Amazon Is Closing Mechanical Turk. Its Workers Were Already Using AI.

Amazon is closing Amazon Mechanical Turk on September 30, 2026, ending a 21-year run for the marketplace that at its peak paid more than 500,000 people around the world to label images, transcribe audio, moderate content, and answer surveys. The platform stopped accepting new customers on July 30. The closure itself was announced on August 25. For the people who earned money there, an entire category of paid work went from viable to worthless in about 2 months.

Amazon's explanation ran to one sentence: "Following an assessment, we've made the decision to close AWS Mechanical Turk, effective September 30, 2026." The same date retires the Mechanical Turk worker type inside SageMaker Ground Truth, which one account of the closure describes as Amazon's complete exit from the human data-collection infrastructure market rather than the retirement of a single product. Jeff Bezos coined the tagline himself back in 2005, and it still describes the business better than anything written since: artificial artificial intelligence. A person somewhere doing the small task the customer assumed a machine had done.

What the decline looked like from inside the work

Krista Pawloski does data work and organizes with Turkopticon, a worker advocacy group that has pressed Amazon on pay and treatment for years. She told reporters the platform had been in decline for some time, with fewer Amazon resources behind it as AI competition grew, and that workers had been drifting to other platforms well before the announcement. From the inside, a dying category looks like a paycheck thinning out across several years, and then a notice with a date on it.

The most instructive number in this story arrived 3 years before the shutdown did. A 2023 study by Swiss researchers found that up to 46% of Mechanical Turk workers in their sample appeared to be using AI themselves to complete assignments, the same assignments whose entire purpose was generating human-labeled data to train AI. The substitution had already happened inside the job. The platform closing is the accounting catching up with a change that started at the desk of the person doing the work.

Switchboard operators had a generation to see it coming

Switchboard operators were once among the largest employers of American women, and automation took that occupation away across roughly a generation. That pace left room for a daughter to choose differently than her mother had, for a school to change what it taught, for a town to build something else on the same street. Mechanical Turk's workers had a quarter.

Switchboard operators had a generation to see it coming. Mechanical Turk's workers had a quarter.

The half-life of skills is what this story is actually measuring

The half-life of skills is the interval between learning something and finding it obsolete, and that interval keeps compressing faster than any retraining program currently moves. Microtask work sat at the short end of it because the skill was easy to acquire, which is precisely what made it easy to automate. The stories worth following in agentic AI right now are all versions of this one: as an agent absorbs a task, it absorbs the skill attached to that task, and the person holding the skill finds out on the company's calendar rather than their own.

The work itself moved up a tier. Scale AI, Mercor, and Prolific now carry much of the AI data economy, and they recruit programmers, scientists, and domain experts to evaluate and refine what models produce. That tier is alive and paying well. The tier that anyone with a laptop and a spare hour could enter is the one that closed, and those entry rungs deserve more attention than they get, because rungs are where careers start.

The question this puts in front of every leadership team

I spent 20 years inside a Fortune 100 watching which work got contracted out first, and it was always the work that had been made simple enough to describe in a paragraph. That simplification is a kindness right up until it becomes a specification an agent can read. If your company runs on contingent labor, gig contractors, offshore microtask vendors, or a services partner billing by the ticket, this arrives with a date attached rather than as a thought experiment. Which of that work will an agent do competently within 18 months? What do the people doing it now get told, and how far ahead of the decision do they get told it?

A 2-month notice period is lawful almost everywhere. It also determines whether someone's next 12 months read as a transition or as a fall, and that difference rarely shows up in the business case that authorized the change. Read across the numbers behind AI job loss and the same pattern repeats: the aggregate employment figures stay calm while individual categories disappear on schedules nobody published.

If you are the person doing the work rather than the one budgeting for it, the question points inward and stays the same. What in your week could a competent agent handle by next spring, and what have you learned in the last 12 months that it could not? Asking that early is the whole advantage, and asking it does not mean you are behind. Most of the leadership teams I speak with are working through the same question with better slides and no better answer.

Mechanical Turk ran for 21 years and gave 2 months of notice. Both numbers came out of the same decision, and only one of them made it into the announcement. The number that matters now is how much notice the people around you get, and who in your company decides what that number is. Put it on the agenda at your next leadership meeting, before the calendar puts it there for you. And if you read this differently, I'm easy to find.

Here is what makes Alex a credible voice on this topic: Alex shaped Cisco's $1.1 billion innovation portfolio over 20 years and now advises leadership teams on what happens to a workforce when AI absorbs a task faster than the retraining budget moves.

Talk with Alex about your workforce plan →