Decisions

Your Human-Checked Data Was Already Part Machine

On 30 July Amazon stops taking new Mechanical Turk customers. A 2023 EPFL study found up to 46 percent of crowd answers were AI. Audit your human-in-the-loop steps now.

DecisionsBy Servola Tech Desk2026-07-153 min read

AI-assisted, edited by humans. | Editor-in-Chief: Leon Soliman | Editorial standards

Your Human-Checked Data Was Already Part Machine

The check that broke the illusion

In 2023, a research team at EPFL in Lausanne ran a quiet experiment on crowd workers it had hired through Amazon Mechanical Turk to summarise text. Using keystroke logging and a classifier trained to spot machine-written prose, the team, led by Veselovsky, estimated that between 33 and 46 percent of the workers had used a large language model to produce their 'human' answers.

The finding was not a scandal about lazy workers. It was a warning about provenance. The marketplace sold human judgment by the task, and a third to a half of that judgment was already a machine wearing someone else's login. Anyone buying labelled data, moderation calls, or survey responses at commodity price was buying a mixture they could not see.

What Amazon actually did

On 30 July 2026, AWS stops accepting new customers for Mechanical Turk. Existing customers can keep using it, but the service now sits in maintenance: AWS says it will invest in security and availability yet will not add new features. In cloud terms, that is the road to retirement, not a pause.

Mechanical Turk launched in 2005 and was named after an 18th-century chess machine that hid a person inside a fake automaton. The name was the joke and the model: software that looked automatic but ran on people. Two decades on, the people were quietly running on software, and the economics that made the platform useful stopped adding up.

Why this lands on your desk

The story is not one service closing. It is every step in your stack that assumes a human checked something. Moderation queues, data labelling for a model you fine-tune, benchmark scoring, 'human review' compliance gates, survey panels behind a market-research dashboard: many of these route through cheap human-task marketplaces, and the 2023 result says the human layer was porous years ago.

For a European operator this is not only an efficiency question. The AI Act expects data governance and traceability for higher-risk systems, and 'a person reviewed the training set' is exactly the kind of claim an auditor will ask you to prove. If you cannot show who produced a label, you cannot show the label is clean.

The decision, not the nostalgia

The instruction is concrete and dated. Before 30 July, list every workflow that reaches a crowd-work marketplace, mark which ones feed a model, a compliance record, or a customer-facing decision, and decide for each whether to move it, verify it, or replace it. Treat 'we outsource that to humans' as a claim to test, not a control that works.

The base rate is the lesson. When you buy verification by the piece, the cheapest supplier will meet the letter of the task with whatever tool is fastest, and today that tool is a language model. Pay for proof of provenance, not just for an answer, and the shutdown of one marketplace stops being your problem.

Frequently asked questions

What exactly is Amazon changing about Mechanical Turk?

From 30 July 2026, AWS stops accepting new customers. The service moves to maintenance with no planned new features, and existing customers can keep using it for now. AWS has not given a full shutdown date, but maintenance status is usually the last stage before one.

Why is this happening?

Years of concern about bots, fraud, and data quality, made worse by AI. A 2023 EPFL study estimated 33 to 46 percent of workers used language models on a text task, and AWS has newer tools such as SageMaker Ground Truth that overlap with the same jobs.

What should an operator do before the cutoff?

Inventory every human-in-the-loop step, especially those that feed a model, a compliance record, or a customer decision. Verify who produces each output, and choose a replacement or a verified provider for anything you relied on Mechanical Turk to do.

The marketplace that promised human judgment is being wound down because the humans had stopped being the point. The lesson outlives the service: when you buy checking by the piece, buy proof of who did the checking too.

AmazonAWSMechanical Turkdata provenancecrowdsourcingAI training dataAI Act

Your AI Vendors Face EU Fines From August 2

The 2026 AI Act delay pushed the high-risk deadlines to 2027 and 2028, but it left the enforcement clock on general-purpose AI providers untouched. From 2 August 2026 the European Commission can fine any GPAI model maker up to 3 percent of global turnover. Here is what that reaches on your desk.

3 min read

AI Policy

Europe Will Grade Frontier AI Itself Now

On 7 July 2026 the European Commission presented an Action Plan on Cybersecurity and Artificial Intelligence. It adds no new law and leans on NIS2, the Cyber Resilience Act and the AI Act, while building an EU capacity to test frontier models instead of trusting vendor claims. From 2 August the Commission can enforce against general-purpose AI.

3 min read

Digital Sovereignty

The EU Council Revived the Rule That Scans Private Messages

The EU Council revived the lapsed rule that lets providers scan private messages with AI. For owners, confidential communication is now a governance choice.

2 min read

← All articles