In short
A Stanford study found an average false-positive rate of 61% across seven common AI detectors for text written by non-native speakers - for native-speaker text it was under 10%. The reason: detectors measure predictability (perplexity), and simpler, more formulaic language - typical for language learners - produces exactly that signal, regardless of AI use.
The mechanism behind it
Earlier AI detectors mainly relied on two signals: perplexity (how surprising the word choice is) and burstiness (how much sentence length varies). People writing in a foreign language tend to reach for more common words, simpler sentence structures, and more formulaic phrasing - for good reason, that's just normal language learning. Statistically, though, it looks exactly like what a language model produces.
What this means for your practice
- Don't simply drop AI detectors for non-native writers. Instead: treat the score as a weaker signal, never as the sole basis. A Deep Analysis across several texts from the same person can help identify unusual outliers.
- Never use a purely statistical tool as an AI detector. Only a specifically trained AI model can deliver reliable results here - in our own tests, purely statistical detectors failed one after another.
The consequence for grading: a high score on a linguistically simple but factually correct non-native text is more likely a sign of language-acquisition level than of AI use. Weigh that explicitly before you enter the 5-step process.
Frequently asked questions
No, but it means you should read the score with far more caution than for a native-speaker text - and get a second check with passage-level reasoning instead of trusting the number alone.
The mechanism is language-independent - it applies wherever perplexity is used as a signal. The same structural caution applies to German-as-a-foreign-language text as to the TOEFL essays in the original study.
The Deep Analysis uses a statistical tool and a neural network to analyze the text. Every result produces a kind of "fingerprint" as a unique signature. This makes it easy to recognize multiple texts from the same person.
By Simon Schmidlin