Amazon will permanently close Mechanical Turk on September 30, 2026

As MTurk's largest requester for the past 10 years, I learned about the shutdown at the same time as workers. Our lead AWS contact moved to Bedrock and SageMaker Model Evaluations 2 to 3 years ago, and MTurk was left with essentially no team after stored-value accounts moved to native AWS billing.

I think the idea has more potential now: agents could hire multiple people for physical-world tasks. There are also training and tuning uses, while Mercor is valued around $20B for a similar labor-market model.

I do not think the underlying model is disappearing. Companies still hire people in bulk for labeling, transcription, RLHF, and moderation, but they now target large AI companies rather than people seeking repetitive online work.

I am not surprised. MTurk's unskilled tasks are the sort of work AI can do well enough that human verification may erase the cost advantage, while verifying AI output now needs domain experts.

I suspect robotics control will create a new category. A deployed robot that cannot fold a T-shirt could ask a crowd worker for help for $1, unless companies instead use a small group of highly trusted operators.

The last I heard, people in developing countries relied heavily on MTurk for AI grunt work. This shutdown probably means that data work has moved to other services.

I assumed many MTurk tasks had already shifted to LLMs and OCI. The platform's repetitive work seems like a direct automation target.

I earned a couple thousand dollars doing tasks between meetings, though Amazon Payments was awkward to redeem. I basically wrote much of someone's doctoral thesis by summarizing psychology papers for a dollar each.

MTurk saved me when I was isolated in a small-town newspaper job in 2005. I started Turking.com around the early community, but the initial excitement and available work fell off after a few months.

I tested MTurk through Bedrock Ground Truth a few months ago with simple MNIST digit recognition. I spent $10 and got accuracy only marginally better than guessing, which made me question research built solely on MTurk results.

I am surprised Amazon let MTurk fade while Mercor, Scale, and similar companies are hiring people for data labeling. The timing looks especially odd given current AI training demand.

I used MTurk to transcribe my father's handwritten letters and journals. Workers sometimes left notes about enjoying his travel stories and the people in his life.

I once used MTurk during the search for Jim Gray, when people inspected satellite imagery of huge areas of the Pacific. That task seems trivial now, but it needed human review at the time.

Without seeing AWS's financials, MTurk had looked like an AWS outlier for years. AI aside, the closure is not very surprising.