The artificial intelligence ChatGPT composes essays, communications and poetry that are confusingly similar to human texts. Researchers have now trained a model to reliably assign certain scientific texts to humans or AI based on typical characteristics. The hit rate is more than 90 percent. However, this has so far only worked for a narrowly defined data set. Experts therefore criticize that the transferability to other publications is doubtful and the detector can be fooled with little effort.
Since the company OpenAI published their AI system ChatGPT in November 2022, the question of how to distinguish between human and AI-generated texts has come into the public eye. Because the generative AI based on adaptive neural networks can generate masses of fake product reviews, write seemingly journalistic texts and write entire homework for students. There is also a risk of attempts at deception in scientific publications. For certain areas of application - for example restaurant reviews - there are already detectors that are supposed to recognize whether a text was written by an artificial intelligence. So far, however, reliable detection systems have not been available for most areas.
Human or AI?
A team led by chemist Heather Desaire from the University of Kansas has now trained a model that is intended to recognize whether certain scientific texts were written by a human or by ChatGPT. As a type of text, they focused on so-called perspectives from the journal Science. These are short articles in which researchers provide a classification of a specific research topic or result. ChatGPT texts on the same topic served as a point of comparison. The request to the chatbot was always to write a 300 to 400 word summary on the topic in question.
Using a training data set of 64 real perspectives and 128 texts created by ChatGPT, the researchers then trained their model on typical features of human and computer-generated perspectives. "By manually comparing numerous examples from the training set, we identified four categories of features that proved useful in distinguishing human texts from those of the chatbot," report Desaire and her team. According to this, people tend to write longer and more complex paragraphs, vary sentence lengths more, use certain punctuation marks such as brackets, colons and dashes more often and use certain words more often than ChatGPT.
Doubtful transferability
The researchers then tested their model on two data sets, which also consisted of science perspectives and ChatGPT texts created in the same way. Since the test data sets were similar to the training data set, the model achieved a hit rate of 100 percent. If the decision was only made on the basis of excerpts from the respective texts, the accuracy was still 92 percent.
From the point of view of the language technology professor Chris Biemann from the University of Hamburg, who was not involved in the study, this almost perfect classification is an indication of so-called overfitting: This means that the model was tailored so much to a specific data set that it works very well on this one - "but only on this one data set," says Biemann. The writing team also concedes that the approach "was designed for a narrower range of writing. "It remains to be determined to what extent the model is generally transferrable."
Attempts at cheating and editing are not taken into account
Another weakness of the study: the request to ChatGPT was only to write a general summary. The researchers did not tell the software that it was supposed to be a scientific text - which explains, among other things, why many of the distinguishing features used in the model referred to typical scientific formulations and complex sentences. "If ChatGPT were asked to write an introduction for a research article in the style of a specific scientific journal, the content would likely be more difficult to discern, both for this model and for others," the researchers said.
But while the model itself only works for the type of text it was trained on, Desaire and her team hope their method could be useful in other contexts as well. "We strived to create an accessible method so that even high school students could create an AI detector for different types of text with little guidance," says Desaire. "There is a need to delve into AI writing and you don't need a computer science degree to contribute in this field."
Biemann, on the other hand, considers the approach to be outdated. "The approach itself can in principle be extended to all types of detection, but countermeasures by the deceiver are not considered and are limited to simple hints to ChatGPT," he criticizes. Even an adjustment of the prompt or a minimal human post-processing of the generated text could therefore render the model unusable. Although there are already more advanced approaches, it is not yet possible to reliably distinguish whether a person has written a scientific text themselves or simply slightly adapted a computer-generated text.
Source: Heather Desaire (University of Kansas, USA) et al., Cell Reports Physical Science, doi: 10.1016/j.xcrp.2023.101426