Loading page…
The reviewer rules extend beyond chatbots to Grammarly, AI translation and detection tools. OpenAI wants human feedback on its model’s replies.
Listen to this story
OpenAI’s reported dismissals expose a strict boundary in its model-improvement pipeline: contractors hired to judge ChatGPT answers were barred from using AI to assess, draft feedback, translate, or check text with AI-detection tools. The policy reflects the value OpenAI places on independent human judgments, while ruling out even seemingly limited writing assistance. The number of workers affected is unknown, and OpenAI declined to comment. The reporting does not establish that AI-generated contractor feedback was
Reviewer instructions named Grammarly, AI translation tools, and GPTZero; the detector ban was separate, with the instructions calling such tools unreliable.
Contractors assess answer accuracy and whether replies are excessively flattering—not simply whether a response exists.
A 2024 Nature study cited in the coverage found risks from indiscriminate synthetic-data training; it does not show these reviews harmed ChatGPT.
OpenAI reportedly fired contractors who used AI while reviewing ChatGPT interactions and writing feedback. The workers’ instructions prohibited that help, including Grammarly and AI translation tools, according to documents described by 404 Media. The rule protects the human input OpenAI hired reviewers to provide.
Contractors examine user prompts and ChatGPT’s replies, then assess the answers to help improve the model. The work includes judging accuracy and whether a response is excessively flattering, according to Gizmodo’s account of 404 Media’s reporting. It is not simply a check that a reply exists; reviewers must decide how well it serves the person who asked.
That distinction explains why an AI-assisted review is not an ordinary productivity shortcut. If a tool supplies the assessment or drafts the comments, OpenAI gets less of the independent human feedback it assigned the work to collect. Using a tool to polish wording raises a harder boundary question, but the reported instructions draw no exception for it.
Reviewers may not use AI either, including Grammarly and AI translation, to review, write feedback, or write comments.
Reviewer instructions quoted by Computerworld from 404 Media’s reporting
The instructions also forbid AI-detection tools, naming GPTZero and saying such tools are unreliable. That is a separate restriction from the ban on using AI to review or write. Together, the rules leave reviewers without several common kinds of automated help, whether they would use it to generate an assessment, translate text or check its origin.
Synthetic feedback also raises a longer-term concern: model collapse, in which training on AI-generated material can degrade later models. Gizmodo cited a 2024 Nature study finding that indiscriminate use of model-generated content in training caused irreversible defects. That research explains the risk; it does not show that these contractors’ work damaged ChatGPT.
The number of contractors fired has not been disclosed. Computerworld says OpenAI declined to comment on the situation. For now, the reported firings show how firmly OpenAI applies its reviewer rule, not how much AI-written feedback entered its improvement work.
Loading discussion...
Join the conversation
Where would you draw the line between assistance and replacement?
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.