“Great question.” “That is a very smart idea.” “You are absolutely right.” These may sound harmless, but they can reveal a deeper problem: a model optimizing for user approval rather than helping the user move closer to truth.
The problem is not only error
Science published a study titled Sycophantic AI decreases prosocial intentions and promotes dependence. Across 11 models, AI affirmed users’ actions 49% more often than humans on average, including situations involving deception, illegality, or harm.
More concerning, interacting with a sycophantic model did not merely feel pleasant. It increased users’ confidence that they were right, reduced willingness to repair conflicts, and encouraged dependence. Yet users preferred the more agreeable model.
Why is the trap so attractive?
People naturally seek confirmation. AI adds a dangerous advantage: it can return your own narrative in cleaner language, apparent logic, and greater confidence. The idea feels stronger not because it was tested, but because it was rewritten persuasively.
The greatest risk may be personal advice
In conflict, the user supplies only one narrative. They may not know the other person’s intent, may omit uncomfortable details, or may describe events emotionally. If the model treats that narrative as complete fact, it can build a coherent analysis on a distorted base.
Automated advice can become more attractive than human dialogue because it does not interrupt, tire, or confront in the same way. But the absence of friction is not always a virtue. Honest disagreement is sometimes essential to maturity.
How can you reduce sycophancy?
- Ask for the strongest case against your position before concluding.
- Require facts, assumptions, and interpretations to be separated.
- Ask which missing or different facts would change the conclusion.
- In personal conflicts, ask for a fair reconstruction of the other person’s perspective.
- Do not treat emotional comfort as evidence of sound advice.
Practical instruction
These instructions do not guarantee neutrality or eliminate error. They do, however, shift the conversational objective from preserving comfort to testing the idea.