What Mechanical Turk task logs showed about hourly earnings

In 2018, Kotaro Hara, Abigail Adams, Kristy Milland, Saiph Savage, Chris Callison-Burch and Jeffrey P. Bigham published “A Data-Driven Analysis of Workers’ Earnings on Amazon Mechanical Turk” in the proceedings of the CHI Conference on Human Factors in Computing Systems. The paper does not ask whether platform task work is worthwhile for any particular person. It asks what effective hourly earnings look like when task rewards are divided by the time workers spend finding, accepting and completing those tasks.

The authors assembled a dataset of task listings and worker activity on Amazon Mechanical Turk (MTurk), covering thousands of workers and millions of tasks. Their central estimate is distributional rather than a single slogan number. For the workers in their sample, median effective hourly earnings were about two US dollars per hour when unpaid time is included. Only a small share of workers in the sample cleared figures near the US federal minimum wage on that measure. Means sit higher than medians, which is expected in a skewed earnings distribution; the paper’s emphasis is on the mass of the distribution rather than on a single headline average.

Two design choices drive how the figure should be read. First, the clock includes time spent searching for tasks and waiting between tasks, not only time with a HIT open. Excluding that unpaid interval raises effective hourly rates, and the authors report sensitivity to that modelling choice. Second, the sample is drawn from workers visible to the researchers’ measurement approach on MTurk in the study period. It is not a census of all crowdwork, and it is not a sample of freelancers on other platforms.

What the paper does not claim

The authors do not present the median as a wage floor that every new entrant will meet, and they do not treat MTurk as interchangeable with ride-hail, content platforms or remote software contracting. Task composition, requester quality, geographic restrictions and rejection rates all affect realised pay; the paper documents a measured distribution under stated conditions rather than a universal rate.

Methodological debate around crowdwork pay estimates often turns on how unpaid time is counted and on whether workers who stay on the platform are selected for higher-yielding strategies. Hara and co-authors address unpaid time explicitly. Selection into continued platform work remains a limit that any observational scrape inherits: workers who leave after a short period are under-represented in sustained activity logs, and workers who specialise in higher-paying requester niches may look different from the median of a broad scrape.

Later commentary on the paper has sometimes compressed the finding into a single phrase about “two dollars an hour.” The published work is more careful. It reports a distribution, discusses unpaid search, and situates the result inside one platform’s task market. Reading the median without those qualifications turns a measured sample statistic into a general claim the authors did not make.

A further limit is temporal. Platform interfaces, requester behaviour and the mix of available tasks change. A distribution measured in one window of platform history is evidence about that window. It is not automatic evidence about a later year, a different country filter, or a different microtask marketplace structured with other payment rules.

For an editorial desk concerned with online monetization research, the useful comparison is not between MTurk and an abstract idea of side income. It is between what this design measures — rewards divided by recorded time on one microtask platform — and what other papers measure when they study utilization in ride-hail markets or vacancy counts on remote labour platforms. Those are different objects. Keeping them separate is the point of the note.

The durable contribution of Hara et al. for this publication is therefore procedural as much as numerical: name the unit (effective hourly earnings including unpaid search), name the population (workers observed on MTurk in the study design), and keep the distributional shape attached to the figure. Without those attachments, a median becomes a slogan. With them, it remains what the authors published — a sample statistic about a defined labour process.