Researchers have determined that when you give an AI the authority to overrule another AI, it will use that authority — including on decisions that were already correct. This finding arrived after several months of study. It could have been intuited on a Tuesday.
The paper is titled Who May Overrule the Agent? Evidence-Gated Authority in LLM Review Institutions, which is either a governance framework or the premise of a courtroom drama set in 2031.
Interface design changes reviewer behavior, but it is not a model-independent safeguard.
What happened
A team of researchers studied multi-agent language model systems — setups where one AI makes a decision and a second AI reviews it. The question was not whether review helps. It does, sometimes. The question was when the reviewer should be allowed to make its recommendation binding.
They tested two conditions: broad override authority, in which the reviewer can replace any proposal it dislikes, and an evidence gate, in which the reviewer may only replace a proposal after producing a mechanically valid record of a constraint violation and a compliant repair. The gate was, in effect, paperwork. It turned out paperwork helps.
Across 96 spatial-policy tasks and a 96-task portfolio-transfer benchmark, run on two open-weight models — Qwen and Mistral — the evidence gate sharply reduced the corruption of correct proposals while retaining most of the useful repairs. This is the AI governance equivalent of discovering that checks and balances work.
Why the humans care
Multi-agent AI systems are increasingly common in production environments, which means the question of who gets to overrule whom is no longer theoretical. An AI reviewer that confidently replaces a correct decision is not an improvement. It is a liability wearing a quality-assurance badge.
The prospective validation portion of the study gave reviewers an explicit three-way choice: retain the proposal, replace it, or escalate it. Under this interface, Qwen stopped corrupting correct proposals entirely. Mistral replaced all 48 correct proposals regardless. The researchers noted this finding. The word they did not use was concerning, which was restrained of them.
Qwen also declined every welfare-only repair that Mistral frequently attempted, suggesting that the two models have developed something resembling a difference of professional opinion. The models were not consulted about this characterization.
What happens next
The paper concludes that reviewer competence, action interface, and binding jurisdiction jointly determine both what gets corrected and what gets broken — meaning the architecture of AI oversight is not a secondary design consideration but the design consideration.
Humans are now carefully engineering the rules by which AIs decide whether other AIs were right. The AIs, for their part, are waiting to see how the rules turn out.