Product1 distinct publisher3 min readPublished
The banner says the closure is permanent. Requesters who wired human review into their stack get 35 days, and the firms that replaced MTurk do not sell what MTurk sold.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
The thing being switched off is not really a website. It was a labour primitive with an API attached: Amazon's own documentation advertised more than 500,000 workers in 190 countries, no minimum project size, payment only once you accepted the work, and the option to restrict a task to what it calls Master Workers [5]. That shape is what let a two-person evaluation team buy 400 judgements on a Thursday, and what let a lab field a survey without raising a purchase order.
The successors named in the reporting, Scale AI, Mercor and Prolific, recruit workers to train models [8]. That is a different product sold a different way. Where MTurk priced a Human Intelligence Task at a few cents, the newer firms sell curated workforces and expert annotation, and the market for supplying labelled data has grown rather than shrunk [15]. The work survives. The self-serve floor under it does not, and it is the floor that the smallest requesters were standing on.
There is an AWS-specific hole here too. Amazon marketed MTurk as the annotation source for SageMaker and promised customers the fastest turnaround for human review and labelling, with workers available 24 hours a day [9]. Anyone who took that at face value and put a human review step inside a SageMaker workflow now has a supply side that ends on 30 September, and a banner that points them at an FAQ [1].
The awkward part is provenance. A 2023 study by academic researchers in Switzerland found up to 46% of Mechanical Turk workers were using AI models to complete their tasks [11]. Teams that built datasets on the platform, including the Allen Institute for AI work on common sense knowledge that Amazon quotes on its own page [13], had one clean remedy available for a suspect batch: re-run the task with tighter qualifications and compare. After the marketplace closes, that remedy closes with it.
Amazon is saying very little. The homepage still sells content moderation, data deduplication, image categorisation and survey research in the present tense beneath the closure notice [6], the notice itself is boilerplate about regularly evaluating programs [3], and beyond it the company has published no separate statement, with a spokesperson offering that Amazon remains focused on delivering the best possible experience for its customers [4][14]. Krista Pawloski of the advocacy group Turkopticon told CNBC that the platform had been in decline for years as Amazon underinvested and rivals recruited its workers away [7], and that insurance and travel companies still using it are now scrambling [12]. Amazon had already stopped accepting new MTurk customers a month before the notice [10], which is the clearer signal in hindsight.
Do the arithmetic on the window. Thirty-five days is five calendar weeks [2], and counting back from 30 September puts the notice around 26 August [1]. Five weeks is less time than many enterprises need to get a data processing agreement signed with a new vendor, let alone qualify a worker pool and revalidate label quality against an existing gold set. The workers have less recourse than that: Pawloski says some still do this full time and are concerned [16], and there is no counterparty to renegotiate with when the marketplace itself is the thing being retired.
Ranked by verification strength, evidence, and original report placement.
The closure banner reads: "We regularly evaluate our programs, tools, and services and make adjustments based on those assessments."
Amazon's own product documentation puts the marketplace at more than 500,000 workers across 190 countries, with no minimum project size, round-the-clock availability, no payment until the requester accepted the work, and the ability to restrict jobs to Master Workers.
Krista Pawloski, a data worker and organiser with the advocacy group Turkopticon, told CNBC that MTurk had been in decline for years, that Amazon put fewer resources into improving it while rivals arrived, and that many turkers moved to competing platforms.
Pawloski said some insurance and travel companies still rely on MTurk for data work and are now scrambling to find alternative platforms; they have 35 days from the announcement.
Amazon spokesperson Montana MacLachlan said the company remains focused on delivering the best possible experience for all its Amazon and AWS customers.
A banner on the MTurk homepage says Amazon has decided to close AWS Mechanical Turk effective September 30, 2026, that it will permanently close, and directs workers and requesters to an FAQ page to prepare.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Primary notice quoted, but one publisher and no Amazon detail
The core fact rests on a primary artifact - Amazon's own homepage banner, quoted verbatim with a date and the word 'permanently' - plus an on-record spokesperson statement and a named CNBC report. That is strong for the shutdown itself. It weakens beyond that: the cluster has a single publisher, Amazon issued no standalone statement, transition terms exist only behind an unquoted FAQ, and the scale and decline claims come from vendor documentation and one advocacy source rather than independent measurement.
Real but unquantified current dependency
There is genuine usage evidence: a documented marketplace of more than 500,000 workers in 190 countries, a named customer testimonial from the Allen Institute for AI, and reported ongoing reliance by insurance and travel companies plus near-full-time workers. But every current-state number is missing - Amazon has disclosed no active worker or requester count, intake was already closed, and the platform is described as years into decline, so present-day adoption cannot be sized.
Vendor copy still overstates a platform being switched off
The story itself is soberly framed - a dated shutdown, a five-week window, an explicitly stale headcount - so the gap is small. It is positive rather than zero because the marketing surface overstates reality in two directions at once: the homepage still sells a global on-demand 24x7 workforce in the present tense underneath its own closure banner, and the 500,000-worker figure keeps circulating for a marketplace the same article describes as years into decline with up to 46% of workers running tasks through AI models.
Boilerplate from Amazon, advocacy framing from the main human source
Incentives are visible on both sides and disclosed. Amazon gives only a boilerplate assessment line and a customer-experience statement, which serves it to say nothing about why a service is being retired or what requesters and workers get. The main human source is an organiser with Turkopticon, a worker-advocacy group with a direct interest in framing the closure as the end of years of neglect. Named alternatives - Scale AI, Mercor, Prolific - stand to gain migrating demand. The publisher also threads in links to its own prior coverage.
Shutdown fact solid, consequences thinly sourced
High confidence that the closure and its date are real, because the banner is quoted and a named reporter at a major outlet carried it. Lower confidence on everything the story turns on operationally - how many requesters and workers are affected, what the FAQ offers them, whether the 35-day framing is exactly right for all contract types, and how the decline actually progressed - all of which rests on one publisher, one advocate and undated vendor documentation.
invest
Nvidia's Perplexity talks move its money one layer further from its own chips1 distinct publisher
invest
Amazon puts Prime Air in nearly 500 cities, and the arithmetic does not tidy up1 distinct publisher
invest
Amazon takes Prime Air to nearly 500 towns: scale arrives, with a five-pound ceiling1 distinct publisher
science
Starwashing: treat space environmental claims as marketing until someone audits them1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 26, 2026