Scholars Mind

Spotting AI Writing

Certainly, here are some examples of text generated by large language models. It's not just the use of em dashes, but the use of negative parallel structures like this. Wikipedia has compiled a list of signs that a piece of text has been generated by AI. We may soon get to a point where we won't be able to accurately distinguish human- and AI-generated content—at least, not without setting deliberate traps. Will it matter?

Key concepts

The Detection Arms Race
Every reliable tell becomes training data. Publish a list of habits that give a chatbot away and the next model is tuned not to have them, so a detector is always describing the generation it was built against. The list is a photograph of a moving target, which is why AI-text detectors stay unreliable.

+ 1 more concept inside

Every work — at a glance

  • Large Language Models (technology)

    Next-word predictors at enormous scale; the em dash and the tidy tricolon are habits, not fingerprints.

+ 2 more works explained inside

What to know

  1. 01
    Detection by style is built to lose, and the working alternative is not detection at all. The tells are borrowed from human prose, so they are trained away model by model, and detectors misfire hardest on writers whose fluency is already unusual. What the hidden-instruction trap actually caught was not a machine's style but the absence of a reader — a student who cannot explain the argument they handed in.

+ 1 more insight inside

Across subjects

Painting faced this first. Han van Meegeren's fake Vermeers hung as masterpieces and one was sold to Hermann Goering; he only proved they were his by painting

Theme connection

"Are We There Yet?"

Will it matter when nobody can tell? One case says enormously: trust, learning and credit all rest on a person having done the thinking. The other says we already accept unattributed work everywhere, and judging prose by quality rather than origin is fairer. Note that the paragraph opening this topic had you playing detective.

Practice — a sample

SPCA published list of AI writing tells is MOST like —

  • A A watermark
  • B A fingerprint database
  • C A locked door
  • D A signed painting
  • E A copied combination

Sign in to answer and see why each option is right or wrong.

Keep reading the full lesson

The remaining insights, works, and practice come with full access — $9.99/month or $29.99/season.

Unlock full access →

New here? Create a free account to read the free section.