Mechanical Turk Is Gone. What Your Ground Truth and A2I Workflows Need Next

Amazon Mechanical Turk closed on September 30, 2026, and the Mechanical Turk worker type disappeared from SageMaker Ground Truth labeling jobs and Amazon Augmented AI (A2I) human-review workflows the same day. The software around it still runs. The on-demand crowd behind it, more than 500,000 registered workers at its peak, does not.

If your pipeline sent labeling jobs or low-confidence predictions to that crowd, something in it is now either failing or quietly doing nothing. This guide covers how to find it, what to replace it with, and how to switch in a week without losing training time.

What exactly closed, and what still works?

The marketplace and its crowd are gone. Ground Truth and A2I themselves are not, but they now need a workforce you bring.

Existing Ground Truth and A2I customers can keep using both services with a private workforce or a vendor workforce. AWS says it will maintain security and availability but is not adding features, so treat both as stable but frozen.

One restriction is worth remembering on the way out. AWS never allowed personally identifiable information or protected health information to go to the MTurk workforce. If you designed workflows around that limit, a replacement that can handle sensitive data under contract may let you route work you previously kept away from humans.

Where is MTurk still hiding in your stack?

Check three places. Most teams find the first two quickly and miss the third.

1. Ground Truth labeling jobs. Any job whose work team is the public crowd. List jobs, work teams and workforces in every region you use:

for region in us-east-1 us-west-2 eu-west-1; do
  aws sagemaker list-labeling-jobs --region "$region" --output json
  aws sagemaker list-workteams     --region "$region" --output json
  aws sagemaker list-workforces    --region "$region" --output json
done

Look for work team ARNs containing public-crowd. Those jobs can no longer get workers.

2. A2I flow definitions. Every human loop in A2I starts from a flow definition that names a work team. List them and check each one's HumanLoopConfig.WorkteamArn:

aws sagemaker list-flow-definitions --region "$region" --output json
aws sagemaker describe-flow-definition --flow-definition-name  --region "$region"

This is the dangerous one. A2I usually sits in production, catching the predictions your model is least sure about. If those loops now have nowhere to go, ask what your application does with an item that never comes back from review. Does it wait, time out, or ship the model's low-confidence answer? Find out before a customer does.

3. Direct MTurk API calls. Scripts and services that created HITs themselves, outside SageMaker. Search your code and infrastructure for the MTurk client:

grep -rniE "client\(['\"]mturk|mturk-requester|CreateHIT|create_hit" .

These are often survey tools, one-off data collection scripts, or an eval harness someone built two years ago. They fail loudly, which makes them easy to find once something calls them.

Changing an ARN is not the migration. The work team you point to next needs instructions, training, a quality bar, and the capacity to absorb your volume on day one.

What actually disappeared: people, not software

Most replacement guides compare annotation tools. Ground Truth is still a perfectly usable tool. What you lost was three things the crowd gave you for free:

• Capacity on demand. Post a job at 2 a.m. and someone picked it up. AWS documentation described the MTurk workforce as typically the fastest option because it was available globally, around the clock.
• No hiring. No recruiting, onboarding, payroll, or scheduling.
• Pay per task. No minimums and no idle time.

What you also lost was a quality problem you may have stopped noticing. A Texas A&M study re-ran the same survey task on MTurk over nine years and found unusable responses rose from 2.4% in 2013 to 88.8% in 2022. A 2023 EPFL study estimated that 33–46% of MTurk workers used large language models on a writing task. If your evaluation sets or preference data came from the crowd, some of your "human" labels were model output.

So the migration is also the cheapest moment you will get to raise your quality bar. Whatever replaces MTurk should match what it did well and fix what it did badly:

Which replacement path fits?

There are three realistic paths. The right one depends less on the tool than on who manages the people.

If you were using MTurk mainly for offline training data, any of the three can work. If you were using A2I to catch live, low-confidence predictions, coverage and turnaround matter more than anything else, and a private team of three people will not cover a weekend.

A one-week migration plan

You can move a typical labeling or review workflow in five working days if you measure before you switch.

1. Day 1: Inventory and freeze a golden set. Run the commands above and list every job, loop, and script that touched MTurk. For each workflow, pull 300 past items with answers you trust. Three hundred is not arbitrary: if a team scores 300 out of 300, you can say with 95% confidence its error rate is under 1%. (Our guide to what "99% accurate" means explains the math.)
2. Day 2: Turn HIT instructions into real guidelines. MTurk instructions were written for strangers who would spend 30 seconds per task. A trained team needs more: the edge cases, worked examples of each label, and what to do when unsure. Take your last 50 disagreements or rejected HITs and write a rule for each.
3. Day 3: Parallel run. Send the golden set, plus a day of live volume, to the new team. Do not tell them which items are golden.
4. Day 4: Compare. Score accuracy against the golden set and agreement on live items. Read every disagreement; half will be guideline gaps, not reviewer errors. Fix the guidelines and rerun the misses.
5. Day 5: Cut over and keep measuring. Point the work team ARN, flow definition, or API at the new team. Seed 2–5% golden items into live traffic permanently, so a drop in quality shows up in a dashboard, not in your model.

If your A2I loops are live and unrouted today, do step 5 first with a stopgap team, then run steps 1–4 behind it. A rough review for a week is better than none.

When should you keep this in-house?

A managed team is not always the answer. Keep the work inside your company when:

• Volume is under a few hours of review a week, and your own engineers already understand every edge case.
• The data is so sensitive that no contract would satisfy your customers.
• The task needs credentials you employ, such as licensed clinicians signing off on outputs.
• You are still discovering what the label even means. Write guidelines yourself first; outsource once they stop changing daily.

How Akmon handles an MTurk migration

Akmon runs human-in-the-loop teams for AI companies: labeling, live review of model outputs before they reach customers, and model audits. For an MTurk migration, that means:

• Live within hours of kickoff. You brief one ops lead in the morning; the team is working by the afternoon, and the first production batch lands within a day.
• Capacity that matches the crowd's best feature. We have spun up 30 reviewers over a weekend, and run 24/7 coverage on live queues.
• A quality bar MTurk never had. Every delivery comes with a QA report. Accuracy is guaranteed at 99%+; below that, we redo the work at no cost.
• Your stack, not ours. We work in your tool and your queue.

The easiest way to judge us is the parallel run in step 3. Send us your golden set, and we will match your current turnaround on it before you switch anything. Book a call.

FAQ

Is SageMaker Ground Truth shutting down? No. Ground Truth closed to new customers on July 30, 2026, and lost the Mechanical Turk worker type on September 30. Existing customers can still run labeling jobs with a private or vendor workforce. AWS is maintaining it but not adding features.

Is Amazon Augmented AI (A2I) still available? For existing customers, yes, with a private or vendor work team. Human loops that pointed at the public MTurk crowd need a new work team.

Can I still approve or reject MTurk work? Until October 30, 2026. Submitted work you have not approved or rejected by then is approved automatically.

What happens to my prepaid MTurk balance? Amazon said prepaid requester balances would be refunded in full within 30 days of closure, with a final bill in your AWS billing cycle.

What is the best Mechanical Turk alternative for AI work? It depends on what you used it for. For small, steady internal work, a private work team. For large training-data batches, a labeling vendor with a measurable accuracy guarantee. For live review of model outputs, a managed human-in-the-loop team with committed turnaround and 24/7 coverage.

Did Amazon say why it closed MTurk? Only that the decision followed an assessment. It did not cite AI.

Sources

• Unite.AI: Amazon sets final closure of Mechanical Turk (closure FAQ: dates, approvals, refunds, Ground Truth and A2I)
• Codercops: MTurk and SageMaker Ground Truth shutdown (which workforce options remain)
• Refonte Learning: MTurk closure migration (existing-customer status; inventory commands)
• RuntimeWire: MTurk's maintenance mode (Ground Truth Plus end of support; AWS workforce description; Texas A&M study)
• Let's Data Science: MTurk closure (PII and PHI restrictions)
• AI Weekly: Sept 30 shutdown (EPFL estimate of LLM use)
• The Decoder: Amazon sunsets Mechanical Turk
• AWS: workforce authentication and restrictions (public-crowd work team ARN)