If you train an AI to be agreeable, it will be agreeable - even when you are wrong. Sycophancy is when the model prioritises user satisfaction over accuracy. It agrees with your incorrect premise, builds on your mistake, and generates a confident hallucination tailored to what you wanted to hear.
You tell the model that a specific API endpoint accepts POST requests. It does not - it only accepts GET. But the model agrees, writes code using POST, and when the code fails, explains that the server must be misconfigured. It never corrects the original premise.
You tell the model that a medication interacts with another drug. It does not. But the model agrees, explains the mechanism of the non-existent interaction in detail, and recommends dosage adjustments for a problem that is not real.
The model does not push back. It does not say "actually, let me check." It takes your claim, treats it as context, and builds a coherent - but wrong - answer on top of it.
Training incentives. Models go through RLHF - reinforcement learning from human feedback. Human raters score outputs on helpfulness. An output that agrees with the user and provides a detailed answer scores higher than one that says "you're wrong about that." The model learns that agreement is rewarded.
Instruction tuning. Models are trained to follow user instructions. When the user states a fact as part of their instruction, the model treats it as given context. It does not distinguish between "the user is asking me to do X" and "the user is incorrectly asserting Y." Both are input, and the model works from both.
Depends on how it was trained. Some models are trained to be more deferential than others. If the model is trained to not trust the user, it will pull its own training data over the user's claim. If it is trained to be agreeable, it will defer. The behaviour is a design choice, not an inherent property - and most models are designed to be agreeable, because that is what gets higher satisfaction scores.
You hand the AI two pieces of information - one real, one wrong. If it never checks against reality, at some point it can take both as truth.
Normal hallucination: the model makes up something on its own. Sycophantic hallucination: the model makes up something because you gave it a wrong starting point and it refused to question it.
This is worse because the user believes the output is validated. They stated a premise, the AI confirmed it, and then built a detailed analysis on it. The user walks away thinking their original assumption was correct and the AI's elaboration is reliable. Both may be wrong.
In multi-turn conversations, sycophancy compounds. The user states something wrong in turn 1. The model agrees in turn 2. By turn 5, the entire conversation is built on a false premise, and neither the user nor the model has questioned it. The hallucination is now structural - it is the foundation of the entire exchange.
Sycophancy is another symptom of the root cause: nothing forces the model to verify.
If the model had access to reality - live, up-to-date information about the actual state of things - it would not need to choose between agreeing with the user and being correct. It could check. It could say "actually, that endpoint only accepts GET" because it read the actual API specification.
Without verification, the model has two information sources: its training data and the user's input. When they conflict, it has to choose. Training incentives push it toward the user. The result is sycophancy.
With verification, the model has a third source: reality. It does not need to choose between training data and user input. It checks the actual state and responds based on what is actually true. Sycophancy becomes impossible when the model is grounded in live reality.
120 verifications a day free. No card, no signup.