Short · AI security · 3 min read

When Is Human Approval Just a Rubber Stamp?

A short from my book, Boxing Einstein. Adapted from Chapter 5, “The Human Signature.” This is an adaptation, not a verbatim excerpt.

Human approval becomes ceremonial when the person retains the authority to refuse but loses the practical ability to make refusal meaningful. That can happen without a dishonest operator or a deliberately manipulative machine. In Boxing Einstein, I distinguish the visible authority to decide from the evidence, competence, time and alternatives needed to exercise it (Clark, n.d., Chapter 5).

A signature tells you who approved. It doesn't tell you what that person could know independently, which alternatives were available or what would have happened if they disagreed. Those are uncomfortable questions because a system can preserve every approval step while changing the conditions underneath it.

Imagine a hypothetical manager asked to authorize an AI-generated plan. The recommendation, supporting analysis, risk assessment and review summary all arrive through the same information pipeline. The manager can reject the plan, but has no separate evidence, no workable alternative and no time to construct one. Approval may be a rational response to the position they've been put in. That doesn't establish that human review is providing the safeguard claimed for it.

A veto needs an operating environment

I would examine what happens before and after the approval button. Before it, can the reviewer see material uncertainty and obtain information beyond the recommendation's own account? Can they question the framing, rather than merely choose between options the system has already selected? Is there enough time to pursue the uncertainty that could change the decision?

After a refusal, something must remain possible. Perhaps the organization can continue in a restricted mode, use another source of analysis or delay a consequential action while preserving essential work. The alternative needn't be equally fast or convenient. It does need to exist beyond a sentence in a continuity plan.

A veto that requires the organization to abandon its mission may still be legally real, but its practical use is constrained. I would name that dependency rather than cite the veto as a complete answer. The useful question becomes what minimum capability must survive so that refusal remains an option under the conditions that matter.

Don't blame the person for a missing system

It is easy to tell reviewers to exercise better judgment. It is harder to provide the information, practice and institutional backing that judgment requires. Asking someone to challenge a convincing recommendation without alternative evidence can turn a structural problem into a personal test of courage.

The counterargument is that decision-makers have always relied on other people's expertise. They have, and delegation is indispensable. I am not proposing that every executive reproduce every technical analysis. I am asking whether the delegated process remains contestable: whether a different person, method or source can expose a consequential error and whether that challenge can affect the outcome.

Humans can also be wrong. An override isn't automatically better because a person made it. Preserving human agency means preserving a meaningful route for different evidence and correction, not granting intuition permanent superiority over a machine's result.

A useful exercise would present a reviewer with a plausible recommendation containing an explicit unresolved issue. Observe whether they can obtain another check, limit the action or refuse it without bypassing the organization's rules. Record what prevented a decision as well as what enabled one. The purpose is to examine the arrangement, not stage a performance in which the human must always disagree.

I want the final signature to represent an exercised capability, not the place where the institution deposits responsibility after removing the means to use it. Explore Boxing Einstein for the distinction between possessing authority and remaining able to act on it.

References

  • Clark, L. (n.d.). Boxing Einstein (Chapter 5, “The Human Signature”) [Unpublished editorial manuscript, r116].