Workers admit using chatbots to train AI models

workers using – Whistleblowers say people paid to supply training conversations and tests for new AI models are increasingly using chatbots like ChatGPT to speed through tasks—often in ways they believe can evade detection. Researchers warn that repeatedly training on AI-gene

For weeks, Alice* sat down to do the kind of work that’s meant to make future AI smarter: high-quality conversations and carefully structured tests. Then she started doing it the easy way.

“Its very widespread; every company I’ve worked for has had explicit guidelines around it and they clearly do try to catch people out, so I think they do care. But I don’t think they can stop it,” she said.

Alice’s shortcut, she explained, was to use ChatGPT to complete training tasks—despite company rules that forbid it. Her defense was disarmingly practical: the output only gets flagged if you’re careless.

“It’s easy to get away with as long as you instruct chatbots to avoid the usual telltale signs of AI output. like a preponderance of em-dashes. ” she said. “It’s only the sloppiest of users that get caught. Anyone with a modicum of awareness around AI hallmarks can tell their output not to use them. and at that point what are you going to do?”.

Alice’s view of the problem isn’t just technical—it’s economic. “If these companies want quality data, then they should offer quality contracts,” she said. “Instead they’re low-balling struggling people. employing them for the barest possible amount of time and tossing them aside as projects are finished with no warning.”.

She said she felt “not in the slightest” guilty about using ChatGPT, framing it less as fraud than as a response to the incentives and uncertainty of the job.

Another worker, Bob*, described a different kind of pressure: he moved from doing the training work to being responsible for catching others doing the same thing.

Bob previously worked for a training platform called Outlier. In his account, he initially carried out tasks using AI in a way he says was illicit. He was then promoted to a leadership role where part of the job was to detect that behavior.

“Management vacillated between light tolerance to outright banning,” Bob said.

Outlier, owned by Scale AI, uses a tool called Hubstaff that takes screenshots of workers’ desktops at random intervals to confirm they are performing tasks as assigned. Bob said he would look for evidence that AI models were open.

“People would have it [AI models like ChatGPT] open in other tabs, or minimised, so obviously we could see it in the task bar,” he said. “Even stuff like folders on their desktop with names gave it [AI use] away.”

Scale AI did not respond to a request for comment. Outlier also did not respond to a request for comment. Scale AI’s website says the company carries out work for technology giants like Meta and Cisco—neither of which responded to requests for comment. Bob also said he had worked on projects for Google, which did not respond to a request for comment.

A third worker, Carol*, has worked across several platforms and described how the rules work only on the surface. In her telling, she began by using AI to check her work for guideline violations—because contravention could mean expulsion from a project and the loss of earnings.

“I was terrified of not having an income source, and then after that, it just became easier to run everything through LLMs,” Carol said.

Carol described a workflow where one model helps produce material and another helps package it into the files needed for the task. “For a lot of the projects that I do now, it’s creating scenarios,” she said. “So I will use one LLM to help me create the scenario and then I’ll use a different LLM to help me create the files that go along with the scenario. I do feel guilty but like I said. in the beginning it was more about trying to make sure I wasn’t making any errors.”.

But the longer she relied on the models, the more uneasy she became. “I do worry that I’m actually making it [AI] worse. I thought using the models to train themselves negates some of the value,” she said.

Her worry lands close to a warning coming from AI research. Mark Lee at the University of Birmingham, UK, said research has shown that AI models “collapse” when they are recursively trained on AI-generated content.

“When this happens, the abilities of the model drop dramatically and they become less useful,” Lee said. The phenomenon is sometimes called AI cannibalism or AI inbreeding.

“That’s the kind of worst-case scenario,” Lee said. “And that’s probably not what’s happening in the real world. There’s still a few humans. And if you have like 10 per cent human data, it mitigates it, it avoids model collapse.”

Even so, Lee said cheating of this kind isn’t consequence-free. “Rather than it being catastrophic, you’ll see that the AI isn’t as good at doing human-like tasks. It’s an issue, because I think the models aren’t as good as they could be.”

The picture that emerges from the accounts is stark: people paid to produce high-quality training data are increasingly relying on chatbots to generate the very material that those systems will learn from. The systems built to prevent shortcuts—guidelines. monitoring tools. desktop screenshot checks—appear to be in an ongoing chase. not a fixed solution.

And for the workers involved, the decision is rarely framed as a single ethical line. It’s tied to low pay, limited time, and the constant threat of losing an income source. For researchers. the fear is more technical but no less serious: less human signal. and models that may end up worse at the tasks they’re being trained to handle.

AI training LLMs ChatGPT whistleblowers synthetic data AI cannibalism AI inbreeding Scale AI Outlier Hubstaff University of Birmingham Mark Lee