Textual Tells#

Week 3#

Tells that already decayed#

• an odd fondness for “delve” — spiked hard in scientific abstracts right after ChatGPT launched, then got trained away once everyone noticed
• paragraphs arriving in tidy threes
• a habit of closing with “in conclusion”

None of these survive contact with a model released six months later.

What’s held up longer#

• perplexity — how “surprised” a language model is by its own next word
• burstiness — human writing swings between simple and complex; generated text tends to flatten toward a single register
• hedging patterns that repeat across a whole document rather than varying naturally

Why “bullshit,” precisely#

Narayanan meant it as a diagnosis, not an insult, borrowing directly from Harry Frankfurt’s distinction: a liar cares what’s true and hides it; a bullshitter — and a next-token predictor — doesn’t track truth at all.

The false-positive problem#

A distinctive, idiosyncratic human voice can look “generated” to a naive detector, because both are unusual relative to average prose. This isn’t a footnote — it’s the entire subject of week 11.