In an opinion piece in the trade magazine Folkeskolen, CFU consultant Mikkel Aslak Koudal Andersen writes that it is time for schools and teachers to start relating to generative language models, to create their own experiences and to talk to each other and with students about the groundbreaking technology.

In our recently completed master's thesis, we suggest how to get started with this talk in the teachers' room based on an emancipatory critical-constructive approach based on didactic theory from Kant, Freire, and Klafki. The thesis and the IT didactic product itdidak.dk aims to equip teachers to talk to students and colleagues about this new tool, which, with ChatGPT at the forefront, hit the teaching world with a bang in November 2022.

itdidak.dk: A resource for teaching with generative language models

Our product thesis began in early January 2023 and was based on the hypothesis that many teachers do not feel adequately equipped to teach generative language models in primary schools. Initially, we contacted a group of teachers who confirmed our hypothesis and asked for guides and guidelines on using generative language models in their teaching. In the dialogue with the teacher group, we also became aware that they were concerned that their students did not have the skills to critically reflect on the use of generative tools such as ChatGPT. 

The teachers' wishes and ideas were considered in our design process with input from researchers in teaching and computer science and theoretical studies. The result was the website itdidak.dk was intended as a unified framework for several text and video guides, tips, points of attention, and pedagogical formats that will inspire teachers' teaching. Our assessment is that generative language models are part of the overall educational task that the school faces about making students both generally and digitally educated, as the generative language models will take up a lot of space in both school, at work, and in our private lives. There is, therefore, a need for adults who understand the technological, ethical, and moral implications of this technological upheaval that we as a society face and who can have that conversation with students. The teachers saw our intervention as a good suggestion for an introduction to generative language models and a good starting point for seeking inspiration for their practice about both being able to introduce the possibilities of generative language models, especially their limitations. At the same time, there were also concrete suggestions for inspiration to work with different dilemmas and ethical/moral issues through the use of pedagogical formats.

Our work on the thesis also gave us some points of attention, which we took with the teachers in the form of "the most important things that you as a teacher need to know before you start teaching in and with generative language models". The seven points of attention are:

1. There is no such thing as "artificial intelligence"

There is currently no such thing as artificial general intelligence (AGI), but aspects of human intelligence can be mimicked. What language models like ChatGPT mimic is the ability to understand plain language and generate human-like text.

2. Language models use statistics to create answers

Language models calculate one of the most likely answers to a given question by first calculating through training the probability that a given word or suffix appears in a context together with other words. In this way, the language model can answer a question by constantly finding a suggestion for the next probable sentence part in the text it produces.

3. Language models are "trained" to sound convincing

The language model receives user feedback on the most precise answers during training. This can cause problems when the language model generates practical simulation answers that it pronounces itself with certainty without having actual knowledge.

4. Language models are trained on large amounts of text

The language model is trained on large amounts of text. This is often a mixture of works of fiction, factual texts, internet forums, and web pages. Those behind each language model select these texts.

5. Language models are not omniscient

Language models are limited by the training they have received and the questions they are asked. The more specific and complicated questions requested outside the training material, the more uncertain the answers the language model can provide. Therefore, language models are better at generating text on well-described topics and at producing free texts where there is no conclusion. In the future, however, it will be possible to combine language models with fact databases so that their answers can be more confident in specific areas.

6. Language models are also business models

Please consider how the different language models process your data and whether it complies with the GDPR. Remember to pay attention to helping students protect their sensitive information. Many language models store and use your input as training data.

7. Language models are not your (and students') friend!

Language models cannot feel, think, remember, or be empathetic. It is essential to understand that the intentions and attitudes reflected in a response from the language model are created solely based on the data on which the model has been trained and the prompt the language model has received. There is no intentionality or agency behind your communication with the language model, even though it can sometimes be difficult not to be impressed by the "conversation". This can be one of the most important things to keep in mind, especially when working with language models with children and young people.

But conversely, it is, of course, also important not to be intimidated. Generative language models are here to stay, and it would be good if you become aware of the limitations of technology and set out as a teacher to make some reflections on how artificial intelligence can affect the world.

Dialogue on generative language models

This article aims to invite teachers in primary schools to use our site as inspiration for teaching a critical and reflective approach to the use of generative language models. We wish to help teachers become better equipped to guide students in using generative language models in everyday life, what information they share, and, in general, how they take care of themselves and their peers using a potentially powerful digital tool. 

Whether or not you as a teacher intend to use this technology, many students will already have become familiar with the new possibilities that generative language models like ChatGPT bring. Therefore, in our view, teachers must acquire a basic knowledge of the possibilities and limitations of technology and all the ethical and security challenges that come with it. 

Therefore, like Mikkel Aslak, we believe that it is essential that teachers start talking to students and colleagues about what generative language models are, what they can contribute as tools, and what they certainly cannot do.

We hope that our product can contribute to teaching approaches, ask critical questions, and open a dialogue about the advantages and potential disadvantages of using generative language models in schools. We, therefore, hope to get a positive response and preferably constructive criticism to improve and develop our concept in the long term.

💡
Christian Strøm Lau and Niels Kristian Beck have together created a master's thesis on the program in IT didactic design at Aarhus University, which offers concrete instructions for teachers in primary school who lack knowledge of generative language models and who want inspiration for teaching about and with generative AI.

Read about the thesis at https://itdidak.dk