By Amazon has confirmed it: AWS Mechanical Turk closes September 30, 2026. The crowdsourced task marketplace that Jeff Bezos once called “artificial artificial intelligence” — humans doing what software couldn’t — shuts down in five weeks. SageMaker Ground Truth and Amazon Augmented AI go with it. New signups already ended July 30. If you have active annotation pipelines, research studies, or RLHF workflows running through MTurk right now, there is no soft landing here. You are migrating — the only question is whether you do it on your timeline or Amazon’s.
Why It Died: AI Contaminated Its Own Training Data
Here is the irony worth sitting with. MTurk existed to provide human signal for tasks machines couldn’t do. Then the machines got good enough that workers started using them to complete the tasks faster. A 2023 study from EPFL — formally titled “Artificial Artificial Artificial Intelligence” — found that between 33 and 46 percent of MTurk workers were using LLMs to complete text tasks. The financial incentive was obvious: faster completion, more tasks, more pay.
The practical consequence: any MTurk-sourced text annotation, preference ranking, or NLP benchmark dataset collected after November 2022 — when ChatGPT launched publicly — carries an unresolved provenance question. Is this human signal or AI signal? A 2023 Data Integrity Consortium study found AI-contaminated datasets degraded downstream model performance by up to 38 percent in sentiment analysis tasks. You cannot retroactively fix it. Label provenance was never tracked at the source level.
The platform that trained AI on human judgment ended up with humans outsourcing that judgment to AI. Amazon apparently ran the numbers and decided the signal-to-noise ratio no longer justified operating the service.
What Is Actually Being Shut Down
Three AWS services close on September 30. Make sure you account for all three:
- Amazon Mechanical Turk — the main crowdsourced task marketplace
- Amazon SageMaker Ground Truth — the managed labeling service that routed work to MTurk and vendor networks
- Amazon Augmented AI (A2I) — the human-in-the-loop review workflow service
Amazon has not announced a formal migration path from any of these to a replacement. If you were planning to shift MTurk work to SageMaker Ground Truth, that option is gone too.
Where to Go: Migration Options by Use Case
There is no single drop-in replacement for MTurk because MTurk served too many different use cases. Pick based on what you were actually doing:
Enterprise ML Data Labeling
Scale AI and Appen handle large-scale classification, transcription, and multimodal labeling with structured quality-control pipelines. Note: Scale AI saw some customer defections after Meta acquired a 49 percent stake in June 2025, and Mercor has grown partly by absorbing that migration traffic.
RLHF and Expert Annotation
Mercor and Surge AI operate in the expert-labeling tier — vetted annotators for reinforcement learning from human feedback. Surge AI prices at roughly $85 to $200-plus per expert hour. This is not cheap-crowd pricing, which is the point: the quality signal is worth it when you are fine-tuning frontier models.
Academic Research and UX Studies
Prolific is the clearest successor for researchers. Participants are verified, demographically representative, and paid at minimum UK wage equivalent. The platform charges a 33 percent fee on top of participant pay. CloudResearch was built partly as an MTurk alternative for academic work and handles existing study continuity well.
Internal and Hybrid Approaches
For sensitive or proprietary datasets, AI-assisted pre-labeling combined with human spot-checking from an in-house team is increasingly viable. The cost math has shifted enough that this is worth modeling against third-party vendor pricing before you commit to a platform.
What to Do Before September 30
Five weeks is not long when workflows have dependencies. The practical checklist:
- Audit every active MTurk workflow — what is still running, what volume, what downstream systems depend on it
- Assess your existing MTurk datasets — particularly any text annotation or preference data collected after November 2022
- Select your replacement based on use case, not just price per task
- Migrate before September 30 — there will be no extension
- Implement label provenance tracking in your new pipeline from day one
For a detailed comparison of migration options, LXT has a useful breakdown of the major alternatives with feature comparisons.
The Broader Shift
MTurk’s shutdown ends the era of cheap, anonymous crowd labor as a viable input to AI training. The market is now splitting into two distinct tiers: fully automated pipelines where AI labels AI (with all the compounding reliability risks that entails), and high-quality vetted human annotation where provenance, payment, and qualification are tracked explicitly. The undifferentiated middle — a crowd you paid cents per task and hoped was human — is gone. Given that it may not have been mostly human for the past few years anyway, that is probably the right outcome.
