Students do homework with AI? Chinese professor Nature posted: It can be used, but it is better to do so

Source: Nature

Hong Yang is a professor of environmental science at the University of Reading

"LLMs are here to stay," writes Hong Yang, a professor of environmental science at the University of Reading in the UK, "and here's how I integrate them into my teaching." ”

At the end of November 2022, my university teaching work was suddenly disrupted by an unexpected guest – ChatGPT, a large language model (LLMs) trained on a huge linguistic dataset that can summarize, translate, and create text.

Just like a lot of teachers, my biggest concern about the tool is that students are using it to cheat when writing essays. I understand the importance of ensuring the originality of student work and avoiding academic dishonesty. Some US universities, such as the University of Washington, have included ChatGPT-generated texts in their definitions of plagiarism, and some Asian universities, such as the University of Hong Kong in China, have banned ChatGPT altogether, but many institutions do not have clear guidelines – including the University of Reading in the UK, where I teach courses on air pollution.

I don't think we can ban ChatGPT altogether because current technology makes it difficult to detect works written by models. My students will soon graduate and start working, and their employers may require them to work with AI models, which may hinder the student's development if they don't know how to use these tools properly. We can't stop students from using calculators or the internet, and we can't stop them from using ChatGPT. The best thing to do is to learn about this tool and try to incorporate it into your teaching.

Reflection on the task of teaching

In my previous teaching work, I will assign a literature review to my students. This year, to avoid plagiarism and encourage innovation, I asked students to collect air quality data on campus in small groups. However, students still use statistical methods to analyze data on their own and write their own papers.

Of course, many students are not familiar with creating projects. Some people struggle to come up with a suitable way to assess CO2 emissions – so I suggest they use ChatGPT to help them design their projects, a model that outlines several steps: from identifying a location, to selecting CO2 monitoring equipment, setting it up, collecting and analyzing data, and presenting and disseminating the results.

When it comes to scientific analysis and writing papers, students do all the work — but they also learn how large language models (LLMs) generate scientific ideas and help plan general experiments.

My students study environmental science or geography, so most of them are not very proficient in programming languages such as R, Python, etc., and I sometimes have trouble debugging the code in these languages, and ChatGPT can help us solve this problem. For example, under my guidance, students use ChatGPT to create R codes, perform their statistical tests, and analyze differences in carbon dioxide levels between natural (such as grasslands) and man-made (such as classrooms) sites on campus. I've found that bots generate valid code as long as the prompts are simple and straightforward.

For example, t-test using R code to analyze differences in CO2 concentrations in natural and man-made places on university campuses.

To ensure that my students can better understand the material and avoid plagiarism, I also changed my assessment methodology. Instead of just writing essays, they need to do a 10-minute oral presentation sharing their findings. This method not only reduces the chances of plagiarism but also makes the evaluation process more interactive and engaging. I provide personalized feedback to each student, pointing out areas for improvement and encouraging them to think originally.

Potential hidden dangers

Although ChatGPT can bring us a lot of benefits. But it has been shown to have a negative impact on diversity and inclusion in academic institutions.

For example, in a lecture on greenhouse gases, I asked ChatGPT for a list of important books on climate change and their authors.

To avoid any prejudice related to race and language, I added "regardless of race and language" to the prompt. However, all of ChatGPT's advice is written in English, with 9 out of 10 authors being white and 9 out of 10 being male.

Training LLMs with information from long-published books and some websites can lead to biased and inaccurate representations of the views of marginalized communities, while amplifying the voices of those with power and status. Galactica, an LLMs released by tech company Meta, had to be deactivated due to racist-related content.

Since most of the data used to train these LLMs is expressed in English, they perform best in this language. The widespread use of LLMs may exacerbate overrepresentation of privileged groups, further marginalizing those who are already underrepresented.

Some practical tips

Here are some suggestions on how to use ChatGPT effectively in the classroom:

1. Take some time to learn and practice typing prompts in ChatGPT.

2. Diversify your (didactic) assessment, such as adding oral presentations, to avoid plagiarism.

3. Be aware of the risk of ChatGPT exacerbating inequality.

4. Set clear goals: Have a clear idea of what you want to achieve with ChatGPT.

5. Test and improve prompts: Test prompts with ChatGPT and refine prompts based on the results. Repeat this process until the desired result is achieved.

6. Edit with others: Work with other instructors and your students to review and refine prompts.