They use automated graders now. It's more likely the communication style was rewarded by overly strict grading rubrics.

(ie "honest caveats" wording comes from a reward for being honest about drawbacks instead of covering them up)

Note Opus 5.5 is much less likely to do it.

I'm speculating but there must be someone who writes a style guide for how Claude is supposed to speak then they design a way to reward for that style? So the problem could be they employ rationalist adjacent people who think this is peak A.I. design.