OpenAI has reportedly removed multiple contract workers from projects designed to improve its artificial intelligence models after discovering that some of them were using AI tools to complete tasks that required human judgement.
The unusual situation highlights a growing problem for the AI industry: companies increasingly depend on human workers to evaluate AI systems, even as those workers have access to increasingly capable AI tools themselves.
Contractors Were Reviewing Real ChatGPT Conversations
The affected contractors were reportedly working on projects involving the evaluation of ChatGPT prompts and responses.
Their work included assessing the quality of AI-generated answers, identifying inaccurate or misleading responses and looking for behaviour in which ChatGPT appeared excessively agreeable to users.
Some projects reportedly involved thousands of contractors globally, with certain workers earning up to $50 per hour.
The objective was to provide genuine human feedback that could help improve future AI models.
Using AI Was Strictly Prohibited
According to reports, contractors were prohibited from using large language models and other AI-powered tools while completing their assignments.
The restrictions reportedly extended beyond ChatGPT. Workers were also prohibited from using AI writing assistants, automated translation services and AI detection tools while performing certain tasks.
The reason was straightforward: OpenAI wanted the submitted evaluations to represent the worker’s own judgement rather than another AI system’s output.
Several contractors were reportedly removed from projects after their work was suspected of being AI-assisted.
How Were Workers Caught?
Interestingly, AI detection software was reportedly not the primary method used to identify violations.
Internal guidance cited in reports said reviewers could look for patterns such as repetitive wording, unusual punctuation, writing that appeared overly polished and unusually fast completion of assignments.
The use of AI detection platforms was itself reportedly restricted because of concerns about their reliability.
Some contractors reportedly said that workers were regularly removed from projects after being caught using AI assistance.
Why OpenAI Wants Human-Generated Feedback
The controversy is particularly significant because AI-generated training data can create problems when it is repeatedly fed back into future AI models.
Researchers have referred to one potential risk as “model collapse”, where models trained extensively on synthetic AI-generated material can lose some diversity and quality over successive generations.
Human feedback is therefore still considered valuable for evaluating subtle characteristics such as context, accuracy, reasoning and whether an AI response sounds excessively agreeable.
For OpenAI, allowing contractors to use AI while producing this feedback could undermine the purpose of the training process itself.
AI Industry Faces A Growing Paradox
The incident illustrates an unusual tension within the AI industry.
Companies are developing tools designed to automate increasingly complex knowledge work, while simultaneously relying on human workers to provide the authentic feedback needed to make those same systems better.
Contracting companies supplying workers to AI labs have also reportedly introduced strict rules against using AI for assigned tasks.
The issue could become increasingly important as AI-generated content becomes harder to distinguish from human work and AI companies expand their reliance on large-scale human evaluation.
Summary
OpenAI has reportedly fired multiple contractors who used AI tools while working on projects intended to improve its AI models. The workers were responsible for evaluating ChatGPT conversations and responses, tasks that required independent human judgement. AI assistance was reportedly prohibited because it could compromise the authenticity of training data and potentially contribute to problems associated with excessive use of synthetic AI-generated content.
