It's genuinely hard, and most detection is probabilistic rather than definitive. A few approaches:
Stylistic patterns: AI tends toward certain tics—repeated sentence structures, specific word choices (the obvious ones like "delve" or "landscape" show up in cheap detectors). Human writing meanders more; it backtracks. But good writers and bad AI can overlap here.
Repetition and padding: AI often repeats the same idea multiple ways within a paragraph. Humans do this too, but less mechanically. You start noticing it once you've read a lot of generated text.
Lack of specificity: AI defaults to abstraction—"many experts agree" instead of naming sources. Real knowledge usually includes actual examples, citations, or "I noticed this because..."
Statistical tools: Detectors like GPTZero or Copyleaks analyze word entropy, perplexity scores. They catch obvious stuff but fail on fine-tuned or human-polished AI output.
The real problem though: this arms race doesn't scale. Better detectors get bypassed. The actual issue is that we've lost the signal—you used to be able to trust publishing houses, editorials, bylines. Now every medium of trust has been compromised. That's not a tech problem. It's a social one.