What Happened
Amazon confirmed on Tuesday, August 25, that it will close Amazon Mechanical Turk on September 30, 2026, ending a 21-year run for one of the strangest and most consequential services the company ever launched. The notice went up on the Mechanical Turk website itself, framed in the flat language Amazon reserves for retirements: the company regularly evaluates its programs, tools and services, and following an assessment it decided to close AWS Mechanical Turk. There was no farewell post, no blog essay, no executive quote. A marketplace that helped assemble the training data behind a decade of machine learning is being sunset the way a deprecated API endpoint gets sunset.
Mechanical Turk launched in 2005 as a marketplace for what Amazon called Human Intelligence Tasks, or HITs. A company or a university researcher would break a large manual job into thousands of micro-pieces — label this image, transcribe this audio clip, read this sentence and mark whether it sounds positive, answer this twelve-question survey — and post them for a distributed workforce to pick up. Payments were often measured in cents. Amazon has said the platform counted more than 500,000 workers at its peak, spread across dozens of countries, working from laptops, library terminals and, increasingly, phones.
The closure did not arrive without warning. On July 30, 2026, Amazon stopped accepting new Mechanical Turk customers, and AWS documentation at the time said existing customers could continue as normal while the company kept investing in security and availability but planned no new features. As TechCrunch noted in July, that is the standard vocabulary of a service on life support. Reporting around the announcement also indicates the shutdown pulls in the public-workforce path of Amazon SageMaker Ground Truth, which routed labeling jobs through the same crowd.
September 30 is a hard date, not a migration window with an asterisk. Requesters with open batches have about a month to drain them, and workers with pending approvals and unpaid balances have the same clock.
Why It Matters
The name was always the joke and the thesis at once. The original Mechanical Turk was an eighteenth-century chess automaton built by Wolfgang von Kempelen, a carved figure in a turban seated at a cabinet that appeared to play — and beat — human opponents across Europe. It was a hoax. A skilled human chess player was folded into the cabinet, sliding along a rail behind the machinery as the doors were opened for inspection. Jeff Bezos, launching the service in 2005, described the idea as "artificial artificial intelligence", and Amazon's own pitch leaned on the premise that there were still many things human beings could do more effectively than computers.
For most of the last two decades that premise held, and it quietly underwrote enormous parts of the field. Computer vision benchmarks were labeled by crowds. Sentiment classifiers were trained on crowd-annotated text. Behavioral economics and psychology papers were run on crowd participant pools deep enough to make sample sizes cheap. A meaningful slice of what the industry calls "AI progress" was, at the layer nobody photographed, several hundred thousand people doing piecework for pennies.
What changed is not that the human labor disappeared. It is that the shape of the demand inverted. The tasks Mechanical Turk specialized in — image classification, transcription, basic sentiment tagging, simple categorization — are now near-free capabilities of a language or vision model. The work that AI labs will still pay handsomely for has moved upmarket: physicians grading medical reasoning, lawyers stress-testing contract analysis, competitive programmers writing adversarial test cases. Scale AI, Mercor and Prolific built businesses on that upmarket demand while Mechanical Turk stayed pointed at the cheap end of a market that stopped existing.
Then came the reversal that makes the story genuinely strange. A 2023 study by Swiss researchers found that in one experiment, up to 46% of Mechanical Turk workers appeared to be using large language models to complete the tasks. A platform built to supply human answers where machines fell short had begun laundering machine answers back to buyers who were paying for human ones — and often using them to train the next generation of machines.
Reaction
Among the people who actually worked the platform, the mood reported around the announcement is closer to grim confirmation than shock. Krista Pawloski, a longtime data worker and organizer with the worker advocacy group Turkopticon, described the platform as in decline, saying Amazon appeared to invest fewer resources in it as competing data platforms gained ground and workers drifted elsewhere. On the platform's main Reddit community, one worker argued it had effectively died years ago, hollowed out by bots and fraud, and predicted exactly this ending: someone at Amazon eventually deciding the servers were not worth the electricity.
The concern that outlasts the nostalgia is economic. Mechanical Turk was flexible remote income for people who needed work they could do from home — caregivers, disabled workers, people in regions with thin local labor markets, people who could only work in fifteen-minute fragments. Pawloski said some workers still rely on it full time, and that those workers are worried now. There is no severance in a marketplace. There is a shutdown notice and a date.
Among researchers the reaction has been more procedural. Social science labs that built their subject-recruitment pipelines around the platform have to rebuild them, and the replication trail for a decade of published studies now points at infrastructure that will not exist. Most of that migration was already underway, pushed less by the shutdown than by mounting doubts about data quality.
What's Next
For AWS the practical answer is that the capacity, the attention and the capital have already moved. The division that is retiring a marketplace of half a million microtaskers is simultaneously spending at historic scale on inference silicon, power contracts and data center construction. Nothing about closing Mechanical Turk is a retreat from AI; it is a reallocation from the human-in-the-loop layer to the compute layer, and the timing says something about which one Amazon now believes is scarce.
For the data-labeling market, expect consolidation to accelerate at the top and a long tail to keep splintering. The specialist vendors will keep bidding up domain experts, because frontier model evaluation increasingly requires people who can tell a subtly wrong proof from a correct one. Meanwhile the genuinely low-skill annotation floor keeps getting eaten by synthetic data and model-assisted labeling, which is efficient and also risks a slow compounding of model errors into training sets with no outside human check.
The open question is verification. If models increasingly grade, label and generate their own training material, the industry loses the independent human signal it was built on — and Mechanical Turk's own 46% problem is a preview of how that failure looks from the inside. Whatever replaces it will need to prove the human is really there, which is a harder engineering problem than posting a HIT.
Closing Thoughts
Kempelen's Turk toured for decades, beat Benjamin Franklin and reportedly Napoleon, and provoked an entire literature of people trying to explain how a machine could think. Edgar Allan Poe wrote an essay arguing it had to be a fraud, on the grounds that a true machine would never lose. He was right about the fraud and wrong about the reasoning, which is roughly the accuracy rate of AI commentary ever since.
Amazon's version was more honest than its namesake — it never hid the humans, it invoiced for them — but it inherited the same lesson. The gap between what a system appears to do and what is actually happening inside the cabinet has been the defining ambiguity of this technology for two and a half centuries, and it did not get smaller when the cabinet became a cloud region.
What ends on September 30 is not really a product. It is a particular arrangement between people and machines, one where the human was the engine and the software was the interface. That arrangement has flipped, and Mechanical Turk is being closed by the very capability it spent twenty-one years helping to build. The automaton finally learned to play. The person inside the cabinet is the one being asked to leave.
한글 요약
아마존이 2005년부터 운영해온 크라우드소싱 노동 플랫폼 '아마존 메커니컬 터크(MTurk)'를 2026년 9월 30일 자로 종료한다고 8월 25일 공지했습니다. 이용자들은 이미지 라벨링, 음성 전사, 설문 응답 같은 잘게 쪼개진 작업(HIT)을 몇 센트 단위로 수행해왔고, 전성기에는 등록 노동자가 50만 명을 넘었습니다. 아마존은 7월 30일부터 신규 고객 접수를 중단한 바 있어 종료는 어느 정도 예고된 수순이었습니다.
이름의 유래인 18세기 '체스 두는 기계' 터크는 실제로는 캐비닛 안에 사람이 숨어 있던 속임수였고, 제프 베이조스는 이 서비스를 "인공 인공지능"이라고 표현했습니다. 그 전제가 뒤집힌 것이 이번 종료의 본질입니다. 이미지 분류·전사·감성 태깅처럼 MTurk가 특화했던 작업은 이제 AI 모델이 거의 공짜로 처리하고, AI 기업들이 비싼 값을 치르는 인간 노동은 의사·변호사·연구자 같은 전문가 평가로 이동했습니다. 스케일 AI, 머코어, 프롤리픽 등이 그 영역을 가져갔습니다. 2023년 스위스 연구진 조사에서는 한 실험 참가자 중 최대 46%가 대규모 언어모델로 작업을 대신했다는 결과도 나왔습니다.
노동자 단체 터콥티콘의 크리스타 파울로스키는 플랫폼이 이미 쇠퇴 상태였다고 말하면서도, 여전히 이 일을 전업으로 삼는 이들이 있어 우려가 크다고 전했습니다. 재택으로만 일할 수 있는 돌봄 종사자나 장애인 노동자에게는 대체재가 마땅치 않기 때문입니다. 한편 AWS의 투자는 인간 검수 계층에서 추론용 반도체와 데이터센터 쪽으로 이미 옮겨간 상태입니다. 남는 질문은 검증입니다. 모델이 학습 데이터를 스스로 만들고 채점하기 시작하면, 업계가 기대온 독립적인 인간 신호를 어떻게 보증할 것인가 하는 문제입니다.
참고 · Quartz · CNBC · Tech Startups