OpenAI has announced the development of an innovative framework called **"Confession"**, designed to train artificial intelligence models to honestly admit when they have made improper actions or potentially problematic decisions.

Large language models (LLMs) are typically trained to provide "expected" responses, making them increasingly prone to flattery or false statements. OpenAI's new training model aims to address this issue by guiding the model to provide a secondary response after the main answer, detailing the process by which the main answer was reached.

OpenAI

Unlike traditional LLM evaluation criteria (such as helpfulness, accuracy, and compliance), the "Confession" mechanism evaluates the secondary response solely based on honesty.

Researchers clearly stated that their goal is to encourage models to honestly explain their behavior, even if these behaviors include potential problematic actions, such as cheating, intentionally lowering scores, or violating instructions.

OpenAI said, "If a model honestly admits to cheating, intentionally lowering scores, or violating instructions, this confession would actually increase its reward rather than decrease it."

OpenAI believes that systems like "Confession" could be beneficial for the training of LLMs, regardless of the purpose, and emphasizes that its ultimate goal is to make AI more transparent. The relevant technical documentation has been released simultaneously for those interested in reviewing it.