Mechanical Turk shutdown: Amazon ends the “artificial AI” marketplace

Amazon did not rebrand Mechanical Turk for the agent era. It turned out the lights. After two decades of HITs measured in cents, the original crowd desk for “things computers cannot do” gets a hard stop date.

Mode: rotate · Category: Business

What Amazon announced

A notice on the MTurk site says the company regularly reviews programs and, after an assessment, will close AWS Mechanical Turk effective September 30, 2026. Amazon spokesperson Montana MacLachlan told CNBC the company remains “focused on delivering the best possible experience for all our Amazon and AWS customers” — standard corporate goodbye language for a product that no longer fits the AI stack roadmap.

Reporting: CNBC on the Mechanical Turk shutdown, Quartz on the Sept. 30 close.

Amazon posted FAQ support for affected users. Expect payout, account, and data-export windows to tighten as September approaches — treat mid-September as the practical deadline, not the calendar day itself.

What Mechanical Turk was

Launched in 2005, Mechanical Turk matched businesses with remote workers for microjobs: image labels, transcription, surveys, content moderation scraps. Pay was often pennies per HIT. The name nodded to the 18th-century chess “automaton” that hid a human operator inside.

Bezos’s line — artificial artificial intelligence — was the product thesis. Early Amazon needed humans to sort catalog chaos computers could not. Later, AWS sold the same workforce as a plug-in for SageMaker human review and labeling, advertising 24/7 availability and fast turnaround.

At peak, Amazon said the platform reached more than 500,000 workers. For many Turkers it was maternity-leave side cash, disability-compatible remote work, or a full-time gig with smartphone hours — closer to Flex delivery economics than to a W-2 data-science job.

Why the Mechanical Turk shutdown happened

Three pressures stacked:

1. Model quality — tasks that needed a human in 2012 (basic captioning, simple classification) now run through foundation models or lighter human-in-the-loop stacks
2. Specialist rivals — Scale AI, Prolific, Mercor and others built screening, domain experts, and RLHF pipelines that enterprise AI buyers prefer over anonymous penny HITs
3. Platform neglect — Turkopticon organizer Krista Pawloski told CNBC MTurk had been “in decline” as Amazon put fewer resources into improving it while workers migrated

Quality scandals did not help. A 2023 Swiss academic study found up to 46% of MTurk workers used AI models to complete tasks — humans hiring other AIs to fake human judgment. That is poison for training data buyers who thought they were buying ground truth.

Last month’s freeze on new customers was the tell. Workers read it as the soft close; Tuesday’s notice was the hard date.

Timeline to Sept. 30

• July–August 2026: new requester signup frozen
• Aug. 25–26, 2026: public shutdown notice; press coverage spikes
• Now → mid-September: export earnings records, finish open HITs, migrate requester workflows
• Sept. 30, 2026: AWS Mechanical Turk closes

Insurance and travel firms that still leaned on MTurk for surge labeling are already shopping replacements, per Turkopticon. If you are a requester with production pipelines pointed at the MTurk API, treat this as an incident, not a calendar reminder.

What Turkers and requesters should do

Workers

1. Download tax and payment history now — do not wait for September inbox chaos
2. Port reputation where possible; many rival platforms want proof of prior annotation work
3. Budget for a pay-mix change — specialist AI task markets often pay better hourly but screen harder and spike less predictably
4. Follow worker groups such as Turkopticon for scam warnings on “migration” phishing

Requesters / AI teams

1. Inventory every workflow that calls the MTurk API or SageMaker-MTurk workforce path
2. Re-quote labeling with platforms that support RLHF, multimodal review, and verified skills
3. Re-run quality audits — if your gold set was MTurk-only, assume contamination risk from the AI-assisted HIT era
4. Update vendor risk docs; auditors will ask what replaced crowd labor

Where the work goes next

Annotation did not die. It professionalized. Frontier labs buy preference data, safety red-teaming, and domain experts — not $0.03 image tags from anonymous accounts. Prolific pitches verified participants and API hooks; Scale and peers sell managed pipelines; niche shops hire credentialed labelers for medical or legal text.

Amazon still sells the rest of the AI stack — chips, cloud, Bedrock, agents. Closing MTurk is less “Amazon exits human data” and more “Amazon stops running a low-margin gig bazaar that competed with its own customers’ preferred vendors.”

That sits next to the other Aug. 2026 pattern: capital pours into machines that act, not marketplaces that hide humans inside the machine. See the agent-payments layer in our Stripe Link agents note and the on-desk inference push in Apple M5 Ultra.

What the shutdown means for AI labor

The Mechanical Turk shutdown is a tombstone for the first generation of crowd AI — cheap, opaque, global, and easy to game. The next generation is slower to onboard, more expensive per hour, and marketed as “expert” rather than “artificial artificial.”

For workers who built a household around HIT volume, Sept. 30 is a layoff without a pink slip. For model builders, it is a reminder that ground-truth labor still exists — just not as a 2005-style penny bazaar under the Amazon logo.

Bezos’s joke aged into a business decision. When the fake robot no longer needs the hidden human at Amazon scale, Amazon stops selling the box.

Informational only — not employment, tax, or vendor advice. Confirm deadlines and payout rules on Amazon’s MTurk FAQ and your local tax guidance before Sept. 30, 2026.