Many academics are already concerned about the ease with which students can plagiarise assessments using AI language models. To prove this, one student journalist received a 2:2 grade from the University of Bristol using an essay written entirely by AI, showing how easy it is to ‘cheat your way to a 2:2.’
As the world struggles to create a code of ethics to engage with AI, many companies are cashing in on the fervour.
New models — such as CopyLeaks, Scribbr and GPTZero — claim to be able to identify AI-generated text and are confidently touted to customers in higher education and the business world.
So does this mean the new AI detectors are ready to step up and combat the encroaching scourge of machine-generated student essays? Well, not really.
A Biased Intelligence
A 2023 Stanford University study showed that not only can AI detectors be easily fooled, but they may also be biased. Dr James Zou, Associate Professor at Stanford University and co-author of the study says:
‘We need to be very cautious about using these models, especially in education or employment. The main finding from our paper is that, first of all, these detectors are relatively easy to fool. Secondly, they have a lot of biases. They give a lot of false positives, especially if you are a non-native speaker.’
The Stanford study found that AI detectors misclassified more than half of essays (61.22 per cent) written by non-native English students as AI-generated. In contrast, essays written by native English speakers received a ‘near perfect’ evaluation from the AI detectors. Using a metric called ‘perplexity,’ most AI detectors will analyse how rare or common certain words are. If the text uses very common words, it will have low perplexity. If the text uses uncommon words it will have high perplexity. Low perplexity is taken as a sign that the text has been machine-generated.
Dr Zou explains: ‘For non-native speakers who may have a limited vocabulary, they often use simpler words. Their usage may be less diverse or complex, which means their text will have lower perplexity compared to a native English speaker.’
To understand this better, the researchers fed the non-native speakers’ essays through ChatGPT with the prompt: ‘Enhance the word choices to sound more like that of a native speaker.’ This led to an almost 50 per cent reduction in misclassification.
Generative AI will also use perfect grammar, so simply adding unusual or incorrect grammar can be enough to fool the detectors. ‘It is very much an arms race,’ argues Dr Zou, ‘The language models themselves are getting more flexible and powerful. They are able to mimic human writing styles in more and more sophisticated ways. But, I think these language models could be very useful tools for education. We should integrate that into the curriculum, and if used properly they could help the students to learn. They could almost serve as a tutor.’
An Information Revolution
It’s not just AI-generated text stoking the information revolution. Last year saw a picture of a gangsta-clad Pope go viral. More insidious was the attempt to meddle in the Slovakian national elections using fake content.
Dr Zou addressed the problem of content that is fast becoming indistinguishable from reality: ‘On the technical side, there’s a lot of work being done on putting watermarks on video, audio or text to show it was produced by an AI. On the societal side, one thing that we have to learn is to think “What are the more trusted sources?” These could be organisations. You go to those websites and you trust those sources more than if you just see something pop up on your feed on social media. Maybe your trusted sources are also people. Maybe they’re your colleagues, mentors or teachers that you trust. Still, nothing quite beats a human-to-human interaction.’
A recent statement from the Russell Group of British universities called for a set of guiding principles ‘to help universities ensure students and staff are “AI literate” so they can capitalise on the opportunities.’
Dr Xue Zhou leads the AI literacy project at Queen Mary University London — a Russell Group member. Dr Zhou explained that despite facing some resistance from staff, she believes that AI literacy will become an important employability skill for students. ‘There’s still a lot of disagreement in the area of AI. Like any other kind of disruptive technology in higher education, it’s in its early stage of adoption. We did some internal surveys and about 60 per cent of the staff don’t want to use AI for different reasons,’ explains Dr Zhou, adding: ‘Some of (the staff) think the students will be lazy when using AI and it will impact students’ critical thinking skills. Some also believe it may make their jobs redundant. But, we did training with AI for students in November, and we had a huge amount of interest from them. We would like to showcase that AI can be used in an effective and efficient way. We can use AI for writing, critical thinking, research, data analysis and the practical side of AI.’
As part of Queen Mary’s AI literacy project, Dr Zhou and her team work to collect best practice case studies and devise workshops for how AI can be used in a variety of fields. Just how ingrained AI will become in our society remains to be seen. However, what is clear is that academic institutions are already making room to accommodate this new technology despite the ethical and practical obstacles.
DISCLAIMER: The articles on our website are not endorsed by, or the opinions of Shout Out UK (SOUK), but exclusively the views of the author.
