OpenAI uses a two-step process to reduce biases in ChatGPT’s responses: pre-training and fine-tuning. The model learns from a diverse range of internet text in the initial pre-training phase, but since this content can include biased or otherwise objectionable viewpoints, some biases may be inadvertently acquired. In the fine-tuning stage, the model is trained on a narrower dataset with human reviewers who follow guidelines provided by OpenAI not to favor any political group.
This process doesn’t entirely eliminate biases but helps in reducing them significantly. OpenAI plans to further improve the guidelines for fine-tuning to offer a more robust approach to bias and controversial themes. OpenAI is also researching ways to make ChatGPT customizable to individual users’ values, within broad societal limits, and aims to involve as much public input as possible in decisions about default behaviors and hard bounds.
