Mechanical Turk Closes - AI Ate Its Own Training Data
Amazon will stop accepting new customers for Mechanical Turk on July 30, 2026, retiring the crowdsourcing platform that spent 21 years training the AI models now replacing it.
7 min read
Amazon has placed Mechanical Turk on its "Services in Maintenance" list and will stop accepting new customers on July 30, 2026. Existing customers can keep using the platform, but no new features are coming and the door is closed to anyone who hasn't already signed up. After 21 years, the service that crowdsourced human intelligence to train machine learning systems is being quietly shelved by the machines it helped build.
TL;DR
- Amazon Mechanical Turk closes to new customers July 30, 2026 after 21 years
- 33-46% of workers were already using LLMs to complete tasks, per a 2023 study
- Active worker pool had shrunk to ~100K despite marketing claims of 500K+
- AWS is pointing customers to SageMaker Ground Truth and third-party vendors instead
- The platform that trained early AI is being retired because AI no longer needs it
What Mechanical Turk Was
Crowdsourcing micro-tasks at scale
Amazon launched Mechanical Turk in November 2005. The name comes from a 18th-century chess-playing automaton that turned out to have a human hidden inside - a commentary on the illusion of machine intelligence that was meant to be playful. Workers on the platform completed what Amazon called Human Intelligence Tasks (HITs): labeling images, transcribing audio, confirming addresses, moderating content, answering survey questions. Each HIT paid a few cents.
At its peak, the platform claimed more than 500,000 available workers. Independent research consistently put the active pool closer to 100,000. Most workers earned $2 to $6 per hour. The most experienced, who had mastered the platform's quirks and could filter for the best-paying tasks, topped out around $8 to $15 per hour on good days.
The 2018 AI pivot
In 2018, AWS reframed Mechanical Turk around AI training data. The company integrated it with SageMaker so that machine learning teams could use MTurk workers to review and label training datasets directly from their AWS pipelines. This was the platform's most commercially significant moment: it wasn't just a place for academic surveys anymore, it was infrastructure for the AI industry's data supply chain.
For several years it worked. Companies building image recognition models, natural language classifiers, and content moderation systems could spin up a HIT campaign, pay per annotation, and get labeled data back within hours.
Crowdsourced labor platforms like MTurk relied on distributed workers completing small digital tasks for micro-payments.
Source: pexels.com
Under the Hood
The HIT model
Requesters posted tasks as HITs with a reward amount, a deadline, and qualification filters. Workers browsed available HITs, accepted them, completed the work in a web interface, and submitted for approval. Requesters could approve or reject submissions. The money moved through Amazon Payments.
The API let engineering teams automate the whole process:
import boto3
client = boto3.client('mturk', region_name='us-east-1')
response = client.create_hit(
MaxAssignments=10,
AutoApprovalDelayInSeconds=604800,
LifetimeInSeconds=86400,
AssignmentDurationInSeconds=3600,
Reward='0.05',
Title='Image Classification Task',
Keywords='image, classification, labels, AI training',
Description='Look at an image and select the correct label from the list.',
Question='<HTMLQuestion xmlns="http://...">...</HTMLQuestion>',
QualificationRequirements=[]
)
hit_id = response['HIT']['HITId']
That Reward='0.05' is not a typo. Five cents per task. At that rate, a worker completing 60 HITs per hour earned $3. Amazon took a 20% commission on top of the reward amount, so requesters paid six cents per $0.05 task.
The infrastructure layer
For requesters, the value proposition was simple: structured, on-demand human judgment with programmatic access and a global workforce. For anyone building an AI training pipeline in 2018-2022, MTurk was often the first call. The data labeling tools that came after it - Scale AI, Labelbox, Remotasks - largely learned from what worked and didn't work here.
The Collapse in Three Signals
LLMs eating the labelers
The most structurally damaging problem wasn't fraud - it was that MTurk workers started using LLMs to complete their tasks. A 2023 study published on arXiv estimated that 33 to 46 percent of crowd workers on the platform were using large language models to automate abstract summarization tasks. Researchers detected this through a combination of keystroke logging and synthetic text classifiers.
The irony is brutal: the platform tasked with producing human-labeled data to train AI was producing AI-produced data labeled as human. Datasets built this way aren't useless, but they aren't what requesters were paying for, and the quality assumptions behind benchmark comparisons built on MTurk annotations quietly broke down.
Worker account closures
Users on the platform's subreddit reported Amazon closing worker accounts with minimal notice and no detailed explanation. This wasn't a handful of edge cases - the pattern was widespread enough to accelerate the worker exodus, which further reduced the quality of available tasks for requesters. Fewer good workers meant more tasks sat uncompleted or were picked up by low-quality workers.
A shrinking, unreliable pool
By the time Amazon made the maintenance announcement, the platform had lost both sides of its marketplace. Requesters had moved to alternatives with better quality controls, managed workforces, and domain specialization. Workers had migrated to platforms with higher per-task rates or better protections. The active pool of workers may have been well under 100,000 by 2025, against a task volume that had also dropped sharply.
Modern AI data labeling now uses managed vendor workforces and automated quality pipelines rather than open crowdsourcing.
Source: pexels.com
What AWS Offers Instead
AWS isn't leaving the data labeling market - it's just consolidating it into products with more control. SageMaker Ground Truth supports three workforce options: a private team (your own employees), a vendor workforce (pre-vetted third-party annotation companies), or MTurk for generic tasks. After July 30, the MTurk option in that list becomes less useful for anyone who hasn't already registered.
| Platform | Workforce Model | Cost Model | Best For |
|---|---|---|---|
| Amazon MTurk | Open crowdsource (~100K workers) | $0.01-$2 per task | Closing to new customers July 30 |
| SageMaker Ground Truth | Private + vendor + MTurk | Custom per project | AWS-native ML pipelines |
| Scale AI | Managed professional vendors | Enterprise contract | High-accuracy labeled datasets |
| Labelbox | Private or vendor workforce | Per-seat or per-label | Annotation workflow management |
| Remotasks | 200K+ specialized contractors | $3-8/hr equivalent | AI training, content moderation |
The market has moved toward managed vendors and private workforces for anything where quality matters. Open crowdsourcing at MTurk's price points is where you go when you need volume and can tolerate noise - a positioning that became harder to defend once LLMs could produce plausible-looking labels faster and cheaper than any human would for $0.05.
Where It Falls Short
Calling this a natural death undersells what was lost. MTurk's open model meant anyone could spin up a labeling job without a sales call, a minimum spend commitment, or a vendor contract. Academic researchers used it for surveys. Independent developers used it to bootstrap training datasets without enterprise budgets. That access point is gone now, and nothing available today replaces it at the same price and accessibility.
SageMaker Ground Truth's vendor workforce requires working through AWS-vetted annotation companies. Scale AI and Labelbox are enterprise products. Remotasks has its own application process. The gig-economy, self-serve, pay-as-you-go model that made MTurk useful for scrappy projects doesn't have a direct successor.
The platform's retirement also marks the end of a specific research infrastructure. Thousands of NLP benchmarks, psychology studies, and HCI experiments were run on MTurk because it offered fast, cheap, and varied respondents. That entire methodological tradition now needs to find a new home - one that doesn't assume $0.05 will get a meaningful answer.
Amazon gave no official reason for the maintenance decision. The most accurate summary: a platform built on the premise that human judgment was cheaper and more reliable than machines couldn't survive the moment that premise stopped being true. The AI training data wall is real, and MTurk was one answer to that scarcity for over a decade. Its replacement, in practice, is synthetic data and LLM-generated labels - the same technology that made the human labels unreliable in the first place. As of July 30, new teams will have to start somewhere else.
Sources:
- Amazon will stop accepting new customers for Mechanical Turk - TechCrunch
- Amazon's Mechanical Turk to stop accepting new customers - The Register
- Artificial Artificial Artificial Intelligence: Crowd Workers Widely Use Large Language Models for Text Production Tasks - arXiv 2023
- Amazon winds down Mechanical Turk - Shopifreaks
