Several contractors working on OpenAI’s AI training projects have reportedly been fired or removed from assignments after using artificial intelligence to complete work they were hired to perform, according to an investigation by 404 Media.
The development has drawn attention because some of the affected workers were hired specifically to provide human feedback on AI-generated responses, making human judgment an important part of the training process.
404 Media reported that it obtained internal documents and spoke with three contractors working across different OpenAI-related projects. One internal document indicated that some of the projects can involve more than 10,000 contractors.
The contractors’ work can include reviewing AI-generated responses, evaluating their quality and providing feedback that is used to improve OpenAI’s models.
However, internal instructions reportedly prohibit the workers from using AI themselves while carrying out those tasks.
One document reviewed by 404 Media instructed contractors not to use AI tools, including AI-based writing, translation and detection services, when reviewing or commenting on other contractors’ work.
The reason is straightforward: if an AI system generates the feedback that is supposed to represent human judgment, the resulting material could undermine the purpose of the training process.
The contractors are also reportedly trained to identify signs that another worker may have used AI. These indicators can include repetitive language, unusual punctuation patterns and unusually fast completion times.
One contractor told 404 Media that people were being dismissed for using AI and described it as one of the quickest ways to lose the job.
“In a group of thousands there are tons that have been caught.”
Another contractor reportedly used AI while working on OpenAI’s training projects and later shared a termination letter with 404 Media. The letter said the employer had identified concerns about the “authenticity” of the contractor’s work.
The contractor told the publication that they had turned to AI for assistance and that doing so ultimately led to their removal.
The restrictions highlight a broader issue facing companies developing large language models. Human-generated feedback is used to evaluate and improve AI systems, but the growing availability of generative AI makes it increasingly difficult to guarantee that every piece of supposedly human feedback was actually produced by a person.
The concern extends beyond OpenAI. Researchers have warned about the possibility of model collapse, a phenomenon in which models trained repeatedly on AI-generated material can lose some of the characteristics and diversity found in human-generated data.
404 Media separately reported that OpenAI uses contractors to review real ChatGPT conversations and rate the quality of the chatbot’s responses. The publication said those conversations can sometimes contain sensitive personal information, although OpenAI says it attempts to remove personal information before prompts reach reviewers.
That human review process is intended to give OpenAI additional information about how its models respond to real-world requests and where their answers can be improved.
The reported restrictions therefore create an unusual situation. OpenAI encourages people to use AI across workplaces, while some workers helping develop its own models are prohibited from using AI when performing their specific training duties.
Mercor, an AI-training company involved in employing contractors working on OpenAI projects, has confirmed to 404 Media that its contracts prohibit workers from using large language models to complete assigned projects.
The reported dismissals do not mean OpenAI has stopped using human contractors to improve its models. Rather, they demonstrate how companies are attempting to preserve the distinction between human feedback and AI-generated feedback in a training pipeline that increasingly relies on both humans and machines.
How did this story make you feel?
Join the conversation
Share a thoughtful response. Submissions may be reviewed before appearing.
