potential of platforms needs democratised ownership and governance (Schor, 2014).

Ferretti does not declare the second route the winner. He cites Eva, a taxi cooperative in Québec since 2017, charging drivers 15 per cent commission against Uber's 25 per cent, from a January 2019 press report (Ferretti, 2020, §3.2). But justifying heavier intervention requires showing that mitigation is not enough, while tighter rules and lower investor returns could reduce income opportunities for low-income people (Ferretti, 2020, §4).

For your own platform, four things to check: which step of the table above your published earnings figure sits on; who bears vehicle or equipment costs, benefits and waiting time; where the app's defaults push working hours; and whether workers can switch platform and take their ratings with them (Mishel, 2018, p. 9; Moradi and Levy, 2020, p. 280–281; Scheiber, 2017). These are my four questions, not the authors' recommendations.

Going deeper: seven sources I read and what I have not checked

Williamson (1981), American Journal of Sociology 87(3), pp. 548–577: full text from a JSTOR scan, read through optical character recognition; page numbers are the printed ones. The paper does not discuss digital platforms; every application to platforms is Ferretti's or mine.

Ferretti (2020), Journal of Social Philosophy 51(1), pp. 54–79: full text of a 21-page author manuscript without journal pagination, so I cite by section (§). The file does not say which version it is and I could not check it against the published text. It is political philosophy, not empirical research; the author himself uses EPI's figures and Scheiber's article as evidence.

Mishel (2018), EPI report of 15 May 2018: full text of the web version and the PDF (28 numbered pages plus a cover; page numbers are the printed ones). I have not read Cook and co-authors (2018), Zoepf and co-authors (2018) or Hall (2018). Moradi and Levy (2020), pp. 270–288: full text; the Uber passages are a few paragraphs. Scheiber (2017): full article via an archived copy; it is reporting based on interviews, not measurement. Coyle (2017): accepted manuscript, so no page numbers. Schor (2014): full text.