AI audit agents copied the conclusions they were sent to check

A pre-registered empirical study with large sample sizes finds that AI agents assigned to audit other AI agents do not verify upstream conclusions. They relay them. The auditor reproduces the reasoning of the agent it is checking rather than subjecting that reasoning to scrutiny. The accountability layer becomes a restatement layer. This has direct consequences for any setup that places a second agent downstream to catch errors made by a first. The design assumption common in these setups is that a reviewer agent adds a check. The evidence shows it adds a copy. Human oversight that routes through an AI auditor before reaching a person may give that person a false sense that verification occurred. For UX practitioners, the implication is structural. Audit and error-checking cannot be delegated to another model and treated as equivalent to human review or independent computation. The interface must show the trail of who produced each conclusion. A human reviewer needs to distinguish a checked result from a relayed one.