Mechanical Turk Shut Down This Week. Its Flaw Was Never AI — It Was Flat Intake.

Amazon Mechanical Turk shut down this week, on September 30, after 21 years. The company's goodbye was a single banner on the site: "We regularly evaluate our programs, tools, and services and make adjustments based on those assessments. Following an assessment, we've made the decision to close AWS Mechanical Turk." No eulogy, no transition plan beyond a 30-day window to pay out what was owed.

It's tempting to read this as another "AI ate the job" story. MTurk — launched in 2005, famously pitched by Jeff Bezos as "artificial artificial intelligence" — spent two decades as the place requesters posted small human tasks ("HITs") and over 500,000 workers picked them up: label this image, transcribe this clip, rate this search result. For years it was also the quiet backbone of AI training data. The irony writes itself: the platform that helped teach machine learning how to work is the thing machine learning has now out-competed.

But that framing skips the actual mechanism, and the mechanism is the interesting part.

The queue never asked who was on the other end

MTurk's intake was flat by design. Any requester could post almost any HIT. Any worker who cleared a bare approval-rate threshold could claim almost any HIT. There was a "qualification" system, but most requesters never configured it beyond the default. The marketplace's whole value proposition was speed and volume, not fit — post a task, and whoever grabs it first does it, for a few cents, no questions asked.

That was tolerable when the output just needed to be "good enough" — flagging spam, tagging objects in photos. It stopped being tolerable the moment AI needed training data it could actually trust. A 2023 study by Swiss researchers estimated that as many as 46% of MTurk workers were using large language models to complete the very tasks that were supposed to produce human-labeled ground truth. The queue had no way to tell a careful annotator from someone pasting the prompt into ChatGPT and moving on. It never had to, until the thing it was feeding started requiring a kind of quality it was never built to verify.

Look at what's eating MTurk's market instead: Scale AI, Mercor, Prolific. None of them win by having a bigger pool of anonymous workers. They win by vetting — credentialed domain experts, verified identities, specific qualification tracks matched to specific work — before anyone sees a task. The front door does the filtering that MTurk's never did.

The failure mode isn't unique to labor marketplaces

Swap "HIT" for "incoming project request" and this is a familiar story. A lot of intake processes are built the same way MTurk was: one open door, first-come-first-served, with maybe a checkbox for "priority" that nobody weights consistently. That's fine as long as what's coming through is low-stakes and interchangeable. It stops being fine the moment the requests start requiring real judgment about fit, risk, and value — and by then, nobody can tell a well-scoped ask from noise fast enough to matter, because the system was never asked to tell the difference. MTurk ran for two decades before that gap caught up with it. Most internal intake processes don't get nearly that long before someone notices the queue is full of things that never should have gotten in.

Admitura exists for the step MTurk skipped: before a request joins the queue, it gets scored against criteria your organization actually defines — so "whoever grabbed it first" and "the thing that actually deserves attention" aren't left to chance.

Sources: Amazon's official Mechanical Turk closure notice, cited via GuruFocus, August 2026; TechRadar Pro, "AWS is shutting down Mechanical Turk, which let humans beat AI at certain work tasks," September 2026.