Amazon is closing Mechanical Turk on September 30, ending the crowdsourced-labor marketplace that helped train the machine-learning era

Amazon confirmed on August 25 that it will permanently close Mechanical Turk on September 30, 2026 -- shutting down, after 21 years, the crowdsourced-labor marketplace that quietly did more to build the machine-learning era than almost any single piece of software still running. The company's own explanation, posted to the platform itself, is two sentences long: "We regularly evaluate our programs, tools and services and make adjustments based on those assessments. Following an assessment, we've made the decision to close AWS Mechanical Turk, effective September 30, 2026."

Amazon launched Mechanical Turk in November 2005 as a marketplace for what it formally called Human Intelligence Tasks -- CAPTCHA solving, sentiment labeling, surveys, content moderation -- paid by the task, often for pennies apiece.(The name is a wink at an 18th-century hoax: a "automaton" that appeared to play chess on its own while concealing a human operator inside its cabinet.) Jeff Bezos, promoting the service at launch, called it "artificial artificial intelligence" -- software standing in for a human doing work a computer of the day couldn't. In 2018, Amazon folded the platform into SageMaker as a data-annotation tool, and its customer base shifted from marketing researchers toward machine-learning teams that needed humans to label the training data their models would learn from.

The proximate cause, reported consistently across the coverage, is a data-integrity problem AI itself created. A 2023 study found that 33% to 46% of Mechanical Turk workers were using large language models to complete the very tasks they were being paid to do -- writing survey answers, labeling images, transcribing audio -- with an AI system standing in for the human judgment the platform was supposed to be selling. That is close to a category failure: a marketplace whose entire value proposition was authentic human judgment on tasks machines couldn't yet handle, undermined by workers using the machines those tasks existed to train.

What actually closes on September 30 is narrower than some of the coverage suggested. Mechanical Turk the marketplace shuts down outright. SageMaker Ground Truth and Amazon Augmented AI -- the tools AWS customers use to run their own labeling and human-review pipelines -- keep running, but lose their built-in option to route tasks to Mechanical Turk's public crowd; anyone still using that integration has to switch to a private or vendor workforce by the same date.

The gap Mechanical Turk leaves is not going unfilled -- it is going to a different kind of company. Scale AI, Mercor, Prolific and Surge AI have spent the last several years building the business Mechanical Turk never adapted into: instead of an anonymous crowd bidding pennies on microtasks, they recruit vetted specialists -- engineers, scientists, domain experts -- to do the harder, better-paid work of evaluating and correcting frontier-model outputs, the kind of RLHF and preference-labeling work that underpins post-training on today's large language models. Mechanical Turk's anonymous, lowest-bidder model was built for an earlier era of AI training data; it was not built for what labeling AI well now actually costs.

None of that six-line timeline says anything about the people who actually did the tasks. 500,000+ (workers, across 190 countries, at Mechanical Turk's peak) Krista Pawloski is one of them: she started "turking" in 2008 during a maternity leave, went full-time in 2012 after a layoff left her caring for a son with special needs, and has organized with the worker-advocacy group Turkopticon since 2019. She described the platform's appeal as "meaningful income with flexible hours, comparable to ride-hailing or Amazon Flex delivery, with the advantage of being remote and doable on a phone."

“There's some people that still pretty much still do it full time. They're concerned now.” — Krista Pawloski, Mechanical Turk worker since 2008 and organizer with Turkopticon

Amazon's own announcement does not mention workers at all -- no transition plan, no severance, just 35 days between the August 25 confirmation and the September 30 close. A company spokesperson, Montana MacLachlan, said only that Amazon "remains focused on delivering the best possible experience for all its Amazon and AWS customers" -- a sentence about customers, in an announcement about a service whose entire product was other people's labor.

There is a specific irony in how this ends. The workers Mechanical Turk is losing were never just doing survey work -- for two decades they labeled images, transcribed audio, and rated outputs that trained the pattern-recognition and language systems now capable enough to do those same tasks themselves, cheaply enough that even the workers started quietly using them. Mechanical Turk did not just employ people who helped build the current wave of AI systems; when nearly half its own workforce started delegating the work to those systems, it also became one of the first labor markets an actual model displaced from the inside.