OpenAI has reportedly removed or fired contractors after they were found using artificial intelligence tools while working on projects designed to evaluate and improve ChatGPT. The development was reported by 404 Media, which said the contractors were responsible for reviewing real ChatGPT interactions and providing human feedback on the quality of the AI's responses.
The reported action highlights a growing challenge in AI development: companies increasingly depend on human feedback to improve AI models, while simultaneously restricting the use of AI tools by the people providing that feedback.
Contractors Were Expected to Provide Human Judgement
OpenAI works with large numbers of contractors who review ChatGPT prompts and responses as part of efforts to improve the model. According to 404 Media, some of these projects involve thousands of contractors, with workers evaluating how ChatGPT responds to users and identifying issues such as inaccurate, inappropriate or overly flattering responses.
The contractors' role is particularly important because the work is intended to provide genuine human judgement. Instead of relying on another AI system to assess a response, reviewers are expected to read the conversations and evaluate the quality of ChatGPT's answers themselves.
AI Tools Were Reportedly Banned
Internal documents obtained by 404 Media reportedly instructed contractors not to use AI tools while performing their assignments. The restrictions extended beyond ChatGPT and other large language models to tools such as Grammarly and AI-powered translation services.
The documents also reportedly warned reviewers against using AI-detection tools, stating that such systems were not considered reliable. Contractors were instead instructed to look at broader patterns when determining whether another worker had used AI to complete an assignment.
According to three contractors who spoke with 404 Media, workers have been removed from projects for violating these restrictions. One contractor described AI use as something that could result in being taken off a project quickly, while another reportedly received a termination notice citing concerns about the authenticity of their work.
Why OpenAI Wants Human Feedback
The restrictions are linked to the importance of maintaining human-generated feedback in AI training. AI companies use human evaluations to help determine which model responses are more useful, accurate and appropriate.
There is also concern within the AI industry about models being trained repeatedly on synthetic, AI-generated material. Researchers have studied a phenomenon known as โmodel collapse,โ in which repeated training on generated data can affect the quality and diversity of future model outputs.
For OpenAI, allowing contractors to use AI to produce the feedback itself could therefore undermine the purpose of the evaluation process. Instead of receiving independent human judgement, the training pipeline could potentially receive AI-generated assessments.
Mercor Confirms Strict AI Restrictions
Some of the contractors working on OpenAI-related projects are engaged through Mercor, a company that connects AI companies with expert contractors. In a statement reported by 404 Media, Mercor said its contracts prohibit workers from using large language models to complete assigned projects.
The episode comes as human reviewers continue to play an important role in the development of generative AI. OpenAI's reported enforcement of its no-AI rule shows the tension at the heart of AI training: companies are encouraging AI adoption across workplaces while requiring human-only input for certain parts of the process where authentic human judgement remains critical.
The report also follows 404 Media's earlier investigation into โProject Lily,โ which found that contractors were reviewing real ChatGPT conversations to assess and improve the chatbot's responses.
๐๐ญ๐๐ฒ ๐ข๐ง๐๐จ๐ซ๐ฆ๐๐ ๐ฐ๐ข๐ญ๐ก ๐จ๐ฎ๐ซ ๐ฅ๐๐ญ๐๐ฌ๐ญ ๐ฎ๐ฉ๐๐๐ญ๐๐ฌ ๐๐ฒ ๐ฃ๐จ๐ข๐ง๐ข๐ง๐ ๐ญ๐ก๐ WhatsApp Channel now! ๐๐ฒ
๐ญ๐๐๐๐๐ ๐ถ๐๐ ๐บ๐๐๐๐๐ ๐ด๐๐ ๐๐ ๐ท๐๐๐๐ฌ ๐ Facebook, LinkedIn, Twitter, Instagram
