Why AI writing is so generic, boring, and dangerous: Semantic ablation

This is a great description of something I've been thinking about in terms of concepts like regression to the mean, clustering and median filtering - the space of LLM output is much smaller than that of the input, precisely because the LLM works so hard to extract minimal-information patterns from its input.

This is a pattern I observe frequently where the text generator replaces very specific code comments with generic versions. Same for variable names that are written in blood.

I do wonder whether "semantic ablation" is the right term for what is going on. Ablation I think fits because it refers to the removal of layers. And yes they are semantic layers. But ablation refers to the process, not the result. So on its own shortening "ablation of semantic layers" to "semantic ablation" is problematic. It sounds like semantics are used for ablation. "Semantically streamlined" could capture the intended meaning but of course streamlining has a positive ring to it, where ablation sounds problematic as it should. I guess "semantic fusiform" is too obscure :-)