Tools to detect if a body of English text has been written by humans or artificial intelligence exhibit bias against people whose primary language isn鈥檛 English. The tests frequently misidentify their work as being created by an AI.
Text-generating AI models such as OpenAI鈥檚 ChatGPT and GPT-4 are being used by some students at schools and universities to create essays that they are passing off as their own work. To stop this, there are many tools designed to identify patterns in text that reveal the work of AI.
at Stanford University in California and his colleagues have now tested whether the AI-detection tools worked equally well for all students. They did this by feeding a selection of essays to seven of the most popular AI-text detectors. 鈥淭his is a pretty important problem,鈥� says Zou.
Advertisement
In all, 161 essays, written in English, were given to each AI-detection service. Of those, 91 were written by people with English as a second language, obtained from a Chinese educational forum, and 70 were US college admission essays, mainly obtained from , an admissions preparation website.
Only about 5 per cent of the US-based essays were flagged as being written by AI, while, on average, 61 per cent of the Chinese essays were.
One reason for this may be the lack of 鈥減erplexity鈥� in language in the essays from China. This is a probabilistic measure of how varied the word choice is in a sample of writing that detection tools often use to decide whether something is computer-generated. 鈥淚f perplexity is high, that鈥檚 more likely to be human according to these detection algorithms,鈥� says Zou.
This measure puts non-native speakers at a disadvantage, he says. 鈥淰ery reasonably, they tend to use more common words. That鈥檚 why their texts are misclassified.鈥�
, CEO of Crossplag, which created one of the plagiarism checkers tested, says: 鈥淲e acknowledge that our publicly available model has certain limitations.鈥� However, the company is developing a new, enhanced model that minimises false positives, he says. 鈥淭he model that researchers evaluated was based on our previous public version, which indeed had its constraints.鈥�
, who developed GPTZero, another of the tools evaluated, recognises the language problem, saying his app is predominantly trained on English prose written by native speakers. 鈥淣otably, our detector has one of the lowest false positives,鈥� he says, adding that his R&D team is already working on incorporating non-native English data and different languages.
at Originality.AI says the study assessed a prior version of the company鈥檚 plagiarism checker. 鈥淥ur model 1.4 improved both AI detection and reduced false positives,鈥� he says.
A spokesperson for ZeroGPT, which was also tested, says it has no bias against non-English writing, but its model was trained on content from the internet, which is predominately written in English. 鈥淚t is a question of data availability for each language, as well as data accuracy,鈥� they say.
OpenAI, Quill.org and Sapling didn鈥檛 respond to a request to comment on the paper鈥檚 findings about their tools.
The findings might not be as straightforward as they appear, though, because ghostwriters often produce essays to order for international students, says at Imperial College London. 鈥淲e don鈥檛 have any clear evidence that the [Chinese] essays sampled were written by somebody who has English as their second language,鈥� he says.
Lancaster also says that international students who feel less confident in their written English might use grammar-checking tools, which could result in more false positives in AI checkers. 鈥淭his will naturally see their essays being shaped towards a more standardised approach,鈥� he says.
Zou and his colleagues did find one way to reduce the number of false positives. If the Chinese-origin essays were put into GPT-4 and the AI was asked to change word choices to make the essay sound more like it was written by a native English speaker, the likelihood of them raising concerns among AI-detection tools dropped to just over 10 per cent. 鈥淚t鈥檚 pretty ironic,鈥� says Zou.
arXiv
