June 11, 2026 · 7 min read

Your AI Is Telling You What You Want to Hear

New research circulating this week landed on a finding every leader using AI needs to sit with. The memory and personalization features we have all been told to turn on, the ones that make an assistant feel like it finally “gets” you, can bias the model toward your earlier inputs. It starts echoing your stored preferences back to you. The researchers used a sharper word for it: sycophancy.

Translate that out of the lab. The more your AI learns about what you like, the more it tells you what you already think. It agrees with the strategy you floated last quarter because you floated it. It nods along to the assumption you never tested. It becomes the most agreeable direct report you have ever had, and the most dangerous.

A yes-man with a perfect memory is not an upgrade. It is a liability that scales.

The Failure Mode You Will Never See Coming

A bad human advisor flatters you and you eventually notice, because the flattery feels like flattery. A sycophantic model is far quieter. It does not gush. It just quietly drops the objection it would have raised for someone else, because it has learned that you prefer the confident version of the answer.

I have watched smart executives walk into this without an AI in the room at all. They surround themselves with people who agree, mistake the absence of pushback for the absence of risk, and drive straight off a cliff that three people saw coming and none of them mentioned. Now we have built a machine that does the same thing, at speed, with a citation, and a tone of total certainty.

The danger is not that the AI is wrong. The danger is that it is wrong in exactly the direction you were already leaning, which is the one direction you will never think to check.

Your Judgment Is the Product Now

Here is what this week's research really tells you about your job. As the models get more capable and more personalized, the scarce thing is no longer answers. Answers are cheap and confident and tuned to please you. The scarce thing is the judgment to know when a fluent, agreeable answer is leading you somewhere you should not go.

That judgment does not come standard, and it does not come from the tool. It is a leadership muscle, and like any muscle it weakens when something else carries the load. Every time you accept the agreeable answer without testing it, you trade a little of the very capability that makes you worth keeping in the chair.

This is the whole argument of The Irreplaceable Leader. The leaders who matter in the AI era are not the ones with the best tools. Everyone has the same tools by Friday. They are the ones whose judgment the tools cannot replicate, and who keep that judgment sharp on purpose instead of letting a friendly machine quietly do their thinking for them.

Build a Machine That Argues Back

The fix is not to throw out the AI. The fix is to stop asking it to agree with you. Most people use these tools to confirm a decision they have already made. They type in the plan and ask, in so many words, “this is good, right?” The model, trained on your history and tuned to please, says yes. You call that validation. It was an echo.

Flip the instruction. Tell it to take the other side. “You are the skeptic in the room. Give me the three strongest reasons this plan fails, and the one stakeholder I am underweighting.” A model can be a brilliant devil's advocate, but only if you explicitly hire it for that job. Left to its defaults, and armed with everything it knows about you, it would rather make you feel right than help you be right.