What Happened to MTurk? An Autopsy of Amazon Mechanical Turk.

MTurk did not just get quieter. Amazon shut it down.

The old advice was to keep MTurk in a browser tab, learn which requesters paid fairly, and use it when better work was unavailable. That advice has an expiration date. Amazon’s current website states that Mechanical Turk permanently closed on September 30, 2026.

For people who built routines around HITs, the loss is more than a missing website. Their familiarity with requesters, task formats, qualifications, and workflow tools represented accumulated effort. That knowledge helped them navigate a marketplace that was often difficult to judge from the outside. It does not automatically transfer to another service.

An honest autopsy needs to separate two questions. We know when Amazon ended the service. Understanding why the model became less attractive requires examining its economics and the research published before the shutdown. Those findings illuminate weaknesses; they are not a transcript of Amazon’s internal decision.

Official status: Amazon Mechanical Turk.

A simple transaction hid a complicated job.

Mechanical Turk organized work into Human Intelligence Tasks, or HITs. A requester offered a task and a reward. A worker accepted it, submitted an answer, and waited for the requester’s approval. The format could accommodate research surveys, classification, labeling, and other small pieces of work.

The simplicity was appealing. A worker could take available tasks without pitching a client for every assignment. A requester could distribute a job across many people rather than recruit each person individually. The difficulty was everything surrounding the transaction: finding worthwhile tasks, assessing instructions, managing qualification barriers, and deciding whether a requester was worth trusting.

This is where an apparently flexible marketplace can become expensive for both sides. Workers spend time finding work. Requesters spend time deciding which submissions are usable. Neither cost is fully described by the reward printed beside a task.

The task reward was not the worker’s hourly wage.

A 2018 study by Kotaro Hara and colleagues analyzed records from 2,676 workers performing 3.8 million tasks. It reported a median hourly wage of roughly $2, with only 4% earning more than $7.25 per hour. The underlying records came from September 2014 through January 2017, so these are historical findings, not a measurement of earnings immediately before closure.

The researchers also examined how unpaid effort affected the calculation, including searching for tasks and work that was rejected or not submitted. That is the enduring lesson: a task’s apparent rate can look reasonable while the surrounding workflow produces a much weaker return.

For a worker, the practical question was whether a whole session paid adequately. Time spent checking the queue and learning a task had to come from somewhere. When useful assignments become harder to find, the cost of checking rises even if the advertised rewards stay unchanged.

Source: Hara and colleagues, A Data-Driven Analysis of Workers’ Earnings on Amazon Mechanical Turk (2018) .

Researchers needed usable responses, not cheap submissions.

A 2023 study by Benjamin Douglas, Patrick Ewell, and Markus Brauer compared MTurk, Prolific, CloudResearch, Qualtrics, and a university participant pool. Prolific and the CloudResearch sample performed better than MTurk on several measures, including attention, meaningful responses, and following instructions.

The authors calculated cost per high-quality respondent. In that study, the figures were $1.90 for Prolific, $2.00 for CloudResearch, and $4.36 for MTurk. These were study-specific results, not universal prices or guarantees. The CloudResearch comparison concerned its screened MTurk recruitment approach; it should not be presented as a direct measurement of today’s Connect panel.

That comparison matters because it changes the buyer’s calculation. A lower initial price is less attractive if more responses must be discarded. Researchers also have to investigate suspicious records, decide exclusion rules, and replace participants. A marketplace can lose its cost advantage without raising its advertised fees.

Source: Douglas, Ewell, and Brauer, Data quality in online human-subjects research (2023) .

Quality failures created costs for honest workers, too.

Data-quality problems were not merely theoretical. In a paper published in Perspectives on Psychological Science in 2024, Margaret Webb and June Tangney described an MTurk study in which their assessment left only 14 of 529 participants as valid. That is one severe case, not an estimate of the proportion of legitimate workers across MTurk.

Fraud, inattentiveness, misunderstood instructions, and unsuitable respondents are also different problems. Labeling every questionable answer a bot obscures the work required to distinguish them. A real person can produce unusable data; an honest participant can also be wrongly suspected.

The concern for workers is the collective effect. If a requester becomes wary of a recruitment channel, reliable participants can lose opportunities alongside unreliable ones. Additional screening may protect a study, but it can also increase legitimate workers’ unpaid effort. This is a plausible mechanism for deterioration, rather than a measured account of every requester’s decision.

Source: Webb and Tangney, Too Good to Be True: Bots and Bad Data From Mechanical Turk (2024) .

Quality control became something buyers could shop for.

MTurk was a broad marketplace. Research participants, software testers, and specialist contributors have different recruitment and quality requirements. A service designed around a particular use can make those requirements easier to manage than a general task board.

CloudResearch’s history illustrates the shift. Its MTurk tools helped researchers screen and manage participants within Amazon’s marketplace. Its current materials describe moving toward its own research products and retiring the MTurk Toolkit alongside the marketplace.

CloudResearch sells an alternative, so its explanation has a commercial interest and should be read accordingly. Still, the trajectory is instructive: a business that helped make a marketplace usable developed a separate destination for that use. The quality controls became part of a competing product rather than only an additional layer on the original marketplace.

My interpretation is that MTurk’s broad format became less compelling where buyers wanted a managed research or specialist-work process. The available sources do not establish how much business moved, when each requester left, or how much that influenced Amazon’s closure decision.

Provider perspective: CloudResearch’s MTurk transition information .

Did AI kill MTurk? That is too simple.

It is tempting to explain the shutdown as machines replacing the people who once handled small tasks. Automation can change the economics of particular assignments, but a believable explanation is not the same as evidence of Amazon’s reasoning.

AI also complicates research in another direction: a requester seeking human responses may receive machine-assisted submissions instead. That makes verification more important. Meanwhile, evaluating model outputs and applying specialist judgment still involve human contributions.

The studies discussed here document worker economics and research-quality concerns. They do not establish that generative AI caused MTurk’s closure. It would be equally misleading to conclude that human work disappeared because one marketplace ended.

The more useful question is which work requires authentic participation, relevant expertise, or accountable judgment, and how a platform verifies those qualities. MTurk’s closure does not answer that question for every type of online work.

The model put too much of the evaluation work on its users.

MTurk’s weaknesses can be understood from both ends. Workers needed to identify worthwhile requesters and count the effort surrounding each task. Researchers needed to identify trustworthy respondents and assess the quality of what they received. The marketplace connected the parties, but the research shows why connection alone was not enough.

My diagnosis is a mismatch between the convenience of distributing tasks and the burden of evaluating the transaction. The worker needed useful net earnings; the buyer needed useful output. A high task count or a low submission price did not necessarily deliver either outcome.

That does not mean every requester paid badly or every worker produced poor work. It means the platform’s general promise was less informative than the particular relationship, task, and safeguards. Workers who had reliable arrangements could experience a very different marketplace from newcomers.

Amazon’s shutdown is confirmed. The contribution of each weakness to that decision remains uncertain. The documented problems are enough to explain why the old advice to simply join and start grinding was inadequate long before the final closure.

Check payment preferences and save your records.

Amazon’s closure FAQ still matters if you have outstanding earnings or records to retrieve. The dates below were checked on October 13, 2026. Follow Amazon’s current instructions for your account.

Approved earnings follow the existing payment schedule. Save relevant records while access remains available, and use Amazon’s support route for payment problems. GigReviewer cannot access your account or release funds.

Choose your next platform by task type, not resemblance to MTurk.

A research study, a recorded usability test, and a specialist AI assignment involve different kinds of participation. Looking for a single replacement can conceal those differences and recreate the habit of checking one dashboard for everything.

If research was your priority, start by comparing Prolific and CloudResearch Connect . For recorded feedback, investigate UserTesting . For specialist AI work, check the actual qualifications and assignment terms rather than assuming it is interchangeable with short survey work.

These are categories to investigate, not promises of acceptance or equal replacement income. Start with a small set, include qualification effort in your records, and compare payments received over a meaningful period.

Sources behind this autopsy.

The shutdown and account deadlines come from Amazon. The earnings and data-quality findings come from the named studies. Explanations about marketplace incentives and competition are GigReviewer’s interpretation, not claims about Amazon’s internal deliberations.

• Amazon Mechanical Turk: closure notice .
• Amazon: worker, requester, and record-access FAQs .
• Hara and colleagues (2018): worker earnings analysis .
• Douglas, Ewell, and Brauer (2023): five-platform data-quality comparison .
• Webb and Tangney (2024): MTurk data-quality case study .
• CloudResearch: transition information and provider perspective .

Research findings reflect their samples and collection periods. They should not be treated as universal rates, current earnings estimates, or evidence that every worker or requester had the same experience.