AI text detectors ‘biased against non-native English speakers’

Tools designed to detect whether academic writing has been generated by artificial intelligence “inherently discriminate” against non-native English speakers, a study has found.

Researchers tested the performance of seven widely used detectors on 91 essays that had been written by Chinese students as part of their Test of English as a Foreign Language (Toefl) exams. More than half were incorrectly labelled as “AI-generated”, equivalent to an average false-positive rate of 61.3 per cent.

The study – published by Stanford University academics Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou in the journal Patterns – also analysed the detectors’ performance when presented with 88 eighth-grade essays written by American students and found that these were accurately classified.

Campus resource: How to use ChatGPT to help close the awarding gap

“The design of many GPT detectors inherently discriminates against non-native authors, particularly those exhibiting restricted linguistic diversity and word choice,” the authors conclude, adding that they believe the findings emphasise “the need for increased focus on the fairness and robustness of GPT detectors”.

ChatGPT’s emergence late last year sparked the launch of several AI writing detectors, all claiming high degrees of accuracy. Major players such as Turnitin have vied with start-ups and apps created by students to become the go-to detector used by universities concerned about whether students are using AI to cheat in tests.

Detectors use “text perplexity” to spot AI-generated text, the study explains, meaning that they predict what will be the next word in a sentence, mirroring the methods used by the text generators themselves. If words are easy to predict, text perplexity is low and AI is more likely to have been used; if the next word is hard to predict, text perplexity will be high.

Because non-native speakers often have a smaller vocabulary and “exhibit less linguistic variability”, they are more likely to be inadvertently penalised, the study finds.

The authors were also able to fool the detectors by prompting ChatGPT to self-edit its text by adding “more literary language” and therefore increasing the text perplexity. This caused detection rates to “plummet to near-zero”.

“The implications of GPT detectors for non-native writers are serious, and we need to think through them to avoid situations of discrimination,” the study concludes.

Potential repercussions include researchers from non-English-speaking countries being excluded from academic conferences or journals that prohibit the use of GPT, it warns.

“Non-native students bear more risks of false accusations of cheating, which can be detrimental to a student’s academic career and psychological well-being,” the paper adds. “Even if the accusation is revoked later, the student’s reputation is already damaged.”

Non-native speakers might also “ironically” be forced to turn to ChatGPT to develop their writing, the study suggests, because it can be used to “refine their vocabulary and linguistic diversity to sound more native”.

In light of the findings, the authors said it was “crucial” that “more robust and equitable methods” be developed by the companies creating AI detectors and that their use in educational settings be curtailed until then.

“Even for native English speakers, linguistic variation across different socioeconomic backgrounds could potentially subject certain groups to a disproportionately higher risk of false accusations,” they warn.

Detectors should not use a “one-size-fits-all approach” and instead be designed in collaboration with users and be benchmarked against diverse writing samples “that reflect the heterogeneity of users”, it adds.

They should also be subjected to “rigorous evaluation”, and users should be better made aware of their potential flaws.

tom.williams@timeshighereducation.com