In the context of teaching and learning, safe and (legally) compliant access to generative AI systems and models is an important aspect, especially as it is necessary in order to actively engage with the technology.

There are thousands of generative AI systems and models in total. In addition to well-known commercial systems, such as ChatGPT (Plus) from OpenAI or Midjourney, there are also non-commercial, open systems and models. These are becoming increasingly better and now account for real alternatives for many applications.

The decision about which systems and models should be purchased and deployed depends not only on the intended use but also on the performance, user-friendliness, availability and accessibility, as well as sustainability, security, costs, and the distribution of the system or model. In addition, ethical, legal and moral implications must be taken into account.

For example, OpenAI’s commercial system ChatGPT (Plus) or its API could be a suitable option due to its widespread adoption, market dominance, performance, and ease of use. However, owing to insufficient data protection guarantees, the use of ChatGPT (Plus) in teaching and learning contexts – particularly for students – remains highly problematic. Furthermore, the question of appropriate licensing models has not yet been resolved. More broadly, the potential positive and negative consequences of choosing a commercial and de facto monopolistic solution over alternatives such as open-source systems should also be taken into account.

If a decision is made to provide access to commercial systems such as ChatGPT (Plus), three options are possible:

1. The use of individual licences, which shifts large parts of the responsibility to the users (i.e. to the students and lecturers). This may be a viable approach, for instance, when the system is used only for demonstration purposes or when students work exclusively with model-generated outputs.
2. The purchase of organizational or campus-wide licences directly from providers (e.g. Google, Microsoft, OpenAI). In many cases, however, this option is often difficult or currently infeasible, as many questions remain unresolved, particularly with regard to licensing models and the legal framework (e.g. GDPR).
3. The use of third-party services such as fobizz classrooms or DieSchulApp, which offer data protection-compliant access to the systems.
    

As open source systems and models become increasingly powerful, they are quickly becoming a viable, secure, and (legally) compliant alternative for educational contexts. Many current systems and models can even be used locally without internet access (e.g. GPT4All).

Even if the open systems and models do not yet match the performance of their commercial counterparts in every respect (e.g. with regard to plugins and agents), they are already highly usable: they involve fewer licensing and cost issues, and data-protection risks can be reduced to an absolute minimum.