teLLMs

Klaas van Schelven
Klaas van Schelven; May 27 - 2 min read

People like to speculate all day about how to detect AI-generated text. Style, structure, tone, etc.

This is not that list.

Instead, it’s the tiny, literal artifacts that slip through the cracks, and that a human writer could not have left behind.

UI Artifacts

  • “✍️” icon after link titles (the footnote icons from the chatGPT UI)
  • source=chatgpt.com in a pasted URL.
  • Leftover [1], [^1] with no target — citation markers with no link or footnote.

Literal Voice Slips

Sometimes the bot’s voice leaks past the “human writer” and makes it right to the reader:

  • “Would you like a version that’s more…”
  • “Let me know if you want more variations.”
  • “Of course! Here’s a version with…”
  • “I’m happy to revise it if you’d like.”

Formatting Quirks

  • Double-spaced periods: “It is done. Now it begins.”. You’ll have to check the source in HTML since whitespace is collapsed in browsers. And yes humans this this in the typewriter era, so boomers get a pass here.
  • Mismatched code block language. A code block labeled matlab on a Python script on a website where no other matlab code exists.

Finally, I refuse to mention emdashes here. Humans use those – even if they shouldn’t.

Feel free to mention your own favorite artifacts on the platform of your choice; I’ll be happy to add them to this list.