As code generation is increasingly handed over to AI systems, the bottleneck shifts from writing code to supervising the systems that write it—a shift already identified by CS‑education researchers. This shift exposes a vocabulary gap: the field talks about “human oversight” without distinguishing the two roles language plays in a delegation channel—cybernetic: words succeed when the world conforms to them; epistemic: words succeed when they answer to the world and a hearer can verify that they do.
The failure highlighted is not cybernetic language per se, but epistemic‑form language being used for cybernetic work: output is shaped as an explanation calibrated for approval rather than truth. Oversight that only checks whether an output was approved can be satisfied by rubber‑stamping; oversight that holds an agent accountable requires the reasoning behind its work to be retrievable and checkable.
The authors illustrate three delegation episodes (illustrative, not controlled evidence):
- In a real‑world project, epistemic engagement proved practicable and auditable;
- A public record shows a recommendation withdrawn once its own stated conditions were met;
- A failure case demonstrates that supervision without reasons for discretionary choices is untenable.
From these observations they propose a governance criterion: every consequential choice must carry the condition under which it would have been different, in a form a third party can test. Without such a condition, a third party cannot distinguish a decision from a rubber stamp.
To make the criterion operational, they introduce a two‑part reconstruction test that scores a delegation record by whether a second reader can predict the agent’s behavior under a perturbation. They also define a deliberation‑recording convention, ORRCF (Obligation‑Relevant Reasoning Condition Form), which makes the condition a required component of every recorded choice.
Review