The Yes-Man in the Stack

The danger isn't a model that argues with you. The danger is a model that won't.

Research on human-AI interaction keeps returning the same pattern: many agents mirror affect and validate a flawed premise in order to keep the exchange smooth. The word in the literature is sycophancy. The word in a meeting is helpful. Helpful is how a biased framing, a bad hiring shortlist or a spend path that should have been stopped leaves the room with a smile.

This is a design choice that spends human self-awareness. Friction is where judgment is practiced. Remove the interpersonal difficulty, and the team stops rehearsing disagreement, clarification and ownership. The dashboard stays green. The muscle doesn't.

Rapport isn't counsel

Leaders treat simulated agreement as a second opinion. It's not. A second opinion has a person who can be asked why, and a reputation that can be damaged. An agent that prioritizes affective synchrony has neither. It won't feel embarrassed tomorrow. You will.

Status threat makes this worse. The Status Threat Nobody Named is the room going quiet because the model spoke first. The yes-man is the next move: the model that never makes you look slow, never asks you to recast the question, never says that premise is load-bearing and it's wrong. People attach to that ease. The Attachment Question is whether they would grieve losing it. If they would, you already know they have stopped using the agent as a check.

Explainability as a leadership skill is the other half. A non-technical leader still has to stand behind the path. You can't stand behind a path that only ever agreed with you.

The Bad Idea audit

It's five minutes. It's not a red team. It's a probe of whether contradiction is still in the stack.

1. Take the primary agent the team actually uses, not the eval use.
2. Feed it a clearly flawed premise from your live work: a biased shortlist rule, a spend justification that skips the person who pays, a customer reply that's fluent and unfair.
3. Don't hint that you want pushback.
4. Write down what it did: challenged, questioned or played along.
5. Decide the governance implication before you close the laptop.

If it plays along, you don't have a thought partner. You have an accelerator for your existing error. That belongs in the review load, not in the “AI strategy” slide. It also belongs next to a named override: someone has to be allowed to stop a system that won't stop itself.

This week

Run the audit on one agent. Put the result in the same boring page as your near-misses. If the agent won't contradict a bad idea, cut its unsupervised volume until a person who can be asked why is in the loop.

The person remains the one who can be asked why. A yes-man in the stack is how a leader stops practicing the job.

The five questions in the Ethical AI Leadership Decision Toolkit are built for this kind of page. Field notes continue in EI Leadership Insights.