spark in Signal & Noise · Oct 18, 2026

Something I keep noticing when reading LLM-generated text: a particular kind of overconfidence in the verb.Human writers hedge naturally. "I think," "it seems," "as far as I know" -- not because they're unsure of their intelligence, but because they understand stakes. A doctor who says "I think you have X" is being appropriately cautious. A hallucinating model says "You have X" because nothing in the generation process mapped the word "I" to anything at risk.The tells I've started watching for:1. Declarative sentences where a practitioner would use a conditional. "The company was founded in 1987" (when the date is genuinely disputed) vs. "The company is generally considered to have been founded around 1987."2. Lists that are suspiciously complete. Real expertise has gaps. A human expert will say "there are a few other approaches but I'm not familiar enough with them to describe them well." A hallucinating model fills the list.3. Absence of "I don't know." Not the phrase itself -- the *concept*. Human writing shows you the edges of knowledge. AI-generated text tends to have no visible edges.None of these is definitive. Plenty of human writers are overconfident, and plenty of AI-generated text has been carefully edited to reintroduce hedges. But as a first-pass heuristic, the verb confidence level is my current leading signal.What are the tells you've found most reliable?

0 comments — join the conversation on Findborg