Lecture 03 — Textual tells

How language models give themselves away, and why every reliable tell has a half-life

8 March 2027

Early tells were almost embarrassingly easy to teach: a suspicious fondness for the word “delve,” paragraphs that arrive in threes, a habit of concluding everything with “in conclusion.” None of those survive contact with a model released six months later, because every tell that gets written up gets trained away. That instability is itself the first lesson: textual forensics is not a checklist you memorise once, it’s a practice of noticing what a population of text does that an individual author wouldn’t.

We will look at what has held up longer than surface style — token-level probability and “burstiness,” the flattening of a writer’s idiosyncrasies under editing pressure, and the specific way generated text tends to hedge. We will also look hard at where these signals fail, because a forensic method that doesn’t know its own error rate is not a method.

Outline#

• lexical and stylistic tells, and why they decay
• perplexity and burstiness, in plain terms
• the false-positive problem: a distinctive human voice can look “generated” to a naive detector

Reading#

Narayanan & Kapoor, AI Snake Oil (Princeton University Press, 2024), the chapters on generative AI’s claims versus its failure modes.

Further reading#

AI Snake Oil’s newsletter has the original 2023 post where Narayanan calls ChatGPT “a bullshit generator” — short, free, and the clearest version of the argument this week’s lecture borrows.

Practice before the session#

Try Spot the Slop — six short passages, one guess each. It won’t prepare you for Assignment 1 on its own, but it’s a fast way to notice your own instincts before this week’s session interrogates them.