This was also visible in earlier versions of Claude and ChatGPT, when the supervisor kicks in after the answer begins generating. Censoring different content, naturally.