AI trainers are outsourcing their jobs to AI

Training frontier AI requires mountains of clean human data. But as it turns out, the very gig workers hired to clean up that data are secretly outsourcing their jobs to ChatGPT.

A team of researchers from the Swiss Federal Institute of Technology (EPFL) found that a massive chunk of data annotators on Amazon Mechanical Turk are using AI to complete their tasks.

Catching the copy-pasters

Tech companies regularly rely on platforms like Mechanical Turk to handle tasks that machines supposedly cannot do well, such as annotating complex text or summarizing documents. These workers are typically paid pennies and expected to turn around massive volumes at high speed. Naturally, turning to ChatGPT is the easiest way to maximize their hourly rate.

To measure the scale of the cheat, the EPFL researchers hired 44 gig workers on Mechanical Turk to summarize 16 medical research papers.

They tracked the workers through two methods: analyzing their keystrokes to see if they were copy-pasting responses from elsewhere, and running a custom-trained detector looking for ChatGPT's telltale linguistic patterns, such as low vocabulary diversity.

The findings were staggering: between 33% and 46% of the workers used large language models like ChatGPT to do their work. The researchers noted that as AI becomes faster and more accessible, this number is almost guaranteed to rise.

Feeding AI its own exhaust

This is not just about workers gaming their hourly pay; it poses a genuine threat to AI quality.

Large language models hallucinate constantly. When an AI generates subtle factual errors and those errors get handed back to train the next generation of models, the mistakes compound.

Oxford computer science fellow Ilia Shumailov pointed out that there is no straightforward patch for this. Artificial data carries statistical errors and misunderstandings that become nearly impossible to trace once baked into downstream models.

EPFL assistant professor Robert West noted that while this will not kill crowdsourcing platforms overnight, it fundamentally shifts the dynamics and forces teams to rethink how tasks are monitored.

Why it matters

We think this study exposes a glaring loophole in the AI supply chain: the industry's reliance on cheap labor is directly polluting its own training data.

For developers and enterprise users, this is a loud warning. If you are fine-tuning models on cheap crowdsourced data, you might actually be training them on ChatGPT hallucinations without knowing it. Paying rock-bottom rates encourages workers to use bots, which creates an incestuous loop of synthetic text.

This will force AI labs to either abandon cheap crowdsourcing altogether or invest heavily in verified, high-paid domain experts. Until platforms develop bulletproof provenance checks, relying on anonymous online workers for ground-truth data looks increasingly like a liability.