9 min read - Mechanical Turk Is Closing: Make Human Review Portable

A human-review pipeline can look like a small API dependency. Send a task, receive an answer, approve the work. The people, instructions, quality checks and payment operations behind that call are easy to overlook until the service has an end date.

Amazon has now supplied one. Its live closure FAQ says Mechanical Turk will close permanently on 30 September 2026. Current workers and requesters can continue until then. The notice also affects the Mechanical Turk Worker option used in SageMaker Ground Truth labeling jobs and Amazon Augmented AI review workflows. It does not say that every Ground Truth or Augmented AI capability is ending.

For teams that still use MTurk directly or through an AWS workflow, this is an immediate migration deadline. For everyone else, it is a useful design review. Human review should be a portable business process, not an unnamed feature of one platform.

The closure has several clocks

Independent reporting from The Next Web on 26 August corroborated the shutdown notice and put it in context: Amazon launched MTurk in 2005 as a marketplace for small digital tasks that computers struggled to complete.

The full shutdown follows an earlier warning. On 30 June, AWS moved Mechanical Turk and several SageMaker AI features into maintenance, closing them to new customers from 30 July while allowing existing customers to continue. A maintenance notice and a permanent closure are different events. Teams that recorded only the first change need to reopen the decision.

AWS now gives requesters several dates to manage:

The FAQ says prepaid requester balances should be refunded within 30 days and asks requesters to verify their payment information. Those administrative steps matter, but they are not a migration plan. A replacement service cannot reconstruct undocumented qualification rules, examples, reviewer feedback or the reason a team trusted an answer.

Start by finding the real dependency

Search for more than the service name. MTurk can appear in application code, AWS SDK calls, scripts, notebooks, research protocols, finance processes and the settings of Ground Truth or Augmented AI workflows. Ask the people who own data, research, product operations and customer support whether they send work to a crowd or review vendor, even if they do not call it annotation.

For each active workflow, record:

• the business decision or model behaviour the human answer supports;
• the task template, examples, interface and version;
• the worker qualifications, language needs and exclusion rules;
• the input fields and data classification;
• the acceptance, rejection, bonus and dispute process;
• the gold cases, redundancy rules and escalation path;
• the result schema, downstream consumer and retention rule;
• the business owner, technical owner and payment owner.

That inventory reveals the unit that needs to move. It may be a survey, an image-labeling batch, a content-moderation queue, a model evaluation set or a manual fallback inside an automated workflow. Migrating the API without the operating rules preserves the integration and loses the control.

The official Mechanical Turk developer guide documents the main objects to trace, including HITs, assignments, qualifications, approval, rejection and worker management. Use those objects as a discovery checklist, not as the target design for the next platform.

Export evidence before comparing vendors

The MTurk FAQ says HIT and assignment data are normally removed after 120 days. The closure FAQ separately keeps transaction history available until 28 January 2027. Export the records your organisation is entitled and required to retain while the service still works, then test that the export can be read without MTurk.

Keep the smallest defensible set. Task versions, approved results, quality decisions, aggregate performance, payment reconciliation and change history may be useful. A bulk copy of worker profiles or sensitive task contents may create a new problem. Have the data, privacy and legal owners decide what may move, for which purpose, for how long and with which access controls.

Do not confuse transaction history with evaluation evidence. A bill proves that tasks ran. It does not prove that the instructions were sound, that reviewers agreed, or that the output met a release threshold. Preserve the rubric and adjudication record alongside the result.

Our guide to the domain evaluation asset explains why gold cases, expert disagreement and regression tests should outlive a model or provider. The same principle applies to the human part of the system.

Compare replacements with the same test pack

Do not start with a feature grid. Build a small, representative pack from authorised material and run it through the current workflow and each candidate route. Include routine items, ambiguous cases, known failure modes, out-of-scope requests and tasks that require escalation.

Score the whole operating path:

The right answer may be a managed crowd, a specialist review provider, an internal team or a mixed route. Do not assume that the cheapest unit price is the cheapest reviewed result. Rework, disagreement, low task availability and specialist escalation can dominate the total cost.

Avoid replacing one opaque dependency with another. Keep task definitions and gold cases in an organisation-controlled repository. Give result schemas versions. Place the provider adapter behind a narrow interface. Make it possible to replay the same approved test pack without exposing production data or rewriting the whole application.

Verify that human still means human

A platform label does not prove how a task was completed. A 2023 EPFL case study reran one abstract summarisation task on MTurk and estimated that 33% to 46% of participating workers used large language models. The authors explicitly warned that the result might not generalise to less LLM-friendly tasks.

That narrow finding should not be turned into a claim about all MTurk work or all crowd workers. It does expose a control question that survives the platform closure: when a process requires independent human judgement, what evidence makes that requirement credible?

The answer depends on the task. It may include clear disclosure rules, controlled interfaces, planted gold cases, response-process signals, independent duplicate reviews and expert adjudication. If AI assistance is allowed, define where and why. A concealed mixture of model output and human approval is hard to evaluate because the organisation no longer knows which capability it measured.

Test language and regional fit directly

France and Morocco teams may need French, English, Arabic or Darija review, sometimes in the same workflow. A provider's global-worker claim does not establish proficiency in the required language, domain or local context.

Build native-language test items from the approved use case. Check whether instructions remain unambiguous, whether reviewers understand local terms, and whether disagreement rises when a task mixes languages. Record who can adjudicate the difficult cases. Our guide to local-language AI offers a broader evaluation approach for French, Arabic, Darija and Amazigh.

Reviewer location and platform ownership also affect data access and contract questions. Treat those as facts to verify for the specific workflow. Do not infer a residency or compliance outcome from a vendor's marketing page.

A transition sequence that fits the deadline

In the next 48 hours, identify active MTurk credentials, requester accounts, SDK calls, Ground Truth workforce settings and outstanding batches. Name one owner for the closure and one owner for payment reconciliation.

During the next week, freeze task-template changes long enough to create a known baseline. Export allowed records, document quality controls and assemble the test pack. Contact affected internal users and any people responsible for worker disputes or specialist escalation.

Then run the current route and the candidate route in parallel on fresh, authorised examples. Compare agreement, adjudication effort, completion time, failure handling and full cost. Fix the task definition before blaming a new workforce for ambiguity already present in the old process.

Before 30 September, stop creating work that cannot finish safely, switch the production adapter or activate the manual fallback, and verify that no hidden queue still targets the MTurk Worker option. Through 30 October, complete approvals, rejections, bonuses and dispute handling. Before 28 January, verify the final transaction archive and remove credentials that no longer have a purpose.

Sources and limitations

• Amazon's live Mechanical Turk closure FAQ provides the closure, approval, refund, bonus, transaction-history and related AWS-workflow dates used here.
• The 30 June 2026 AWS service-availability notice records the earlier maintenance status and new-customer cutoff.
• The Amazon Mechanical Turk developer guide defines the service objects and operating concepts used in the inventory.
• The Next Web's 26 August report independently corroborates the new closure notice and service history.
• The EPFL paper "Artificial Artificial Artificial Intelligence" supplies the scoped estimate about LLM use in one MTurk summarisation study and states its generalisation limit.

AWS may update closure instructions, and this article does not compare or endorse replacement providers. The transition model is operational guidance, not legal, privacy, labour or procurement advice. Confirm contractual duties, worker obligations, data handling and record retention for the specific organisation and workflow.

Preserve the review system

The most important export is the review system: what people were asked to do, how quality was judged, where difficult cases went and why the business trusted the result. A CSV of old answers cannot explain those decisions.

Mechanical Turk's deadline makes that work urgent for current users. It should make every AI team ask the same question: if the human-review provider closed next month, could the organisation move the process without losing its quality evidence or its responsibilities?