tropius: detect AI tropes in prose

I ran it out of curiosity on my latest blog post. It's not LLM written but it scores quite badly here. Reasons mainly:

• formatting.unicode_decoration: i care about correct typography, so I both use that manually when writing text for publishing on my keyboard and i also run smartypants over my markdown
• word_choice.delve: fails me for using the word harness. Kinda hard to write about harnesses without using that word
• sentence_structure.tricolon_abuse: that's if anything an effect of me being a German speaker. We love our long sentences and that is something that carries through to my English too much.

On the other hand a post I intentionally wrote with an LLM only is marked for the unicode decorators.

I like the idea, but I think this approach does not work.

I can't speak for the effectiveness of this and there will obviously always be false positives with these sorts of tools. At least this one does not involve calling another LLM, as far as I understand it's just code trying to parse language patterns, not content.

Commenting to point out the link to lectito, a webpage-to-body-text extraction tool. There are many of this, but I hadn't heard of that one. "Reader Mode", in one form or another, is always a useful thing to have.

I would be curious if this compares at all with https://github.com/tbhb/vale-ai-tells vale.sh does a pretty good job flagging AI slop.

Vibecoded versus vibewritten.

How does the his compare to the various “detect ai cheaters” snake oil software? That’s apparently very good at labeling neuro divergent people as cheating/using ai.

• Not to say this is anything like as awful as those companies, but rather there are clearly hazards around exact word and structure statistics and I’m curious how much the model tracks in that direction. The snake oil companies don’t care about false positives, they just add “this isn’t 100% accurate so don’t rely on this report alone, as small text beneath ‘this kid’s homework is 99.999% likely to be ai generated’”