In the world of academic publishing (and beyond), generative AI tools present both benefits and risks. Whilst they can boost productivity and, when used responsibly and ethically, complement human endeavour, their improper use can compromise intellectual rigour and cognitive capacity (see, for example, this study and this one) and cast doubt on the veracity and reliability of reported findings.

With increasing suspicion swirling around the misuse of AI, academics and institutions are ever more wary of being flagged for AI use in the writing of their research for fear that such suspicions will be extrapolated to the production of that research. (This concern is particularly salient for authors in the humanities, for whom the writing of research is often substantively constitutive of the research itself.)

Naturally, with the rise of AI and the proliferation of concerns around its (mis)use, “AI detection” tools have begun to spring up across sectors. We have shown elsewhere that these tools are not fit for purpose. They produce high false positive rates, with grave consequences for researchers falsely flagged as having used AI, and their output is highly dependent on even negligible differences in formatting and between manuscript versions.

We have now become aware of a further serious deficiency in these tools that is even more relevant to AsiaEdit’s client base: bias against authors whose first language is not English. A paper by researchers at Stanford University revealed that several widely used AI detectors consistently misclassified non-native English writing samples as AI-generated, whereas native writing samples were accurately identified.

This is probably because most current AI detection tools are based on the degree of “perplexity” in text – the extent to which linguistic expression is unpredictable. More “human”-sounding text is expected to be more complex and unpredictable. As the linguistic expression of non-native English writers (such as the richness of their vocabulary and the complexity of their grammar) is more limited than that of their native counterparts, non-native writers are more vulnerable to being falsely flagged for AI use.

“This finding”, say the researchers, “underscores the necessity for developing and refining AI detection methods that consider the linguistic nuances of non-native English authors, safeguarding them from unjust penalties or exclusion from broader discourse”.  

Another important study conducted at the University of Essex supported this finding, showing that AI detection tools are indeed biased against non-native English speakers (along with other groups of authors who may deviate from expected language norms, such as those with disabilities).