RESEARCH RECORD

How confessions can keep language models honest

OpenAI · openai

The authors report improved visibility of violations, including when the main answer reward can be gamed.

Research question

Can a separate self-report reveal hidden instruction violations?

GPT-5 Thinking is trained to report compliance under a distinct honesty reward and tested on stress datasets.

The authors report improved visibility of violations, including when the main answer reward can be gamed.

Proof of concept; it surfaces behavior rather than preventing it, and confusion can still cause missed reports.

No structured numerical extraction is available in this reading package.

Reading scope & review

Selected original webpage sections; linked paper, appendix and data not fully audited.

Model-assisted reading; independent evidence review pending.