Why AI detectors fail, and what actually catches AI-written work
Most AI detectors read prose and guess whether it sounds machine-written, a signal that gets weaker every time AI models improve. Learnaway takes a different approach: it reads how a piece of work was assembled, not how it sounds, so the evidence doesn't erode as language models get better. Here's why the text-only approach struggles, and what a behavioural signal catches instead.
The arms race problem
Text-based detectors are trained on today's AI output. As soon as a newer model writes more naturally, yesterday's classifier starts missing tomorrow's essays. That treadmill never ends.
False positives on real students
Simple, formulaic, or non-native writing patterns can statistically resemble AI text. Detectors that classify by style routinely flag EAL students and plain writers who wrote every word themselves.
A probability isn't evidence
A percentage score offers little to put in front of a student or a parent. It's a statistical guess, not a fact, and it's hard to defend when challenged.
What holds up instead
Paste events, typing cadence, and time on task don't depend on how convincing the prose sounds. A large paste with no preceding keystrokes is the same tell whether the output reads fluently or stiffly, and the signal doesn't change as AI improves.
Frequently asked
- Are text-based AI detectors always wrong?
- No, but their accuracy degrades as models improve, and they're prone to false positives on non-native or very plain writing. Learnaway isn't a replacement so much as a second signal that doesn't share the same weaknesses.
- Does Learnaway avoid these problems entirely?
- Learnaway never reads the words a student writes, so it can't misjudge prose style; there's nothing to be wrong about. It has its own limits instead: work needs to be written through a Learnaway link, and it always requires multiple signals before raising concern, precisely to avoid the same false-confidence problem.
- Should we drop our existing AI detector?
- Not necessarily. Many schools keep a text-based detector for a broad first pass and add Learnaway's behavioural evidence for the cases that need a fair, defensible answer.