.png)
A list under development
Based on our dialogue with various academic circles and relevant literature [1], KURT has put together a list of tips and measures for how you can meet the challenges of language models in your teaching. The list is far from exhaustive, and we will update it as we receive tips and thoughts from you. Feel free to get in touch if you have ideas or want to discuss any of the points below.
Explain the guidelines
- Although language models such as ChatGPT are not explicitly mentioned in the guidelines for the exam, it is very clear that the answers must be the students' own work. Read more here.
- Copying text from other sources, getting others to write the text for you, or using text generators is considered cheating, and can result in expulsion and a lost opportunity to sit the exam.
You can allow the use of language models and adapt the assessment criteria accordingly
- Emphasize originality and creativity in the answer.
- Pay extra attention to overall and logical structure, consistent language and terms, and a clear explanation of key points in the text.
- For subjects where mathematics is a central part: Teach students to write good mathematical texts.
- Make sure that the students describe and explain what they do in the exam, and that they get practice in doing this in the teaching before the exam.
Not all tasks are suitable for language models
- Interpretation of data sets often requires both the ability to organize data, present data visually, extract meaningful information from the presentations and check the results.
- Let images, video, sound and other media be included in the task formulation, so that the students themselves have to translate this information into text.
- Language models cannot, for example, read graphs or extract meaningful information from a video.
- Ask the students to produce supplementary figures and figure texts. Feel free to focus on current topics.
- ChatGPT is trained on text material up to and including 2021. Focus on niche areas, where you have checked in advance that language models generate academically weak answers.
- Create fictitious scenarios in which the students must apply their subject knowledge to assess various outcomes given certain criteria.
Customize assessment forms
- Correction in group lessons or outside of the lessons, where the group teacher approves the answers after an oral review, does not need to take more time than correction at home. Handwritten answers prepared in the group lessons can be collected for assessment. (These can be scanned and entered into Canvas)
- Focus on the process: Ask for a reflection note where the students describe how they have thought to arrive at the answer.
- For topics that use programming: Use version control tools like GitHub, which monitor the process of developing code responses.
- Ask the students to draw up a project outline before starting the work, in which they explain the problem and objectives.
Practise plagiarism control
- Look for irregular wording or repeating patterns in the text. Language models such as ChatGPT can be repetitive, and have a strange and inconsistent use of language.
- Check the use of sources in the text. Language models tend to include inappropriate and non-existent sources. Look for consistent typos.
- People can be wrongly taught and make consistent mistakes in, for example, and/or, punctuation, then/when, sentence structure and much more. Look for originality.
- Language models will not move far outside the data on which they are trained. Use analysis tools such as GPTZero, ZeroGPT or OpenAI's text classifier. These are likely to change in the coming years.
- If cheating is suspected, follow the university's guidelines.
Read more: