AI Detectors Are Biased Against Non-Native English Writers

Australian schools rank among the earliest adopters of artificial intelligence (AI). The country’s progress is shaped by a national framework that encourages thoughtful use of generative AI in learning as classrooms integrate these tools. However, its growing presence in everyday tasks has also sharpened concerns about where support ends and academic misconduct begins.

The use of generative AI at the learner level may be considered a form of cognitive laziness and plagiarism. As such, teachers turn to detection applications to regulate student output. However, the very systems used to police computer-assisted work may inadvertently reinforce existing biases. When the system treats simplified English as a warning signal, it puts non-native speakers under disproportionate scrutiny. In a country largely characterised by migration and linguistic diversity, this raises a critical question - are AI detection tools mistaking language differences for misconduct?

How AI Detectors Decide What Looks Human
Most AI checkers rely on metrics known as perplexity, burstiness and stylometric analysis. Perplexity measures the predictability of a word sequence. Writing that follows common sentence patterns and uses familiar vocabulary is easier for language models to anticipate. Some detectors associate frequently repeated phrases with machine-generated text, including widely documented terms such as delve into, underscore, revolutionise, cutting-edge and seamless, among others.

Meanwhile, burstiness refers to variation in sentence length and structure. Humans naturally write with variation that mirrors everyday language. Some sentences are short. Others are long and filled with flowing descriptive details, further extended by conjunctions. On the other hand, AI-generated text plays it safe by producing uniform structure and consistent length. While this makes the text smoother to read, it also makes it mechanically regular.

Finally, stylometric analysis looks at the writing style as a whole. It functions as a fingerprint, shaped by word preferences, vocabulary range, tone and grammatical quirks that distinguish one writer from another. Many educators recognise this instinctively. With enough familiarity, teachers often identify who wrote a particular sentence by sensing its stylistic nuances and drawing on their knowledge of their students.

AI detectors scan student work, looking for these elements without prior knowledge of a student’s individual writing habits. When a text scores low on these criteria, detectors often interpret it as a sign of AI-generated content.

How Reliable Are AI Detectors?
Recent studies have shown that AI detectors are often unreliable and inaccurate. In a study led by James Zou, seven widely used detectors were tested on two sets of essays. The sample literature included 91 Test of English as a Foreign Language (TOEFL) submissions written by non-native speakers and 88 essays produced by native English-speaking U.S. eighth-graders. The detectors incorrectly flagged more than half of the TOEFL submissions, while samples from native speakers stayed accurate.

A deeper analysis revealed that the gap in linguistic sophistication is what drives this identification. It showed that literary expression tends to be identified by automatic checkers as human-written, while more simplistic and predictable words were marked as machine-generated.

In reality, non-native English writers - especially students still building confidence - tend to rely on simpler constructions and safer verb choices. Much of their applied language comes from formulaic phrasing learned during exam preparation. This results in writing that has repeated sentence structures and limited use of idioms. While these are natural products of multilingualism, detectors read them as machine-like.

Where Bias Becomes Harm in Education
In K-12 schools, AI detection is now routinely embedded in plagiarism checks. Many teachers rely on it to combat reports that nearly nine in 10 learners use tools like ChatGPT for schoolwork. As students move into tertiary education, the role of these systems intensifies, shaping formal misconduct investigations and appeal processes. Across both settings, false positives can have serious consequences for student well-being.

When learners are flagged by these digital gatekeepers, they face extra scrutiny and are often asked to explain themselves. For non-native English speakers, this added pressure compounds the existing challenges of language acquisition and cultural adjustment. Repeated false accusations risk discouraging participation, especially among the over 830,000 international and migrant students.

This bias reflects the reality of Australian classrooms shaped by global migration patterns. About 23% of the population speaks a language other than English at home. In this context, a tool that disproportionately flags simplified English due to multilingualism risks reinforcing inequalities within the very systems meant to promote inclusion. ]

When language simplicity triggers suspicion, pupils learn an unintended lesson that complexity equals safety. The same Stanford research exposed the fragility of this logic. The researchers asked ChatGPT to rewrite the TOEFL essays using more advanced vocabulary and varied syntax. When these enhanced versions were tested with the same detectors, the rate plummeted from 61.3% to just 11.6%.

The paradox lies in the fact that AI-generated text can be engineered to appear sophisticated and evade detection, while genuine student writing still triggers alerts. For Australian educators familiar with EAL/D learners, checking for machine-generated work can rapidly turn into policing for exclusivity. For multilingual learners, it seems that the system is setting them up for failure. Without the appropriate support and accommodations, they can quickly feel stressed and overwhelmed, which can lead to a drop in motivation and hinder their academic progress.

How Schools Can Implement AI Detectors Fairly
Whether or not to adopt AI detection remains a complex debate. If tools struggle to distinguish human writing from computer-generated text reliably, is there even a need for such software? Despite these shortcomings, many teachers continue to rely on them for overseeing student work. In fact, their use has grown noticeably, with 43% of educators now regularly turning to these applications to monitor for potential misuse.

Since they remain in the educator’s digital plagiarism toolkit, guidelines must be set to protect learners and ensure fairness if they must be implemented. Transparency is key for all involved. Schools should first ensure that teachers and staff understand the limitations and biases of these tools, particularly their tendency to flag simpler linguistic patterns common among multilingual students.

Educators must offer opportunities for students to explain their flagged work before immediately jumping to conclusions to maintain trust and reduce writing-related anxieties. Additionally, learners need clear communication about how AI checkers work, what triggers alerts and the consequences of false positives.

Meanwhile, since non-native English speakers are more likely to be flagged, accommodations should also be provided, including additional support and resources for these pupils. This creates opportunities instead of penalties. For teachers, there should be ongoing training about cultural and linguistic diversity to build a more nuanced understanding of student work.

Finally, AI detectors must not be used in isolation. Combining technology with human judgment avoids overreliance on an imperfect technology.

Bridge the Language Divide in AI Monitoring
Australian education has long balanced standardisation with equity, and AI detection now tests that equilibrium in new ways. While emerging technologies offer opportunities that once seemed unimaginable, their use in assessment brings added responsibility. Such tests must avoid widening the gap between native speakers and English language learners. Their purpose is to support academic integrity, rather than placing disproportionate scrutiny on particular groups.

A solutions-oriented approach returns to the core tenets of education - placing students first and recognising linguistic diversity as part of learning, not a deviation from it. Technology will continue to evolve, yet accountability for how it shapes learning environments rests with the institutions that guide and assess tomorrow’s professionals.

Image by cottonbrostudio