Amazon Mechanical Turk (MTurk) is a global marketplace where requesters publish human intelligence tasks and workers complete them on demand. Available research and Amazon documentation describe a service with a wide range of reported workforce estimates, substantial differences in worker activity and earnings, and detailed controls for measuring submitted work.
Contents
- Workforce size and marketplace reach
- Worker activity and retention
- Geography and language research
- Earnings, wages, and completed HITs
- Time spent, returns, and wasted effort
- Fees, task volume, and requester controls
Workforce size and marketplace reach
MTurk’s reported scale varies sharply by definition and measurement period. In its 2008 launch announcement, Amazon said the service could reach hundreds of thousands of workers. The same Amazon Mechanical Turk launch press release said the workforce came from more than 100 countries. These were launch-era statements about potential reach, not a contemporaneous census of active workers.
A later source, Demographics and Dynamics of Mechanical Turk Workers, found more than 100,000 workers available on MTurk and more than 2,000 active workers at any given time. The study’s survey ran for 28 months, collected more than 85,000 responses, and included approximately 40,000 unique participants. Those figures describe observed availability and participation in that study, rather than a permanent platform total.
Another source, Using Amazon’s Mechanical Turk for Crowdsourcing Research, reports that Amazon had estimated the active worker pool as low as 5,050 workers and as high as 500,000 workers. The range is important: it shows why MTurk statistics should identify whether they refer to registered, available, active, or study-observed workers.
Amazon describes MTurk as a 24/7 global distributed workforce marketplace. Its product details say requesters can engage a distributed workforce to perform tasks on demand. In practical terms, the platform combines continuous availability with work that can be divided into individual human intelligence tasks, commonly called HITs.
Worker activity and retention
The available evidence suggests that participation is uneven. The crowdsourcing research source reports that active workers reportedly work at least two days per week. Within that group, some active workers spend fewer than five hours per week on tasks, while other active workers spend 21 hours per week or more. These are different activity levels within the reported active population, not a single average workload.
The 28-month study of worker demographics also estimated a worker half-life of around 12 to 18 months. In that context, half-life describes the estimated duration over which a cohort of workers remains active under the study’s measurement approach. The researchers found that new worker arrivals roughly balanced worker departures. That pattern points to continuing turnover alongside a replenished participant pool.
Task completion is also limited at the individual level. Introduction to Amazon Mechanical Turk states that a worker can do each task only one time. For a requester, that rule can help prevent the same worker from submitting multiple responses to one task. For analysis, it means that the number of assignments and the number of unique workers are separate quantities.
Geography and language research
The geographic figures available here come from The Language Demographics of Amazon Mechanical Turk. That study surveyed workers for 100 languages and geolocated 4,983 workers by IP address. It found 60 workers who were located in more than one country during the study, while 238 workers could not be geolocated.
The country counts in that sample were:
| Country | Workers in the sample |
|---|---|
| India | 1,998 |
| United States | 866 |
| Philippines | 142 |
| Egypt | 25 |
| Russia | 10 |
| Sri Lanka | 4 |
These counts are study measurements, not current platform shares. They also should not be treated as a complete global distribution: the geolocation process left 238 workers unresolved, and 60 workers appeared in more than one country during the study. The language-demographics research is therefore most useful for showing the geographic diversity and measurement limits of MTurk participation.
The launch-era figure of more than 100 countries and the later language study’s coverage of 100 languages describe different dimensions. One concerns the countries from which Amazon said workers could come; the other concerns the languages included in a research survey. They should not be combined into one geographic percentage.
Earnings, wages, and completed HITs
The earnings evidence comes from A Data-Driven Analysis of Workers’ Earnings on Amazon Mechanical Turk. The study observed 2,676 workers and recorded 3.8 million tasks. Across that dataset, the average worker earned $95.96, while the median worker earned $11.90.
Hourly results were also highly uneven. Using the study’s interval-based method, the median hourly wage was $1.77 per hour and the mean hourly wage was $3.13 per hour. The interval-based wage standard deviation was 25.5. Only 4.2% of workers made more than minimum wage in that interval-based analysis.
The mean and median tell different stories. A mean of $3.13 per hour sits above the $1.77 median because higher observations can pull the average upward. The reported 25.5 standard deviation reinforces that the wage distribution was widely dispersed under the study’s method. These results should remain tied to the study’s dataset and measurement approach rather than being presented as a current MTurk wage forecast.
The same study found a very broad range of completed HIT counts, from 1 to 94,608. The median worker completed 12 HITs, while the mean worker completed 1,512.8 HITs. The HIT-count standard deviation was 9,586.8. As with earnings, the gap between the median and mean indicates that a relatively small number of very high-volume workers strongly affects the average.
| Measure | Reported result |
|---|---|
| Workers observed | 2,676 |
| Tasks recorded | 3.8 million |
| Average earnings | $95.96 |
| Median earnings | $11.90 |
| Median hourly wage | $1.77/hour |
| Mean hourly wage | $3.13/hour |
| Workers above minimum wage | 4.2% |
| Median HITs completed | 12 |
| Mean HITs completed | 1,512.8 |
Time spent, returns, and wasted effort
The earnings study recorded substantial differences in time spent. Workers spent 17.2 hours on average on the HITs in the dataset, while the median time spent was 0.9 hours. This contrast follows the same pattern as HIT counts: a smaller group of workers with much higher activity increases the average above the median.
Workers returned 26.5% of HITs on average. A returned HIT is not the same as an approved assignment or a rejected assignment; it represents work that workers returned under the study’s reporting. That distinction matters when interpreting completion and payment statistics.
At the study’s $1.77-per-hour wage estimate, workers wasted $30.44 on average. This figure is a study-specific result associated with the reported wage estimate and dataset. It should not be generalized to all MTurk work, all task types, or later periods without additional evidence.
Together, the time and return figures show why task totals alone do not describe worker experience. A dataset can contain millions of recorded tasks while the typical worker contributes far fewer tasks and less time than the mean suggests. Median values are especially useful when participation is concentrated among a small number of high-volume workers.
Fees, task volume, and requester controls
The crowdsourcing research source reports that Amazon charges fees of 20% to 40% on task payments, depending on the job type. Academic surveys with more than 10 respondents incur a 40% fee according to the same source. The applicable fee therefore depends on the task arrangement, and a requester’s worker-payment budget is not necessarily the full platform cost.
Amazon’s Mechanical Turk help states that the maximum number of HITs in a batch is 250,000. The Sandbox environment supports a maximum of 500 HITs per batch. Requesters can create thousands of HITs at a time; the Mechanical Turk Requester Website Guide describes a workflow in which they define a template, upload input data, and publish the HITs.
Requesters can specify up to five worker requirements for a project, according to Create an Amazon Mechanical Turk project. These requirements provide a way to narrow eligibility before work is assigned, while the one-time-per-task rule limits repeat completion by an individual worker.
Payment depends on requester approval. Amazon’s help documentation says workers are paid when requesters approve submitted work. If a requester rejects work, the worker is not paid. The platform also exposes approval-related measures: View Worker statistics says requesters can view a worker’s approval rating and see assignments approved over the last seven days, the last 30 days, or all time.
MTurk’s Approval Rate qualification is calculated as the percentage of assignments submitted that were approved, according to Mechanical Turk concepts. The same documentation identifies Assignments Approved as a worker qualification metric. Worker block status has three states—Never Blocked, Blocked, and Unblocked—according to the worker-statistics documentation.
Amazon’s help materials also describe MTurk Masters as awards for workers who demonstrate excellence across a wide range of HITs, with Masters maintained through ongoing statistical monitoring. Finally, requesters can deposit funds into their MTurk account at any time and withdraw funds at any time, according to the concepts documentation. These account controls, approval measures, and task limits define the operational side of the marketplace alongside its workforce and earnings statistics.