My Personal Index of LLM-Speak Heuristics
I dislike finding AI writing. I wish I didn't have to navigate the economics of deciding whether or not to read it. More annoying than that however is when I find AI writing wrapped around appealing content, i.e. when someone has a real insight or conducts an interesting experiment but then has an LLM write about it.
I still haven't quite decided on my own set of categorical rules regarding the valuation of LLM writing, but at least for now I often do still read through it when I find it in the latter mentioned case.
As part of expressing my annoyance and just keeping a handle on things, I'm writing out a list of the heuristics my pattern-recognizing brain has automatically isolated. This is not meant to be an exhaustive list, but rather only the ones which I think may be novel. I likely will update this post with more as they come to me.
I'm even considering writing a browser extension to use XPath and regex to evaluate and call these out on pages I visit.
-
The sentence "the pattern is clear:
conclusion", especially ifconclusionis bolded. Variations include "the pattern is everywhere", "here's the pattern:conclusion", etc.The pattern is clear: this prose is slop.
-
Any use of a bolded term and a colon as a header inline in a paragraph or line item:
Bottom line: using a couple of words as an inline paragraph header is inhuman and a dead-giveaway.
-
Three short, "punchy" (yuck) sentences in sequence that semantically overlap.
This is a solved problem. It has been solved since 2019. Most CLIs still have not caught up. [ref]
-
Where/what/why explicative headings:
What the Problem Is
Why You've Never Fixed It
Where It Happens