A growing number of individuals employed to train the next generation of artificial intelligence models are reportedly circumventing their duties by using existing chatbots to generate the necessary conversational data. This revelation, brought to light by whistleblowers speaking to New Scientist, suggests a widespread practice that could have significant implications for the quality and utility of future AI systems, potentially leading to what some experts term "model collapse."
The current generation of large language models (LLMs) predominantly relies on vast datasets scraped from the internet. However, as these models become more sophisticated and demand even greater volumes of high-quality training data, AI firms have increasingly turned to human trainers. These workers are paid to engage in conversations and tests with nascent AI, with the explicit aim of producing superior data to enhance the intelligence and functionality of future LLMs. However, multiple sources indicate that many of these human trainers are instead feeding prompts into established chatbots, such as ChatGPT, and submitting the AI-generated responses as their own work.
The motivation behind this shortcut appears to be largely economic. Many trainers are employed by third-party companies, often on short-term contracts and for low wages, rather than directly by the major AI developers. This precarious employment situation incentivises workers to complete tasks as quickly as possible, leading them to bypass the intended human interaction. One worker, identified as 'Alice', stated that the practice is "very widespread" and that while companies attempt to catch offenders, it is relatively easy to avoid detection by instructing chatbots to steer clear of common AI linguistic tells.
Another individual, 'Bob', who initially used AI illicitly for training tasks, was later promoted to a leadership role where he was responsible for identifying others doing the same. He described how management's approach varied from "light tolerance to outright banning" and detailed methods used to track workers, including desktop screenshots. Despite these measures, trainers found ways to conceal their use of AI. 'Carol', another worker, admitted that her initial use of AI was to ensure compliance with stringent guidelines, fearing loss of income, before it became a routine shortcut. She expressed guilt and concern that she might be inadvertently degrading the AI she was meant to improve.
The potential consequences of this "AI inbreeding" are significant. If future AI models are trained predominantly on data generated by existing AI, rather than genuinely human-created content, there is a risk of diminishing originality, introducing subtle biases, and ultimately reducing the models' ability to understand and interact with the real world effectively. This feedback loop could lead to a stagnation or even degradation of AI capabilities, undermining the substantial investments made in the sector.
For UK businesses, this trend could mean that the advanced AI tools they plan to integrate may not perform as expected, impacting productivity gains and innovation. Consumers might encounter less useful or more error-prone AI applications. From a regulatory perspective, bodies like the UK's Information Commissioner's Office (ICO) and the forthcoming EU AI Act may need to consider new guidelines on data provenance and quality assurance in AI training to safeguard against such practices. Experts highlight the need for AI companies to offer better compensation and working conditions to ensure data quality, alongside developing more robust methods for detecting AI-generated training data.