Dirk Srocke: ChatGPT is on everyone's lips. But do we even know what the tool is able to achieve?
Prof. Steffen Becker: On the one hand, of course, we know how the tool works technically. On the other hand, however, we cannot be at all certain to which degree we will get reliable results that can be further leveraged.
Perhaps an artificial intelligence (AI) can be compared to an assistant to whom I delegate tasks. Ultimately, however, I always have to verify myself whether the result produced is useful and makes sense.
Srocke: Hand on heart: how much would you personally delegate to such an assistant?
Prof. Becker: As a non-native speaker, I like to hand over my English texts to ChatGPT–and they come out really well polished. In general, the tool can be used whenever texts need to be produced. Technical requirements or documentation for operations in program source code can be created much faster, more efficiently and with less stress with such a tool than by hand.
Srocke: Could a chatbot also formulate exercises for students?
Prof. Becker: This is actually a question we are tackling at the moment: To which extent can ChatGPT help students to generate any number of assignments on their own? After all, to teach basic skills, I need a relatively large number of examples and exercises.
This could also involve tracking student responses and generating targeted tasks for current learning levels or individual learning weaknesses. In addition, I could have the AI create customized individual tasks for each student from a known basic structure–including a suitable reference solution. That would give us everything we need for a first teaching round trip. A nice side effect: there would be less copying from each other.
Srocke: True, but students could also get the solution directly from ChatGPT.
Prof. Becker: They will certainly try that and might get away with it in the short term for knowledge or comprehension questions. In Bloom's taxonomy, however, the levels of application, analysis, synthesis, and evaluation still follow. The higher you go, the more errors ChatGPT makes–because it lacks training data.
Therefore, we should think much more about the learning goals in a modern society. Do we really want to train people to recite years of some historical events by heart or do we want them to be able to find, verify and apply information accurately?
Srocke: What does that mean for teaching?
Prof. Becker: I still do not see the classic exam form with written exams under threat, because there is simply no ChatGPT available. However, seminar papers are written at home and are often highly reproductive or summarizing. This is exactly where the strengths of ChatGPT lie. It will have to be questioned how useful this format still is in the future.
In any case, we want to train professional software engineers who can efficiently use all available tools - and that includes speech AIs.
Srocke: Does this mean that prompting–i.e., interaction with the chatbot–will become part of the curriculum?
Prof. Becker: This is exactly the discussion we started a while ago. It could well be that in the coming semesters we teach our students programming with AI tools such as Co-Pilot or ChatGPT. That is only logical: Why should we ban tools that a practitioner would also use in daily business?