Can existing didactic values be transferred 1:1 into automatically generated learning sequences, or do new standards have to be developed for AI-supported learning units?

Jan Beinke: Since AI-supported systems manifest their own peculiarities and possibilities, 1:1 transfer is not possible as a rule. Some aspects of "traditional didactics", however, can also be applied in AI-supported learning units, such as the formulation of clear learning objectives, the consideration of various learning styles, and the inclusion of feedback and reflection. AI’s strengths, such as personalization and data-based decision making, should be used in the development of new didactic approaches based on the requirements of the individual learners.

Thorsten Krause: In undertaking this, it is important to strike a balance between human and AI-supported interaction in the learning process. AI can provide teachers with support in the identification of learning needs and the adaptation of learning content, while instructors retain responsibility for personal interaction, promoting empathy, and fostering social skills.

Generally speaking, the development of AI-enabled learning units should a collaborative effort among education experts, teachers, and learners to ensure that the end products are effective and pedagogically expedient.

How can these requirements be transferred?

Jan Beinke: Didactic demands are usually qualitative in nature, but there are two possibilities to enable AI systems to meet these requirements: Either they can be designed to meet the requirements implicitly, or they are quantified and given to AI as a criterion for optimization. For example, it isn’t possible to quantify learner autonomy directly, but it can be fostered through the systematic offering of multiple recommendations that guarantee the learners adequate scope of action.

Thorsten Krause: In research on "recommender systems", properties such as the diversity of the offer have already been successfully quantified. Likewise, the degree to which learning content is aligned with the learner’s level of competence can be ascertained via curricula and competence models. These criteria can be considered when generating recommendations. Nevertheless, it is important that experts ultimately evaluate whether the system actually implements these criteria successfully.

To enquire from another angle, are the didacticians required to make the necessary adaptations or the technicians?

Thorsten Krause: Both groups have to adapt in order to optimally exploit the potential of artificial intelligence in education. Didacticians need to adapt their knowledge of teaching and learning methods, as well as of learning theories, to the possibilities and particularities of AI-supported systems. They have to debate didactic approaches and standards that leverage AI technology’s strengths, while maintaining basic pedagogical principles.
For their part, technicians must come to terms with the requirements of didactics and learn how to deploy their technical skills in the development of learning systems that comply with both learner needs and the didacticians’ pedagogical specifications. They are responsible for creating user-friendly, efficient, and secure systems that adhere to privacy policies.

Jan Beinke: Business informatics can act as a mediating element between didactics and technology to optimize AI’s potential in education. The field has long been a component in the process of adapting domain-specific requirements to technical frameworks and has developed a variety of basic tools for this task.
Through their experience in aligning domain requirements to technical circumstances, Business IT specialists are in a position to understand and combine the needs of didacticians with the technicians’ professional skills. They can identify requirements, mutate them into technological solutions, and evaluate the systems developed.

In this mediating role, Business IT specialists support the collaboration between didacticians and technicians in the joint development of AI-supported learning systems that meet both learners’ requirements and the didacticians’ pedagogical specifications - and at the same time are user friendly, efficient, and privacy compliant. At the German Research Center for Artificial Intelligence, we undertake research and industrial projects in teams with this type of structure.

Are there already real-world examples that you would term "best practice"?

Thorsten Krause: Our experience with AI-lifecycle methods such as approaches based on CRISP-DM has been very good. On KUPPEL, we are using it to construct the recommender system for the recombined learning sequences, and on CLEVER, we used it to realize a recommender system for teaching content. In both projects, we collaborate with the didactics experts of Saarland University’s Research Institute Education Digital (FoBiD) and Didactic Innovations GmbH as implementation partners.

It is advisable to use other domains with similar issues as orientation as well: for example, the public broadcasters, who increasingly personalize their online offerings, while having to maintain their focus on their core values.

Do you think there should be a visible structural difference between individually created learning units and automatically generated ones so they can be identified accordingly, or does the creation process no longer play a role in the application?

Thorsten Krause: In some cases, it may make sense to distinguish between individually created and automatically generated content, but in others, it may have little relevance. One argument in favor of labeling the creation process is that learners and teachers develop awareness of how the content was generated and the role AI played in the process. This might help users understand the strengths and weaknesses of AI-powered content better - and prompt them to have reasonable expectations about the quality and accuracy of the information. Furthermore, the use of AI should be transparent, especially in sensitive domains such as education. When possible, future regulations should also be considered here.

Jan Beinke: It is our fundamental belief that content generated by artificial intelligence should be labeled as such. On the other hand, some contend that the creation process is initially irrelevant as long as the learning units are of high quality, didactically meaningful, and effective. In this case, it appears unnecessary to label the creation process since the quality of the learning material is paramount, regardless of whether it was individually created or automatically generated. In the end, technology, in this case artificial intelligence, is not an end in itself, but merely a means to an end: to optimally support learners on their individual learning path.