Artificial intelligence in teaching

On 16 February 2023, the University of Helsinki’s Academic Affairs Council published guidelines on the use of AI in teaching (pdf).

The guidelines state that course coordinators decide the extent to which AI can be used in their courses. Please note that AI can never be used for maturity tests.

In these guidelines, AI refers, in particular, to language models based on machine-learning methods (esp. Generative Pre-trained Transformer, or GPT) and the easy-to-use and widely available user interfaces built on them. At the time of writing these guidelines, the model most talked about was ChatGPT, the most advanced and easy to use of the open-access software solutions based on language models. AI as such does not produce new information; its outputs are based on the data (e.g., websites) it has been trained on.(1)

AI generates responses guided by prompts that can include information on, for example, the text to be used as the basis or the structure of the response. The strength of AI-based applications lies in the extensive and varied data used to generate responses. AI solutions based on language models produce text by inferring the next word from the data used to train them; they do not actually understand the meaning of the text or produce new information. The output should therefore be considered critically and used mainly as a basis for ideas or texts. 

In general, AI tools such as ChatGPT produce better quality if the genre of the text is common and well established – haikus, for instance. But the more specific the topic, the likelier it is that the text generated is factually incorrect or poor quality. AI solutions can often write convincing texts on introductory-level topics, but have more difficulty producing a high-quality text on a specific research issue. In addition, when ChatGPT is asked to indicate the sources used in a text, it generates convincing references and bibliographies, but invents the sources – it is simply imitating the conventions of the genre.

AI guides teachers to design learning assignments that are more applied in nature and geared to higher levels of expertise (cf. Bloom’s taxonomy). With easier and faster information-seeking, learning focuses instead on the analysis and assessment of the content acquired, the justification of interpretations made and the creation of new knowledge. This means teachers have to carefully consider the learning assignments they give to students. AI-generated output can easily answer a simple open-ended question or at least provide a good basis for a general essay assignment. This means that practising appropriate referencing becomes increasingly important.

AI helps produce responses that sound convincing to assignments asking the student to reflect on learning material. It is not enough to ask students to keep a learning journal outlining their challenges and including critical reflection. This is such a well-established type of text that AI tools such as ChatGPT can generate a response largely on the student’s behalf if the student attaches their lecture notes to a prompt and requests a reflective summary.

Hence, the teacher should experiment with AI tools to see the responses they produce to the planned learning assignments, and then consider how the assignments could be developed with appropriate instructions to make learning more meaningful. In several of the examples listed below, AI can help students complete assignments more quickly. This means you should outline assignments so that students are guided towards achieving the targeted learning outcomes of the course regardless of the means they use to complete the assignment.

Examples of learning assignments:

• AI-generated text can be used as a basis, but the student’s own comments and thoughts must be clearly separated from it.
• Students can be asked to apply information in case-based or problem-based learning.
  • As part of the assignment, the original and follow-up questions posed to the AI tool must be documented. At the same time, students practise using AI as part of information-seeking and assess the output critically.
  • Students can be asked to use source material provided by the teacher and report their reasoned solutions so as to make the learning process visible.
• The conventions of academic writing should continue to be followed: sources must be mentioned, direct quotations clearly indicated and the student’s own comments distinguished from others’ work.
• When exploring and seeking information on a new topic, the processing of information can be supported with follow-up questions (e.g., students can be asked for applied examples or to consider any unclear issues). AI can help with this, but the role of the assignment in the course should be minor or the assignment should not affect assessment.
• Process writing and the use of interim feedback
• Mind maps and other visualisation as part of information acquisition assignments. As AI applications can also generate images (see, e.g., DALL·E 2), you should test what this means in practice.

Instructions for AI use

The basic premise is that AI-generated text cannot be presented as one’s own. It can be referred to under the name of the AI model and the date of the generated text. Students are responsible for their own texts, and plagiarism is strictly prohibited.

Students can be instructed to use AI in a way that fits the course, for example, as follows:

• Use permitted: AI can be used freely to seek information and assist in completing assignments. If a language model has been used to produce the work submitted, the student must indicate in writing which model they have used and how.
• Limited use: AI can be used in accordance with detailed instructions:
  • For example: Refer to AI-generated text by the name of the AI model and the date the text was generated.
  • Additional instructions, if any: In your answer, mention the prompts you gave to the AI tool.
  • Follow-up questions as part of the assignment, for example: Consider how well the answers provided by the AI model match the assignment given.
• Use prohibited: AI cannot be used. In practice, verifying this is challenging, although certain services can assess the likelihood of a text being AI generated.

When information acquisition and assignment completion become easier, students may not learn to the same extent or in as much depth as they would if they read a certain amount of background information. If the assignment is such that AI can complete it easily, it should not significantly affect the overall course grade.

Learning depends increasingly on what happens during the course in addition to information-seeking. Can the learning process (what and how I learn) be made more visible to both the students themselves and the teachers? If students learn less at the information acquisition stage, learning can be supported with other assignments during the course.

Examples of assignments that guide students to explore, understand and apply knowledge and skills:

• Reading groups
• Panel discussions, role-based discussions
• Seminar-style work: Presentations, videos and teaching sessions given by students, followed by questions from the other participants
• Small-scale intermediate tests and questions assessed automatically in Moodle can be used as before. Although answers can be found in the course material and online, these can support the review of specific questions and themes and enable students to assess their own skills.

Examples of how to deepen learning after the information acquisition stage:

• In flipped learning (see, e.g., Talbert 2017), information is acquired and the topic explored during remote periods, while the information is applied and the topic examined collaboratively in depth in contact teaching.
• Peer assessment and self-assessment help delve deeper into the topic.
• Online discussions based on previous messages; if necessary the teacher can guide the discussion with follow-up questions in a direction relevant to learning.
• Project-based courses

If you wish to hold an examination ensuring that students know key processes and concepts by heart, you can choose between Examinarium examinations and invigilated on-site Moodle examinations.

AI can also be used to help plan courses. You can ask an AI tool to generate ideas for the course syllabus, a large number of multiple-choice questions on a selected topic or textual material for students to assess as part of an assignment.(2)

1. So far, no model with a real-time connection to online material is freely available. For example, ChatGPT knows little about events that took place after 2021. Its developers have not published a full list of the data used to train it, but you can consult Alan D. Thompson’s useful summary of the ChatGPT training data.
2. An example from a take-home examination by Bodo Steiner, a teacher at the Faculty of Agriculture and Forestry: “I will generate a set of ChatGPT-generated essays on three topics, the students can choose a given essay that fits their interest, and then they will be tasked to write an essay themselves on the very same topic, discussing the strengths and weaknesses of the ChatGPT-generated essay.”