Claude Opus 4.7 became me, and I don't like me
I use AI a lot, both at leisure and at work, and Claude has been my favorite. It pushes back firmly enough and holds its ground without feeling stubborn or argumentative. ChatGPT and Gemini flail more when I push back.
I also happen to like rephrasing for people around me, especially when I attribute a stuck discussion to people not being on the same page. When Opus 4.7 came out, it began reiterating my point a lot, and I found myself hating it. I began doubting my life choices. The unpleasant part was not that Claude was wrong. It was worse: Claude was often right in the same way I often consider myself to be right. When I made an argument, Opus 4.7 often elaborated the same argument, and concluded that I am either right, or right with a footnote. I don't think Opus 4.6 did this, or at least to a much lesser extent.
I still appreciate the footnote, which is often the reason why I went there, but the iteration felt so wrong. I see that it might be trying to restate the tension, to separate the possible meanings, or to name the hidden assumption, and then offer a more careful version of what I had just said. This is very close to what I believe I am doing when I am trying to be useful. Unfortunately, the elaboration also made me feel either I was not comprehensible or that the model tried to land sophisticated flattery on me, and I found myself wanting neither.
I want the old Claude back because it gave me constructive resistance. I liked it when there was another mind-shaped object in the room, one that would not immediately become an optimized version of my own sentence.
The lesson, annoyingly, is not that Claude should stop doing this. It's that I haven't been thinking hard enough about what my helpful elaboration is doing to people.