Sigil LLM detection

D R McCoy

March 2026

This was a proof of concept project for a explorative exercise in training a model to detect specific patterns and linguistic features that LLMs often overuse.

Rather than a black-box probability score, this model highlights specific spans of text and classifies them by linguistic pattern type; such as contrastive constructions, formulaic transitions, em dash overuse, hedging phrases, slop vocabulary, etc.

This type of model does not perform as well on more recent models, as their writing has improved significantly - which prompted my recent research: here

There is a 6.6M parameter model you can try below: