Mechanical Turk is dead. The gig economy it spawned isn't.

What happened

Amazon is shuttering Mechanical Turk on September 30, 2026, ending a two-decade run for one of the earliest and strangest pieces of internet infrastructure. The notice went up on mturk.com and hit the top of Hacker News (333 points) within hours, with requesters and workers sharing screenshots of the shutdown banner and instructions for withdrawing balances before the cutoff.

MTurk launched in 2005 as an internal Amazon tool for de-duplicating product listings — humans doing tasks that were, per Jeff Bezos's original framing, "artificial artificial intelligence." The name was a joke about the 18th-century chess-playing automaton that turned out to have a person hidden inside, and for twenty years it was also a fairly accurate description of how a lot of machine learning actually worked. Requesters posted HITs (Human Intelligence Tasks) — label this image, transcribe this receipt, judge whether this sentence is toxic — at prices typically ranging from one cent to a few dollars. Workers, mostly in the US and India, did them by the thousand.

Amazon has not published a detailed post-mortem. The shutdown page cites a strategic decision and points requesters toward Amazon SageMaker Ground Truth for labeling workflows. Worker payouts and balance withdrawals continue through the end of September; the API stops accepting new HITs before that.

Why it matters

MTurk's fingerprints are on a shocking amount of the modern AI stack. ImageNet — the dataset that arguably kicked off the deep learning era in 2012 — was labeled largely on MTurk. Every psychology and behavioral-econ paper published between roughly 2010 and 2020 that ran "an online experiment with N=400 participants" almost certainly ran it on MTurk. RLHF pipelines at multiple frontier labs bootstrapped their early preference data there before graduating to specialist vendors. If you've used a language model in the last three years, some fraction of what it knows about "good" versus "bad" answers was decided by someone earning $0.08 per comparison on a platform Amazon has now decided isn't worth keeping open.

So why kill it? Three pressures converged. First, the enterprise labeling market got professionalized. Scale AI, Surge, Labelbox, Snorkel, and Toloka built vetted-workforce products with SLAs, quality metrics, and enterprise contracts that MTurk's self-serve marketplace was never designed to compete with. When OpenAI, Anthropic, and Google need millions of RLHF comparisons, they don't post HITs — they sign eight-figure contracts. Second, worker quality collapsed. By 2022, multiple academic studies were reporting that a meaningful fraction of MTurk survey responses were being generated by bots or by workers running LLMs themselves — the automation eating its own training data. Third, the ethical spotlight got hotter. A decade of reporting on sub-minimum-wage effective earnings, wage theft via mass rejections, and the total absence of worker recourse made MTurk a persistent reputational liability for AWS's enterprise sales motion.

The deeper story is that MTurk's original bet — that a fully open, self-serve, low-friction marketplace for human cognition would scale — turned out to be wrong in the same way that eBay-for-anything turned out to be wrong. The winners in every marketplace vertical were curated, gated, and specialized. MTurk stayed a flea market while the actual money moved to boutiques.

The Hacker News thread is heavy with eulogies from academics who ran their dissertations on it, and from workers describing a decade of side income vanishing on five weeks' notice. There's less nostalgia from people who tried to use it after 2020: the consistent complaint is that quality had degraded past the point where the price arbitrage made sense.

What this means for your stack

If you touch MTurk in any pipeline — even a dormant one — audit now. Concrete migration paths, in rough order of switching cost:

For ML labeling: SageMaker Ground Truth is Amazon's suggested landing spot and offers a similar API surface, but pricing and workforce model are different (private workforce, vendor workforce, or Ground Truth Plus managed service). Scale, Labelbox, and Surge are the serious alternatives if you want a vendor that will hand-hold. For open-source-friendly workflows, Toloka (Yandex's spin-out, now independent) is the closest ideological successor to MTurk's marketplace model and has been aggressively courting displaced requesters this week.

For academic survey research: Prolific has been eating MTurk's lunch in this segment since roughly 2019 — higher costs per participant but dramatically better data quality, real vetting, and attention checks that actually work. CloudResearch (formerly TurkPrime, which was itself a quality-layer on top of MTurk) has pivoted to its own participant pool. Budget an extra 2-3x per response and expect your data to actually be usable.

For reproducibility: This is the underappreciated crisis. Thousands of published papers cite MTurk HIT IDs and worker qualifications as part of their methods sections — after September 30, that infrastructure of scientific reproducibility simply ceases to exist, with no archival plan announced. If you're a researcher with active projects, export your HIT metadata and worker qualification data now.

Looking ahead

MTurk's shutdown isn't a surprise so much as the last date on a long death certificate — the platform has been a rounding error inside AWS for years. What's worth watching is whether anyone tries to rebuild the open marketplace model or whether the enterprise-vendor oligopoly cements. Toloka is the obvious contender; a handful of crypto-adjacent projects have been trying to build decentralized labeling markets for years with limited traction. Either way, the era of "I need 10,000 human judgments by tomorrow morning for eighty bucks" is over. The humans behind the AI are still there — they just cost real money now, and they work for someone else.