Amazon is closing Mechanical Turk to new users on July 30. The platform that spent 21 years supplying the human judgment that trained modern AI systems is going into maintenance mode — not because AI improved beyond it, but because the workers providing that “human” judgment started using AI to fake it. The loop closed on itself.
What’s Actually Changing on July 30 #
After July 30, Amazon will stop accepting new requesters (the companies and researchers who post tasks) and new workers (the people who complete them). Existing accounts are unaffected — if you are currently running pipelines on MTurk, nothing changes for you right now. AWS says it will continue investing in security and availability, but new features are off the table.
This is not a shutdown. It is a quiet soft-retirement: the platform remains operational for its existing customer base while Amazon slowly redirects annotation demand toward Amazon SageMaker Ground Truth.
Why the Platform Failed #
MTurk launched in November 2005, named after an 18th-century chess-playing automaton that turned out to be a hoax — a human chess master concealed inside a “mechanical” player. Amazon borrowed the metaphor deliberately: artificial-looking intelligence with a human inside. For two decades, it worked. At peak, 500,000 workers across 100 countries completed HITs ranging from image labeling to preference ranking for RLHF datasets. Over 800 academic papers used MTurk data in 2015 alone.
Then the hoax flipped.
A 2023 study by researchers at EPFL found that 33–46% of MTurk workers were using large language models to complete tasks. The economic logic was unavoidable: GPT-class models outperform the average MTurk worker on standard annotation tasks by roughly 25 percentage points, at about one-thirtieth the per-annotation cost. Workers who discovered this had a straightforward financial incentive to let the model do the work and pocket the difference.
The result was self-defeating. MTurk’s entire value proposition rested on authentic human judgment — the signal that supervised learning and RLHF needed to distinguish good AI outputs from bad ones. Once a significant portion of that signal was itself AI-generated, the platform was selling a product it could no longer guarantee. The Register noted that not even AI can save it — the irony being that AI is exactly what broke it.
Your Migration Options #
If you are building a new pipeline and need labeled data, here is where to go depending on what you need: SageMaker Ground Truth— AWS teams running existing ML workflows. Active learning reduces human review by up to 70% on tasks where the model is confident. The most natural replacement if you want to stay in the AWS ecosystem.Prolific— Research teams and academics who need verifiable human responses. Prolific runs active anti-bot detection and its HUMAINE benchmark uses 27,000+ vetted evaluators. If authentic human signal is the requirement, not just human-looking signal, Prolific is the serious option.Scale AI / Surge AI— RLHF preference ranking and high-stakes AI training data where quality trumps cost. Both operate managed expert workforces. Surge’s client list includes Google, Anthropic, and Cohere.Label Studio— Teams with internal annotators who want to own the toolchain. Open source, self-hosted, and actively maintained.
The Data Provenance Problem You May Already Have #
If your team collected annotation data on MTurk after November 2022 — when ChatGPT became publicly available — that dataset has a provenance question attached to it. Any RLHF preference rankings, NLP benchmark labels, or behavioral study responses from that period may include a material proportion of LLM-generated output presented as human judgment. There is no clean way to retroactively verify which responses were authentic. This is not just an accuracy problem. The EU AI Act requires documentation of training data provenance. If your labeled data was secretly AI-generated and you have certified it as human, that is a compliance exposure worth auditing now rather than later.
What This Actually Signals #
Commodity crowdsourcing for ML annotation is over. Tasks that MTurk handled well — binary classification, basic tagging, simple preference ranking — are now cheaper and faster via LLM APIs, and anyone still routing those tasks through human crowds is paying a premium for a quality layer that may not actually exist. What survives is expert annotation: medical imaging review, legal document analysis, domain-specific labeling that requires genuine expertise rather than willingness to click quickly for a few cents.
The platforms positioned to grow from here are those that can either prove their workforce is human — Prolific’s approach — or guarantee their workforce is expert, not just cheap — Scale AI’s approach. The middle ground MTurk occupied, cheap-and-probably-human, is no longer viable.