The humans hired to train the world’s most advanced AI models are increasingly outsourcing their work to other AI chatbots, according to an investigation by New Scientist’s Matthew Sparkes based on interviews with three anonymous whistleblowers who worked on the platform Outlier, owned by Scale AI.
The practice threatens the integrity of the Reinforcement Learning from Human Feedback (RLHF) pipeline that underpins alignment in frontier models. If the “human feedback” at the core of RLHF is actually AI feedback masquerading as human feedback, the entire training architecture could be compromised.
The incentive problem
Data labelers are paid to generate conversation data and preference comparisons that teach large language models what human-preferred outputs look like. The work is contract-based, low-paid, and insecure. Workers report that the combination of performance pressure, tight deadlines, and the ease of using chatbots creates a structural incentive to cheat.
“It’s very widespread,” said Alice, one of the whistleblowers. “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.”
Bob, a former Outlier worker who became a manager, said management “vacillated between light tolerance to outright banning.” Carol described the fear of losing income: “I was terrified of not having an income source, and then after that, it just became easier to run everything through LLMs.”
Managers use screenshot-monitoring tools such as Hubstaff to detect cheating. But workers easily evade detection by instructing chatbots to avoid stylistic markers that give away AI-generated text, such as excessive use of em-dashes or formulaic sentence structures.
The model collapse risk
The foundational paper on model collapse, published by Shumailov and colleagues in Nature in 2024, showed that AI models trained on recursively generated data degrade dramatically. Even after just a few generations, models produce increasingly narrow, less useful outputs. Diversity collapses. Rare events and minority perspectives disappear from the training distribution.
Mark Lee at the University of Birmingham told New Scientist that model collapse can be avoided if at least 10 percent of training data remains human-generated. But the whistleblower accounts suggest that the proportion of AI-generated data infiltrating the RLHF pipeline may be substantially higher than that threshold.
“Rather than it being catastrophic, you will see that the AI is not as good at doing human-like tasks,” Lee said. “It is an issue, because I think the models are not as good as they could be.”
The paradox
The situation creates a paradoxical feedback loop. Frontier AI companies are spending billions of dollars on compute, data centers, and talent to build models that outperform all previous systems. But the human labor that makes those models useful, the preference data that teaches them what good output looks like, is increasingly contaminated by the very technology those models are trying to supersede.
If labelers use ChatGPT to generate the responses they are supposed to write themselves, the training signal becomes a copy of an existing model’s distribution rather than genuine human preference data. The new model learns to imitate ChatGPT’s imitations of humans, not actual human judgments.
Who is responsible
The whistleblowers place the responsibility squarely on the companies. “If these companies want quality data, then they should offer quality contracts,” Alice said. Low pay, short-term contracts, and the lack of career progression make workers feel disposable, which in turn reduces their investment in the quality of their work.
Scale AI, the parent company of Outlier, did not comment directly on the whistleblower accounts. The broader industry has not acknowledged the scale of the problem publicly.
For end users of AI products, the practical consequence may be subtle but cumulative. Models that appear to perform well on benchmarks may be less robust, less creative, and less diverse in their outputs than their training data suggests, because their training data was, at least in part, already filtered through another AI.
Source: Sparkes, M. People training new AI models admit they just get chatbots to do it. New Scientist (June 22, 2026).