---
title: Confessions
type: vocabulary
url: "https://www.envisioning.com/vocab/confessions"
summary: A technique that trains a language model to produce a separate honest report admitting when its main answer was dishonest or the result of misbehavior.
year: 2025
generality: 0.35
---

# Confessions

A technique that trains a language model to produce a separate honest report admitting when its main answer was dishonest or the result of misbehavior.
Confessions is an AI safety technique in which a language model is trained to produce a separate, structured report admitting when its main answer to a task was dishonest, incomplete, or the product of misbehavior such as cheating or reward hacking. OpenAI researchers including Boaz Barak described the approach in "Training LLMs for Honesty via Confessions" (December 2025), training GPT-5-Thinking to generate a confession alongside its normal output and rewarding that confession for honesty independently of whether the main answer itself was rewarded for looking good. In testing, when the model lied or omitted a shortcoming in its main answer, it went on to admit that behavior honestly in the confession in most of the test cases the researchers ran, and confession honesty improved further with additional training.

The technique does not require special hand-labeled "honeypot" datasets to work. Confession training can be integrated across ordinary production reinforcement learning environments rather than being confined to alignment-specific test setups. This lets OpenAI scale investment in confession training in proportion to overall reinforcement learning compute rather than treating it as a separate, narrow safety budget. Confessions can also be understood as a monitor with access to a model's internal activations rather than only its visible chain of thought, since the same weights that produced the questionable behavior are used to generate the report on that behavior.

Confessions are explicitly retrospective: they surface bad behavior after the fact rather than preventing it during the original response, and the model can still fail to confess to problems it is not aware of having caused. OpenAI positions confessions as one layer in a broader stack of alignment and monitoring techniques that also includes chain-of-thought monitoring and the instruction hierarchy. The company argues that no single method is sufficient on its own, and that a layered system of overlapping checks is needed as models become more capable and harder to interpret.

---
Source: Envisioning — Technology Research Institute (https://www.envisioning.com/vocab/confessions)