AI writing detection for non-English languages, does it even work?

Most AI detection research focuses on English, but what about the millions of students writing in other languages? I tested detection accuracy across five languages to find out.

Languages Tested

I generated essays with ChatGPT in English, Spanish, French, Italian, and German. Then I ran each through Turnitin, GPTZero, and Copyleaks.

English (Baseline)

All three detectors performed well: 85-90% accuracy. This is expected since English has the most training data.

Spanish

Turnitin: 78% accuracy. Copyleaks: 82% accuracy. GPTZero: 65% accuracy. Copyleaks was the clear winner for Spanish, likely due to their specific multilingual focus.

French

Similar to Spanish. Copyleaks led at 80%, Turnitin at 75%, GPTZero at 62%.

Italian

Turnitin: 72% accuracy. Copyleaks: 79% accuracy. GPTZero: 58% accuracy. Italian had a notably higher false positive rate across all tools, possibly due to the complex sentence structures common in Italian academic writing.

German

Turnitin: 74% accuracy. Copyleaks: 77% accuracy. GPTZero: 60% accuracy. German compound words seem to confuse some detectors.

Key Findings

Accuracy drops 10-20% for non-English languages across all tools. Copyleaks is consistently the best performer for non-English detection. GPTZero is the least reliable for non-English content. False positive rates are higher for all non-English languages.

What This Means

If you’re a student writing in a language other than English, be aware that AI detectors are less reliable for your work. This means both that your legitimately written work might get falsely flagged AND that AI-generated text in your language is less likely to be caught. Keep your writing process documented either way.

For schools: consider using Copyleaks alongside Turnitin for non-English submissions, and apply higher tolerance thresholds for non-English AI detection scores.