Back to timeline

Research · December 14, 2017

Placed by the contemporary primary publication. The exact event date is not known; its documented interval appears below.

What data work actually pays: 3.8 million tasks measured

On 14 December 2017 a preprint appeared that measured hourly pay on Mechanical Turk by recording the work rather than by asking: 2,676 workers, 3.8 million tasks. The median came to about $2 an hour, and only 4 per cent earned more than the US federal minimum of $7.25.

Why it matters

Until then low pay in data work rested on self-reports, and their spread - $1 to $6 an hour - left room for any conclusion. Here pay was computed from actual working sessions including unpaid time, which made the gap between what a requester pays and what a worker receives a measurable quantity.

The data came from the Crowd Workers Chrome plugin, used by workers on an opt-in basis, between September 2014 and January 2017. Requesters paid $11.58 an hour on average (median $4.57). Unpaid time eats the difference, and the authors name three sources of it: searching for tasks, working on tasks that are rejected, and working on tasks that are never submitted. The figure $2.11 needs stating precisely. It is not "the median": in the paper the per-worker median lies between $1.77 and $2.11 an hour depending on the method of counting, and $2.11 is the top of that range. The authors' own summary is about $2 an hour at the task level, with an estimate that 96 per cent of workers earn below the federal minimum.

Event record

Event date
December 14, 2017
Timeline date
Primary publication date
Verification
Sources gathered automatically · September 21, 2026
Lines
ID
evt-0517

The day of the first version of the preprint, per the arXiv submission history. The peer-reviewed version was accepted to CHI 2018, held in Montreal in April 2018.

Sources

Related events

Records that link to this one