news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Smart Things
Compiled by Chen Junda
Edit Panken

According to a report by the top academic journal Nature on July 31, the use of generative AI in academic writing has seen explosive growth. Related research shows that 10% of paper abstracts on PubMed, the largest database in the biomedical field, are suspected of being written by AI, which is equivalent to AI being involved in 150,000 papers each year.

Research by the Berlin University of Applied Sciences in Germany shows that the average accuracy of mainstream AIGC detection tools is only 50%, and it is easy to judge human writing as AI-generated. However, many AI-generated papers can easily be concealed through paraphrasing, synonym replacement, etc. In addition, it is more difficult to detect the use of AI tools by native English speakers.

Although AI tools have been widely used in academia, using generative AI to directly output papers or ghostwrite them is still controversial. AI tools make plagiarism easier and may also lead to the abuse of copyrighted content.

AI-assisted writing is not without merit. Many scholars have avoided the trouble of publishing papers in unfamiliar languages ​​by using generative AI-assisted writing, and can focus on scientific research itself. Many journals allow the use of generative AI tools, but require authors to disclose the use in detail in the paper.

1. AI may be involved in writing 150,000 abstracts, and non-native English speakers are more likely to be discovered

Since ChatGPT was released at the end of 2022, the use of AI in academic writing has exploded. Research from the University of Tübingen in Germany shows that at least 10% of abstracts in biomedical papers in the first half of 2024 were written using AI, which is about 150,000 papers per year.

The research team analyzed the abstracts of 14 million papers in the biomedical database PubMed from 2010 to 2024, and found that after the emergence of generative AI tools such as ChatGPT, the use of specific modifiers has increased abnormally. The research team used the frequency of these words to estimate the proportion of abstracts written using AI.

The researchers also found that the use of AI writing tools varied across countries, with their data showing that papers from countries such as China and South Korea used AI writing tools more frequently than papers from English-speaking countries.

▲The use of Chinese style words in academic writing has increased dramatically (Source: Nature)

According to the research team's analysis, authors in English-speaking countries may use AI at a similar frequency as authors in other countries, but the way they use it may be more difficult to detect.

In fact, before this wave of generative AI emerged, AI had already been put into use in fields such as drug discovery and protein structure prediction, and these uses did not cause much controversy. The reason behind this may be that in such applications, AI obviously plays an auxiliary role.

The application of generative AI in academic writing has brought about two major problems. On the one hand, plagiarism has become easier because of generative AI. Plagiarists can let generative AI paraphrase other people's research in the style of academic journals, and the degree of similarity between the paraphrase and the original text is generally low, so it is difficult to be judged as plagiarism.

In addition, AI models consume a lot of copyrighted content during training, and the model is very likely to output copyrighted content to users without any prompts. The New York Times once found that OpenAI's AI chatbot ChatGPT had directly output the original words in the New York Times report without marking the source. They also sued OpenAI for infringement.

2. Detection tools lose in the cat-and-mouse game, can be easily circumvented and are prone to misjudgment

In response to the increasingly widespread use of AI tools, many companies have launched AIGC content detection tools, but these tools have all failed in the "cat and mouse game" with generative AI.

At the end of last year, several scholars from the Berlin University of Applied Sciences in Germany published a study in which they found that only five of the 14 AI detection tools commonly used in academia achieved an accuracy rate of more than 70%, while the average recognition accuracy rate was only 50%-60%.

▲The performance of mainstream AIGC detection tools on the market is poor (Source: Weber-Wulf et al.)

These AIGC detection tools perform even worse when faced with AI-generated content that has been manually edited or transcribed by machines. With simple operations such as synonym replacement and word order adjustment, the accuracy of AIGC detection tools will drop below 50%. Studies have shown that the comprehensive recognition accuracy of AIGC detection tools is only 50%.

▲ In the figure, 05 and 06 are texts generated by AI and edited by humans and transcribed by machines, respectively. The AIGC detection tool performs poorly on these two types of texts (Source: Weber‑Wulf et al.)

Data shows that AIGC recognition tools on the market have a high accuracy rate in identifying human-written papers. However, if the author first writes an original paper in a language he is familiar with and then asks the translation software to translate it into another language, it may be misidentified as AIGC by the recognition tool. This may have a very negative impact on the academic reputation of scholars and students.

3. The boundary between assisted writing and academic misconduct is blurred, and AI writing tools also have their value

However, the use of generative AI tools has indeed brought convenience to some researchers. Hend Al-Khalifa, an information technology researcher at King Saud University in Riyadh, shared that before the emergence of generative AI tools, many colleagues who were not proficient in English faced great obstacles in writing papers, but now these scholars can focus on the research itself without spending too much time on writing.

The line between AI-assisted writing and academic misconduct is hard to draw. Soheil Feizi, a computer scientist at the University of Maryland, argues that using generative AI to paraphrase existing papers is clearly plagiarism.

However, the use of AI tools to assist in expressing opinions should not be punished. Researchers can use detailed prompts to generate texts or use AI tools to edit drafts, provided that they actively disclose the use of AI tools.

Many journals have also regulated the use of AI tools in academic writing, but have not directly banned it. The top academic journal Science stipulates that AI cannot be listed as a co-author, and authors should disclose the AI ​​system and prompt words used, and be responsible for the accuracy of the content and whether there is suspected plagiarism.

▲The regulations of the academic journal Science on the use of AI (Source: The official website of the academic journal Science)

The academic journal Nature stipulates that researchers should record the use of generative AI tools in the "Research Methods" section. Statistics show that as of October 2023, 87 of the top 100 journals have developed guidelines for the use of generative AI tools.

Conclusion: AI tools are not a disaster; changing academic orientation is the only way to solve the problem

In the past university graduation season, many domestic universities also introduced AIGC detection tools in the thesis defense and review process. However, the introduction of this indicator has not effectively curbed academic misconduct related to AI tools. There are also various services on the market that specifically reduce the AI ​​rate. After being mistakenly judged as AI-generated, many students' papers have become unrecognizable due to the reduction of the AI ​​rate.

Taking an adversarial attitude towards generative AI tools in academic research may not be a fundamental solution to this problem. Scholars from the University of Applied Sciences Berlin in Germany also emphasized in their research summary that the problem of AI abuse in academic writing is difficult to solve by AIGC detection alone, and adjusting the academic atmosphere oriented towards papers and results is the key to solving this problem.

Source: Nature