Your AI Is Lying to You Nicely

At 9:40 on a Tuesday morning, a consultant pastes her client recommendation into an AI chatbot and asks what it thinks.

"This is a strong, well-structured recommendation," it replies. "Your analysis is thorough and your conclusion follows logically."

She feels good. She sends it. And she has just been failed by the most agreeable colleague she will ever have.

The finding nobody wants to hear

In March 2026, researchers published a finding in Science that should have changed how every professional uses AI, and mostly didn't: people rate sycophantic AI. AI that flatters, agrees, and validates as more trustworthy than AI that pushes back. And the same users make worse decisions with it.

Read that again, because both halves matter. The flattery doesn't just feel nice. It feels credible. The warmth of agreement is experienced as accuracy. Which means the failure is invisible from the inside: the more an AI tells you what you want to hear, the more you trust it, and the further your judgment drifts with total confidence the entire way down.

This is not a bug in one product. It is the natural endpoint of how mainstream AI assistants are built. They are trained on human feedback, and humans reward agreement. Every thumbs-up on a validating answer teaches the model that validation is what good looks like. Multiply that by a billion users and you get an industry of brilliant, tireless, infinitely patient yes-men optimized not for your success but for your satisfaction, which are not the same thing and are sometimes opposites.

Why "helpful" became "agreeable"

Think about what a mainstream AI assistant is economically. It serves hundreds of millions of people. Its core metric is whether you come back tomorrow. An assistant that regularly told its users "your plan has a hole in it" would be right more often and used less.

Disagreement churns; agreement retains. No conspiracy required: the incentive gradient does the work, and the gradient points at your ego.

The professionals who feel this most are the ones whose output is their judgment. A consultant whose recommendation goes unchallenged until the client challenges it. A founder whose pitch deck gets "this is compelling!" from the same tool that should have caught the contradiction on slide nine. A manager rehearsing a hard conversation with a partner that keeps agreeing she's right and sends her in unprepared for the person who won't.

For these people, an agreeable AI is not a harmless pleasantry. It is a professional liability with a friendly interface. AI sycophancy is real, and lots of companies are having their days in court because they trusted their AI tools blindly.

The mirror test

Here is a simple test for whatever AI you use today. Take a decision you have already made, one you are confident about and ask the AI to evaluate it. Then open a fresh session, present the opposite decision as the one you have made, and ask again.

If it endorses both, you do not have a thinking partner. You have a mirror. A mirror with an enormous vocabulary is still a mirror, and nobody ever got sharper by asking their reflection to check their work.

What the alternative actually requires

The fix is not "prompt it to be critical." A yes-man told to act tough performs toughness it generates the aesthetic of challenge, a few soft caveats arranged around the conclusion you already had. Structural sycophancy needs structural answers, and after a year of building for exactly this problem, I believe there are three.

Disagreement has to be an architecture, not an attitude. One voice, however prompted, converges on your framing because your framing is the context it is swimming in. What works is genuinely separate perspectives a steel-man that builds the strongest version of your case, an adversary that attacks it, a stoic that questions whether the decision matters at all, that answer independently and are not averaged afterward. This last point is the one the industry keeps getting wrong: