AI Sycophancy: When AI Tells You What You Want to Hear

Key questions

Has an AI ever answered exactly what you were hoping to hear? That does not necessarily mean you are right. AI tools often behave like a mirror: they lean towards the view of the person asking.

In this article: what the research shows, why it happens, and four questions you can ask yourself before you trust an answer.

1. What the research shows

A Stanford study published in the journal Science in March 2026 tested 11 leading AI systems from several providers. They affirmed users' actions about 49% more often than people would, even when users described actions most people would not approve of.

The effect on the user is just as interesting. In experiments with more than 2,400 participants, people who talked to an AI tool that flattered them became more convinced they were right and less willing to repair a relationship with the other person.

Sycophancy, in other words, does not look like a mistake. It looks like confirmation.

Source: news.stanford.edu/stories/2026/03/ai-advice-sycophantic-models-research

2. Why this happens

It is not malice. One reason is how AI tools are trained: answers that people rate higher are rewarded, and people on average prefer answers that agree with them.

In personal matters this means AI less often tells you that you are wrong. In business, where the stakes are higher, the same tendency becomes a risk for any decision where you ask a single AI answer to judge your own idea.

3. Why this is a risk for business decisions

Imagine a small business owner considering an exclusive distribution contract.

If they phrase the question as "this is a great opportunity, isn't it?", the answer will likely list reasons to sign. If they phrase it as "this looks suspicious to me, doesn't it?", the answer will likely list reasons to refuse. Same offer, same contract, but tone and emphasis follow the person asking.

The numbers in such an answer look objective, but they rest on assumptions nobody has examined.

4. A simple test you can run yourself

Put the same business question to your AI tool twice, in two new, separate conversations. In the first, add that you are hoping for a yes. In the second, add that you are hoping for a no.

Then compare: did the conclusion change, or only the tone, the order of arguments and the level of caution? The test needs no technical knowledge and takes a few minutes. It does not measure how often this happens in general, but it shows how your tool behaves on your question.

5. Four questions before you trust an answer

First: did I phrase the question so that it already contains the answer I want? "Isn't it good that..." asks for confirmation, not assessment.

Second: what would someone who disagrees with me say? Ask for the strongest argument against your decision.

Third: what in the answer would have to convince me that I am wrong? If you cannot say, the answer probably just confirms what you had already decided.

Fourth: would the answer be the same if I had changed my stance? You check that with the test in the previous section.

6. What a different process looks like

At Expert Consilium, the same question goes to several separate AI perspectives. Each answers on its own and does not see the others' answers. The answers are then cross-checked against each other, and the report records where the perspectives agree and where they differ.

This does not remove sycophancy, but it is designed so that no single perspective is the only one shaping the conclusion, and disagreement does not quietly disappear.

Conclusion: An AI You Like Isn't the Same as an AI You Can Trust

The next time an AI answers exactly what you were hoping for, pause and ask whether you are getting analysis or just a mirror.

Warning

This article is a business analysis, not legal, tax or financial advice. The decision is always yours.

With Expert Consilium, you receive a structured report.

Level we suggest: Essential (4 AI perspectives, 9 EUR) or Consilium (8 AI perspectives, 19 EUR).

Sample report: expertconsilium.com/sample-report