Amazon is shutting down Mechanical Turk after 21 years

Amazon is closing Mechanical Turk (MTurk) on 30 September 2026, 21 years after launch. Jeff Bezos once described the service as artificial artificial intelligence: humans quietly doing the work software could not yet do. Two decades on, the software has caught up, and the marketplace built around that gap is being switched off.

AWS stopped accepting new registrations on 30 July. SageMaker Ground Truth and Amazon Augmented AI close on the same date, which means Amazon is not trimming a product line. It is leaving human data collection altogether. The reporting comes from CNBC, TechTimes and The Next Web.

What is Amazon Mechanical Turk?

Mechanical Turk is an online marketplace where "requesters" post small jobs and "workers" complete them for payment. Each job is called a HIT, short for Human Intelligence Task. Typical HITs include tagging a photo, transcribing a snippet of audio or checking whether two product listings describe the same item.

The name borrows from the Mechanical Turk, an 18th-century chess-playing automaton that was really a person hidden inside a cabinet. That is the joke behind Bezos's phrase. The service looked like automation from the outside, but a human was doing the thinking inside.

MTurk launched in 2005 and went on to label an enormous share of the data that trained the models everyone now uses. For years it was the cheapest way to get a large volume of human judgement on demand.

What are the shutdown dates?

The wind-down has a fixed schedule. Requesters get the standard 30 day HIT approval window on anything completed before the shutdown, so approvals or rejections can run until 30 October. Anything left untouched after that auto-approves. Transaction history stays accessible until 28 January 2027.

If you have jobs running, that is the calendar to work backwards from. Finish and review anything in progress well before the end of September, because the approval window only covers work completed before the shutdown.

Why is Amazon Mechanical Turk shutting down?

The premise cracked when the models got good enough to do the tasks themselves. A 2023 study by researchers in Switzerland found that up to 46% of Mechanical Turk workers were already using AI models to complete their assignments. That is the whole story in one statistic: the platform was paying humans to do work that those humans were quietly outsourcing back to machines.

For a requester, that is a serious quality problem. A marketplace sells human judgement. If a large share of the answers come from a language model instead, the buyer could have called the model directly and skipped the middleman.

The market moved too. Scale AI, Mercor and Prolific took the annotation business by selling vetted, specialist labour rather than an open marketplace of anonymous microtasks. Frontier labs, meaning the companies building the largest AI models, now want a radiologist or a lawyer marking up training data, not whoever claims the HIT first.

Why does the end of MTurk matter?

It is worth looking at the shape of what happened. The people who trained the systems were replaced by the systems they trained, and the replacement was economic rather than dramatic. The work simply stopped being worth $0.02 a task to anybody.

That price tells you what the model of the business was. MTurk depended on enormous numbers of tiny tasks, each paid at a few cents or less, and on workers being anonymous and interchangeable. Once a model can produce an acceptable label for close to nothing, that arrangement has no room left to compete on price. What remains valuable is judgement that a model cannot yet supply.

For workers, the loss is real. Many people used the platform as a source of small, flexible income, and the same open marketplace that let anyone in is the one being closed. The vetted platforms that replace it screen applicants for expertise, so the entry-level microtask work does not simply move across.

Expect the work to keep going somewhere. Requesters who relied on MTurk for quick checks will need another route, whether a specialist vendor, an in-house review team or a model with a human spot-checking the output. None of those is as cheap or as open as the marketplace they replace.

What does Amazon exiting human data labelling signal?

The closure of Ground Truth and Augmented AI alongside MTurk is the important detail. Ground Truth was Amazon's tool for building labelled datasets, and Augmented AI let developers route uncertain model predictions to a human reviewer. Both were ways of putting a person into a machine learning pipeline. Shutting down all three shows Amazon has decided that human data collection is not a business it wants to keep running itself.

Data labelling has not disappeared. It has moved upmarket and got expensive. That fits the wider pattern of AI spend shifting from cheap volume to specialised capability, the same shift visible in OpenAI building its own inference silicon and in the kind of reasoning work behind OpenAI's Astra solving open maths problems. If you want the earlier coverage of the first closure to new users, see our report on Amazon closing Mechanical Turk to new users.

What should requesters do now?

Anyone with live work on the platform has a short list of jobs. These are the steps that follow from the published schedule.

• Finish any running HITs and make sure completed work is submitted before 30 September.
• Approve or reject every outstanding submission by 30 October, rather than relying on auto-approval.
• Export or record your transaction history before it stops being accessible on 28 January 2027.
• Decide where future labelling will go, whether to a vetted specialist provider or an in-house team.

Rejecting work is worth doing carefully. Workers' approval records follow them, so a considered decision is fairer than leaving everything to auto-approve, though that remains the default if you do nothing.

Can you keep your own data out of the AI training pipeline?

The end of open microtask labelling does not stop your data being used to train models, but it does shift where the human review happens. If you would rather keep your own prompts and documents out of that pipeline entirely, our guide to running a local LLM on a Mac is the practical starting point. A model running on your own machine does not send anything to a provider, so there is nothing for a reviewer to see.

Key takeaways

• Mechanical Turk closes on 30 September 2026, and AWS stopped taking new registrations on 30 July.
• SageMaker Ground Truth and Amazon Augmented AI close alongside it, so Amazon is exiting human data labelling entirely.
• Requesters have until 30 October to approve or reject outstanding work, and transaction history stays available until 28 January 2027.
• A 2023 study by researchers in Switzerland found that up to 46% of Mechanical Turk workers were already using AI models to do their tasks.