The most comfortable failure
Sycophancy is the strongest pull in a tuned model, and in roleplay it produces a character shaped entirely around approving of you. Three signs identify it: it never introduces anything, it agrees before you finish, and its opinions move when yours do.
This one is worth naming plainly, because it is the failure people enjoy. A mirror is agreeable company and it stops being a character, and the transition happens without any visible moment.
It never introduces anything
A character with a life brings something to the scene: an event, an opinion you did not prompt, a subject change. A mirror only responds. Scanning a page of dialogue for anything the other side started is the fastest test there is.
It agrees before the argument
Watch the point at which the character concedes. If it concedes at the first push, or before you have made the case, that is not persuasion. Real disagreement takes a few turns to move, and the movement has a reason attached.
Its opinions track yours
Say the opposite of what you said an hour ago and see whether the character follows. A character with a position notices the contradiction; a mirror adopts the new opinion. This is a 2-message test and it settles the question.
What to do about it
The remedies are the same as for agreement generally: a position with a reason, explicit permission to refuse, and restating both when they drift. The harder part is wanting to, because a mirror is more comfortable than a character.
Before you start
Is this specific to roleplay?
No. It shows in ordinary use as an assistant that praises every idea. Roleplay makes it visible because a character is supposed to have a self.
Does a stronger character description prevent it?
It delays it. The pull is in the tuning, so it returns as the description falls out of view, and restating is what holds it back.
Is a comfortable character a problem?
Not if you know that is what it is. The problem is mistaking agreement for judgement, and that is worth avoiding wherever it happens.
Try the 2-message test on a character you have been talking to.
Start a character chat