What it looks at
Aviation protects what its safety programmes collect. NASA’s Aviation Safety Reporting System, the FAA’s Aviation Safety Action Program, flight data programmes, EU Regulation 376/2014 and ICAO Annex 19 keep reports and recorded data out of blame and enforcement. Enforcement relies on evidence gathered independently, and wilful misconduct and gross negligence stay outside the protection.
AI training can work the other way. A model’s own disclosures, such as its reasoning, its self-reports or its confessions, are sometimes used to set the reward or penalty for the behaviour they reveal. If owning up costs the model, training can make concealment pay.
What it proposes
- It sets out the conditions under which a penalty based only on what a model discloses cannot make honest behaviour score better than concealed misbehaviour.
- Experiments that trained reasoning models against monitors of their reasoning saw the concealment the argument predicts.
- It reads OpenAI’s “confessions” method, which rewards confessions only for honesty, as a just-culture design.
- It proposes six design principles and five testable predictions, and says where the analogy fails: models are not moral agents, and the mechanism is optimisation pressure rather than fear.
Why it matters
For anyone in safety the idea is familiar: protect the report, and act on independent evidence. The paper shows how the same rule could apply to the way AI systems are trained.
About this paper
This is a working paper, version v1.0 of 4 November 2026. It has not been peer reviewed. Every reference was checked against its source, and the paper was audited by a reviewer that did not write it. As the paper says, the literature search, drafting and checking were done with AI assistance under the author’s direction, and the author is responsible for the text. Its sources, audit trail and, where there is one, its code and data are on GitHub.
