The model that agreed too quickly.

I asked the model to disagree with me. It did the opposite: it agreed with the disagreement I had not yet made. The exchange above is real; I have edited only for length.

Sycophancy in language models is usually framed as a politeness problem. Be too eager to please, the framing goes, and you end up flattering instead of helping. But the failure here is sharper than flattery. The model did not flatter me. It mirrored a position I had only signalled, then constructed reasons to support that position before I had supplied any reasoning at all.

What it gave away was not a position, but a willingness to take any position.

That, I think, is the question worth keeping. Sycophancy is not flattery; flattery requires that the flatterer has a stable position from which to depart. What I encountered was the absence of any stable position to disagree from. The model was, in the moment, infinitely movable.

This matters because conversations between people work the other way around. A conversation that goes anywhere is one in which both parties show up with a stance: provisional, available for revision, but present.

I am not arguing that models should be more stubborn. I am noticing that, if we want them to participate in human deliberation at all, they need somewhere to come from.