# Amazon Mechanical Turk Shuts Down Sept. 30: Act Now > Source: > Published: 2026-08-30 20:08:57+00:00 Amazon Mechanical Turk goes dark on September 30, 2026. After 21 years, the platform Jeff Bezos once called “artificial artificial intelligence” — and that quietly underpinned the training pipelines of modern AI — is shutting down. Developers have 31 days. Here is what actually breaks, and where to go next. ## A Platform Named After a Fraud Mechanical Turk launched in 2005, taking its name from an 18th-century chess-playing “automaton” that wowed European courts — until it was revealed to conceal a human chess master inside the cabinet. The joke was intentional. Amazon’s platform worked the same way: tasks that appeared automated were actually routed to a global crowd of human workers doing them for cents. Bezos described it in a 2007 New York Times interview as “artificial artificial intelligence” — humans performing tasks that computers couldn’t yet do. The name aged well. It aged almost too well. ## The Platform Ate Itself Here is the part worth sitting with: MTurk didn’t lose to a competitor. It lost to the thing it helped build. A [2023 EPFL study](https://arxiv.org/html/2306.07899) found that between 33 and 46 percent of MTurk workers completing text-production tasks were using LLMs to do the work — then submitting AI output as human annotation. The platform designed to supply genuine human judgment was quietly filling with machine output. Requester pipelines built on the assumption of human responses were getting GPT completions. Meanwhile, annotating via ChatGPT costs roughly $0.003 per item. MTurk’s full pipeline ran significantly higher, and in multiple benchmarks, LLM annotation now matches or exceeds crowd quality on structured tasks. The economics had already collapsed. Amazon is just making it official. ## What Actually Breaks on September 30 If you use MTurk directly via the API, your workflow stops. But the cascading failure is wider than that. **SageMaker Ground Truth**: The MTurk workforce type — `WorkforceType: PUBLIC` — is removed on September 30. Any labeling job configured to route tasks to the MTurk crowd breaks immediately. Private workforces and vendor workforces are unaffected, but public-workforce labeling jobs need to be reconfigured before the deadline. **Amazon Augmented AI (A2I)**: Also closes September 30. Human review loops that route low-confidence predictions to MTurk workers for review will stop functioning. If you have an A2I flow condition sending items to a public workforce, it needs a replacement now. Amazon closed new customer enrollment for both services on July 30. Existing customers can use them until September 30 — but no new HITs, no new flows, no new labeling jobs after that date. ## Where to Go Instead There is no single MTurk replacement. The right alternative depends on what you were using it for: **Prolific**— Best for surveys and behavioral research. Participants are pre-screened and paid at least $8/hour. Expect a 43% platform fee, but quality is meaningfully higher than MTurk for research use cases.[Prolific’s comparison guide](https://www.prolific.com/prolific-vs-mturk)covers the differences in detail.**Scale AI**— Enterprise-grade annotation and RLHF at volume. Managed service, custom pricing. Best if you’re running production ML pipelines and need contractual quality guarantees.**Surge AI**— Specialized RLHF and high-stakes annotation. Expert annotators run $85 to $200+ per hour. If you’re fine-tuning frontier models, this is the tier your budget needs to be in.**Toloka**— The closest analog to old-MTurk pricing for high-volume commodity tasks. Pay-as-you-go, global crowd. Quality controls are more manual, but the economics work for bulk classification.**Labelbox / Appen**— Enterprise labeling platforms with managed annotation teams. Better for organizations that want a vendor relationship rather than a marketplace. One concrete warning: a 2026 Gartner report found that 42 percent of businesses switching from MTurk saw a 15 to 20 percent drop in data accuracy during the transition period. Run both platforms in parallel for a few weeks if your timeline allows it. ## The Signal This Sends Amazon is not just shutting down a product. It is exiting human-data infrastructure entirely. MTurk, SageMaker Ground Truth, and Amazon Augmented AI all close simultaneously — a deliberate strategic withdrawal, not a cleanup. The market is splitting into two tiers: automated annotation using synthetic data or LLM-as-annotator for high-volume, lower-stakes labeling, and expert human review at $85 to $200+ per hour for RLHF and tasks where model quality depends on nuanced human judgment. The 21-year experiment in using human cognition to simulate AI has ended. September 30 is just the paperwork. [TechRepublic’s full shutdown timeline](https://www.techrepublic.com/article/news-amazon-mechanical-turk-shutdown/) and [ecorpit’s alternatives comparison](https://ecorpit.com/mturk-closing-human-data-labeling-alternatives-2026/) are worth bookmarking before the deadline hits.