⏱️ Lectura: 11 min
Amazon Mechanical Turk, the microtask marketplace that helped annotate the images behind ImageNet and much of modern machine learning, will shut down on September 30, 2026. The company confirmed it on the service’s own page, without announcing a direct replacement.
📑 En este artículo
The news marks the end of a platform that for more than two decades connected businesses and developers (Requesters) with a global army of remote workers (Workers) willing to solve tasks that machines still couldn’t handle on their own.
TL;DR
- Amazon confirmed the shutdown of Mechanical Turk for September 30, 2026, without giving detailed reasons.
- MTurk launched in 2005 and has spent 21 years connecting Requesters with Workers for remote microtasks.
- The name comes from the 1770 chess-playing automaton that hid a human operator inside.
- Much of the manual labeling behind ImageNet was done by Mechanical Turk Workers.
- The Allen Institute for AI (AI2) used MTurk to build common-sense datasets for its models.
- Food Genius (US Foods Data Science) used MTurk to extract menu data and spot trends.
- Amazon SageMaker Ground Truth loses MTurk as one of its three human workforce options.
- Amazon did not announce a direct replacement service and points users to its FAQ page to prepare.
What Happened
Amazon posted a brief, direct notice on mturk.com: “Mechanical Turk will close September 30, 2026.” The notice adds that the company “regularly evaluates its programs, tools, and services and makes adjustments based on those evaluations,” and that after that evaluation it decided to permanently shut down Mechanical Turk.
Amazon did not publicly explain the specific reasons for the shutdown or announce a substitute service. It did direct both Workers and Requesters currently using the platform to an FAQ page to prepare for the closure, without yet specifying what happens to in-progress tasks (HITs) or to pending payment funds in Requester accounts.
The notice is deliberately terse: there are no figures on active users, no processed task volume, no detailed business reasons. In that sense, it’s a quiet shutdown for a service that was once a pioneer.
Context and History: From an 18th-Century Automaton to Modern AI
Amazon launched Mechanical Turk in 2005 as a crowdsourcing marketplace that let businesses and developers outsource processes and jobs to a distributed workforce that carried out tasks virtually. The name is a historical wink: The Turk was a supposed chess-playing automaton built in 1770 that actually hid a human operator inside the cabinet, passing itself off as an intelligent machine. Amazon flipped the joke: it named a service where, behind an interface that looked automated, there were always real people solving the work.
📌 Note: The name Mechanical Turk is a direct reference to the 1770 automaton that fascinated Europe by making people believe a machine was playing chess, when in reality a human player operated the mechanism from inside.
In its early years, MTurk became a key piece of computer vision research. Much of the manual labeling work behind ImageNet, the dataset that fueled the deep learning resurgence starting in 2012, was done by Mechanical Turk Workers drawing bounding boxes and assigning categories to millions of photographs. That pattern, using humans to produce training data at scale and low cost, was later repeated in hundreds of academic papers and commercial machine learning products.
Over time, MTurk also became a standard tool in social science and experimental psychology: university researchers recruited Workers as participants for surveys and experiments, drawn by the low cost and recruitment speed compared to traditional lab panels.
Technical Details: How Mechanical Turk Works (Worked)
MTurk’s operating model is organized around three pieces: the Requester posts a task called a HIT (Human Intelligence Task), the Worker completes it in exchange for a per-unit payment, and Amazon acts as the intermediary, charging a commission on every paid HIT. Typical tasks include data validation, transcription, content moderation, surveys, product categorization, and, above all, data annotation for training machine learning models.
A HIT is defined with a template, a per-unit payment, a number of assignments, and a time limit. Programmatic access happens via a REST API, with official SDKs like boto3 for Python. Here’s what the task flow looks like at a high level:
sequenceDiagram
participant R as Requester
participant M as Mechanical Turk
participant W as Worker
R->>M: posts a HIT (task + payment + deadline)
M->>W: shows the available HIT
W->>M: submits the completed assignment
M->>R: delivers the result for review
R-->>M: approves or rejects the payment
M-->>W: credits the payment if approvedThat cycle, post, solve, review, pay, is what sustained, for 21 years, both data labeling for machine learning and the business process outsourcing of repetitive tasks: content moderation, database deduplication, or extracting information from menus and websites, as Food Genius (part of US Foods Data Science) described in Amazon’s own materials.
How to Prepare for the Shutdown (or Export Your Data Today)
If you currently manage a Requester or Worker account on Mechanical Turk, it’s worth acting before September 30, 2026, rather than waiting until the last month. Here are the concrete technical steps for Requesters using the API:
# install the official AWS SDK for Python
pip install boto3With your AWS account credentials configured (AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_DEFAULT_REGION), a minimal script to list pending HITs and export their results before the shutdown looks like this:
import boto3
mturk = boto3.client(
"mturk",
endpoint_url="https://mturk-requester.us-east-1.amazonaws.com",
region_name="us-east-1",
)
# list active HITs in the account
respuesta = mturk.list_hits(MaxResults=100)
for hit in respuesta["HITs"]:
hit_id = hit["HITId"]
asignaciones = mturk.list_assignments_for_hit(
HITId=hit_id,
AssignmentStatuses=["Submitted", "Approved"],
)
for asignacion in asignaciones["Assignments"]:
print(hit_id, asignacion["WorkerId"], asignacion["Answer"])This script loops through the account’s HITs and downloads each Worker’s responses (Answer), the essential step before exporting them to a CSV or your own database before the API stops responding. For Workers, Amazon’s recommendation is to review pending balance and approved assignments from the dashboard before the shutdown date.
⚠️ Heads up: Amazon did not specify what happens to Requester balances or to Worker payments left unprocessed after September 30, 2026. The only official guidance available is the FAQ page linked from mturk.com.
For those relying on MTurk as a source of human workforce within machine learning pipelines, the most direct replacement within the AWS ecosystem itself is Amazon SageMaker Ground Truth, which until now offered MTurk as one of its three workforce options, alongside a private workforce or one provided by a third-party vendor. With the shutdown, that specific option within Ground Truth no longer makes sense, and teams will need to migrate to a private workforce or a vendor.
| Alternative | When to Use It | Advantage | Limitation |
|---|---|---|---|
| SageMaker Ground Truth (private workforce) | You already use AWS and need control over who annotates | Integrates directly with your ML pipeline in AWS | You build and manage the annotator team |
| Managed providers (Scale AI, Appen, Surge AI) | You need large-scale annotation with an SLA | Trained workforce managed by the provider | Higher cost per task than MTurk |
| Prolific | Academic research or surveys with a controlled sample | Better demographic control than MTurk | Built for studies, not BPO or mass labeling |
| In-house workforce or freelancers | Low volume or highly domain-specific tasks | Full control over quality and training | Doesn’t scale as fast as a marketplace |
Impact and Analysis
The material Amazon still has published about MTurk includes testimonials explaining why the platform mattered so much to the artificial intelligence industry. Michael Schmitz, Engineering Director at the Allen Institute for AI (AI2), noted: “we use crowdsourcing platforms like Amazon Mechanical Turk to build datasets that help our models learn common-sense knowledge, something that’s often necessary to answer basic questions that are easy for humans but still hard for machines.”
David Falck, Executive Director of Food Genius (part of US Foods Data Science), described it from another angle, that of business process outsourcing: “Workers on Amazon Mechanical Turk respond to our requests to gather information from menus, websites, and other channels. We can leverage that collective human knowledge to better understand customer needs and uncover important market trends.”
Both cases sum up the dual role MTurk played: on one hand, data infrastructure to train and evaluate machine learning models, from image annotation to human-in-the-loop validation; on the other, a cheap and flexible way to outsource business processes that previously required hiring and scaling a temporary team of your own, something costly and slow, as Amazon’s own materials describe.
The shutdown also highlights a shift in eras. When MTurk launched in 2005, the only way to produce labeled datasets at scale was to pay people to label them by hand, one HIT at a time. Two decades later, much of that annotation work is partially automated by AI models that pre-label data and leave humans only the final review, reducing the need for a massive, generic microtask marketplace like MTurk.
💭 Key takeaway: MTurk was born to solve with humans what machines couldn’t do. Its shutdown coincides with an era in which AI models themselves handle much of that work, leaving humans as reviewers rather than primary executors.
What’s Next
Until September 30, 2026, Requesters and Workers can keep using the platform normally, according to the notice posted on mturk.com. Amazon did not disclose a gradual shutdown timeline, for example halting new HITs before that date, nor did it confirm whether the API will be disabled the same day or whether there will be an additional window just for exporting historical data.
There is also, for now, no announcement of a direct replacement service within the AWS catalog. The closest thing is reinforcing SageMaker Ground Truth’s other two workforce options, private and vendor-provided, though neither exactly replicates the open marketplace model that made MTurk popular among academic researchers and small startups with tight budgets.
For the broader data annotation ecosystem, the shutdown accelerates a migration already underway toward specialized providers and hybrid pipelines where an AI model does a first pass at labeling and a human only corrects the doubtful cases.
📖 Summary on Telegram: View summary
Try it yourself: if you have an active Requester account, run the boto3 list_hits script against your account today and confirm how many assignments you still have to export before September 2026.
Frequently Asked Questions
When exactly does Amazon Mechanical Turk shut down?
September 30, 2026. Amazon confirmed it in a notice posted directly on the homepage of mturk.com.
Why is it called Mechanical Turk?
It’s a reference to the 1770 chess-playing automaton that appeared to play on its own but hid a human operator inside. Amazon used the name to describe a service where, behind tasks that look automated, there are always real people solving them.
What happens to pending tasks and payments?
Amazon has not detailed this publicly beyond directing Workers and Requesters to its FAQ page to prepare for the shutdown.
Is there an official Amazon replacement?
No direct substitute service has been announced. Amazon SageMaker Ground Truth remains available with private or third-party vendor workforces, but it no longer offers MTurk as one of its workforce options.
What was MTurk mainly used for in machine learning?
Mainly for large-scale data annotation: labeling images, drawing bounding boxes, validating model outputs, and collecting survey or research data.
What alternatives exist today for data annotation?
Managed providers like Scale AI, Appen, or Surge AI, research platforms like Prolific, or your own private workforce managed through SageMaker Ground Truth.
References
- mturk.com: Amazon’s official notice about the Mechanical Turk shutdown on September 30, 2026.
- Amazon SageMaker Ground Truth documentation: workforce options for data annotation within AWS.
- boto3 documentation for the MTurk client: API reference used to list HITs and assignments.
- Wikipedia: Amazon Mechanical Turk: history and context of the service since its 2005 launch.
- Wikipedia: The Turk (automaton): history of the 1770 chess-playing automaton that inspired the name.
📱 Like this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Imagen destacada: Foto de Campaign Creators en Unsplash
0 Comments