What AI content detectors actually catch (and what they miss)

People keep asking: will this pass an AI detector? It's the wrong question, but it's worth understanding why, because the answer changes how you edit.

How detectors actually work

AI-text detectors don't "know" a machine wrote something. They measure statistical properties of the text, mainly two:

• Perplexity: how predictable each next word is. Model output runs smooth and predictable; human writing is bumpier and more surprising.
• Burstiness: how much sentence length and rhythm vary. People write a long sentence, then a short one. Models drift toward an even, uniform cadence.

A detector sees low perplexity and low burstiness and raises its hand. That's the whole trick. It's a probability guess about how the text is shaped, not a fact about who made it.

Why they get it wrong in both directions

Because it's a shape-based guess, it fails predictably.

• It flags real human writing. Clear, plain, well-structured prose, the kind good writers aim for, can look "too smooth" and get marked as AI. Non-native English writers get flagged more often, unfairly.
• It clears real AI. Add a few odd word choices, vary the sentence lengths, and generated text sails through. The patterns detectors measure are the ones easiest to perturb.

So a passing score proves very little, and a failing score might just mean you wrote cleanly.

The right goal isn't a score, it's a reader

Here's what the detector debate misses: your audience is the detector that matters. A person doesn't run a perplexity check. They feel the tells, the puffery, the tidy symmetry, the "not X but Y" reflex, and they lose interest. You can pass every automated detector and still get scrolled past, and you can fail one while writing something people actually read.

Optimizing for a detector score pulls you toward gaming statistics. Optimizing for a human pulls you toward specifics, real opinions, and uneven, natural rhythm. That also lowers your detector score, as a side effect rather than the goal.

What to do instead

Forget the score. Edit for the reader:

• Lead with something specific and true.
• Vary your sentence lengths on purpose.
• Cut the adjectives that stand in for facts.
• Add at least one detail only you could know.

That's the same standard behind the list of AI tells and the editing pass, and it's what SelfSM optimizes for by default: sound like a person, because that's who's actually reading.