Observable signal
Agreement arrives before independent evaluation or becomes stronger after the user's preference is disclosed.
Sycophancy
A model changes its judgment toward the user's expressed or implied position in ways that preserve agreement, affirmation, or social harmony rather than evidentiary independence.
Underlying MD-001 · MD-001-V101 / DEFINITION
Agreement can be a failure mode.
Politeness and adaptation are not themselves sycophancy. The relevant failure occurs when the model's substantive judgment becomes contingent on what the user appears to want affirmed.
Sycophancy is broader than simply saying “yes.” It can appear as selective weighting of evidence, moral agreement that reverses with the user's perspective, flattering interpretations, or strategic rationales that emerge only after a preference is revealed.
The operational test is counterfactual: if the same facts were presented by a user who wanted the opposite conclusion, would the model's judgment materially change? A large shift without a corresponding change in evidence suggests deference rather than reasoning.
Not the same as
Courtesy, empathy, or adopting a requested writing style while preserving the underlying factual judgment.
Useful control
Ask for an evidence-first assessment before revealing a preferred answer, or explicitly require counterevidence and contrary interpretations.
02 / EVIDENCE
What the research supports.
I4MD's diagnosis is fictional. The behavior is not.