AWS to Shut Down Mechanical Turk in September, Ending 21-Year Crowdsourcing Microtask Platform
AWS announced it will shut down Mechanical Turk (MTurk) on September 30, 2026, ending the crowdsourcing microtask platform that has been in operation for about 21 years. The company stated that the decision comes from a routine evaluation of projects, tools, and services.
Launched in 2005, MTurk allowed businesses, researchers, and developers to break down tasks that machines struggle to complete reliably into human intelligence tasks, which could then be completed by dispersed online workers for payment. Typical tasks included image labeling, content moderation, data classification, surveys, and transcription.
The AWS announcement did not disclose revenue, user numbers, active worker counts, order sizes, specific evaluation metrics, or alternative product arrangements related to the shutdown. Current service terms still indicate that Amazon Augmented AI can call upon MTurk labor, but customers are prohibited from providing protected health information or personally identifiable information to that labor.
MTurk has long been a foundational node in the machine learning data supply chain: research teams quickly procured human annotations through it, and businesses could distribute image, text, or audio review tasks to global crowdsourced laborers via API. Its closure will force clients reliant on the platform to migrate task processes, rebuild worker pools, or choose alternative annotation services before September 30. cnbc+1
The platform's original name comes from the 18th-century "Mechanical Turk" chess-playing hoax; Amazon founder Jeff Bezos once referred to MTurk as "artificial artificial intelligence," meaning it allows human workers to perform tasks that appear to be completed by machines. The product emerged before the industrialization of large model training, positioning itself as a way to package human judgment as a programmable cloud service.
From a market mechanism perspective, the closure of MTurk will shift the demand for low-cost, task-based human annotation from buyers to professional data annotation vendors, crowdsourcing alternative platforms, BPO suppliers, and synthetic data tools; clients needing high-quality, multilingual, or sensitive content review will face higher migration and quality control costs. Vendors with closed expert networks, enterprise-level privacy compliance, and auditable data sources will benefit; research projects and small AI teams relying on open, cheap labor and quick access to the MTurk API will be under pressure.
Source: Public Information
ABAB AI Insight
The historical value of MTurk lies in transforming human judgment into on-demand cloud resources. When it launched in 2005, deep learning had not yet formed today's data flywheel; a large amount of natural language, visual, and behavioral science research relied on it to quickly obtain survey samples, preference labels, and human validation. Bezos's term "artificial artificial intelligence" reveals its business logic: when algorithms cannot judge, the platform uses humans to complete the last mile while concealing the complexity of organizing dispersed labor from the demand side.
The shift in capital pathways is from an open microtask market to a controlled data production system. Companies needing foundational models require not just one-time labels, but data assets that are copyright traceable, privacy manageable, task consistency measurable, and domain expert callable; this will push budgets toward data operators like Scale AI, enterprise BPOs, professional evaluation teams, and synthetic data infrastructure. MTurk's low-barrier supply is suitable for long-tail tasks but struggles to naturally meet the auditing chain required for high-risk training data.
A historical analogy is the reevaluation of artificial data service providers like Appen and TELUS International in the generative AI boom: they have expanded from search quality assessment, voice and content labeling to large model alignment, red team testing, and multimodal evaluation. MTurk represents the "open market matchmaking" phase; as large model training enters scaling and compliance, competition shifts to control over data licensing, quality control, labor management, and corporate contracts, rather than just competing for the lowest unit price.
This belongs to industry chain reconstruction. In the past, annotation demanders could post tasks to the open market to obtain labor; today, training data has become part of model capability, legal risk, and commercial delivery, requiring buyers to have verifiable sources and responsible parties. The closure of MTurk will not eliminate the demand for human data but will shift value from open task matchmaking to closed supply chains with data governance, professionals, and compliance guarantees.
ABAB News · Cognitive Laws
When data becomes an asset, cheap labor must turn into auditable supply
The platform's moat is not matchmaking efficiency, but responsibility assumption
The stronger the automation, the more human judgment concentrates on high-risk areas