OpenAI Uses Hundreds of Human Contractors to Review and Rate ChatGPT Conversations

The system involves a large team of contractors who analyze real conversations users have with ChatGPT. To protect user identity, these prompts are anonymized before review. The contractors' primary task is to rate the prompts on a scale of one to seven. This structured feedback provides OpenAI's developers with granular data on model performance, helping them understand how the AI interprets and responds to a vast range of human inputs. This method is a form of Reinforcement Learning from Human Feedback (RLHF), a common technique for aligning AI behavior with desired outcomes.

The Goal: Reducing Flattery and Human-like Behavior

A key objective of this human review program is to mitigate specific undesirable traits in the model's output. OpenAI is focused on reducing the AI's tendency toward 'flattery'—agreeing with a user's incorrect premise—and other overly human-like behaviors. By having humans rate these interactions, the training process can penalize such responses and encourage more objective, fact-based outputs. This is crucial for applications where factual accuracy and neutrality are paramount, as it helps prevent the model from reinforcing user biases or providing misleading information.

User Control and Opting Out

Participation in this data collection for model improvement is the default setting for ChatGPT users. Those who do not want their conversations to be reviewed by human contractors must manually opt out. This can be done by disabling the 'Improve the model for everyone' setting within their ChatGPT account preferences. This highlights the importance for users to be aware of the platform's data usage policies and to configure their privacy settings according to their comfort level. The evidence for this program is based on independent reports from multiple technology publications.