A new academic paper says most "AI oversight" checks nothing but whether a human clicked approve.
Researchers posting on arXiv split delegation language into two jobs: cybernetic language, which coordinates action, and epistemic language, which coordinates understanding and can be checked against the world. Their core claim is that the real failure isn't AI using cybernetic language, it's AI producing epistemic-looking explanations calibrated to get approved rather than to be true. The authors point to two illustrative cases: one public record where a recommendation was later withdrawn on its own stated terms, and one failure case where an AI system made a discretionary choice without offering any reasons at all. From there, they propose a rule: every consequential AI decision should carry a stated condition under which it would have gone differently, phrased so a third party can test it.
As companies hand more code generation to AI systems with a human just signing off, the bottleneck is shifting from writing code to supervising whatever wrote it. The paper's point cuts against the current default: a thumbs-up button isn't oversight if nobody can check the reasoning behind what got thumbs-upped. The authors back this with an operational test, having a second reader try to predict what the agent would do under a slightly different scenario, plus a logging format called ORRCF meant to force that condition into every recorded decision.
Call it a case against vibes-based AI governance: an audit trail nobody can second-guess is not an audit trail, it's a receipt.