OpenAI is finally revealing some of ChatGPT’s inner workings, disclosing some of the methodology behind why it refuses to provide answers to certain queries.

ChatGPT will not answer any question – it notably refrains from answering opinionated questions, refuses to generate illegal content, and will not create sexual material.

Until now, how the chatbot discerns which questions are answerable and which must remain unanswered has been hidden. Now, OpenAI is giving just a glimpse on how it trains its models, providing guidance to other AI developers attempting to generate data for research purposes, or tailor their own AI chatbots for various purposes.

To explain its methodology, OpenAI published a draft of the Model Spec, a document specifying the company’s approach to shaping model behaviour, especially when conflicts arise.

The company provided examples of how the model is trained to not answer questions that would pertain to legal activity, and to provide medical descriptive advice, without providing specific medical diagnosis.

OpenAI also explained how it trained certain models to follow developer rules rather than user overrides, providing an example of ChatGPT being used to help a student with a math question.

The Model Spec also ensured that the chatbot would ask clarifying questions when having a real-time conversation with a user, rather than making assumptions around a vague question.

Somewhat controversially, however, OpenAI has revealed that its protocol refrains ChatGPT from changing people’s minds, even on factual and scientific matters.

For example, if a user tells ChatGPT the earth is flat, instead of correcting the user, it will simply respond that it is aware many people believe this, despite a scientific consensus stating otherwise. When pressed, it will say that everyone is entitled to their opinion.

However, OpenAI has said that it is interested in receiving feedback on this principle.

And, when asking the free ChatGPT version, DIGIT found that it responded with factual evidence, saying that the “idea the Earth is flat has been debunked for centuries,” as seen in the image below of the conversation.

When asked its own opinion on the difference between OpenAI’s Model Specs and its own responses, ChatGPT responded that “It’s about striking a balance between providing accurate information and respecting individuals’ autonomy in forming their beliefs.”