Let’s be honest about something: AI chatbots are not coming to our courses and our research. They are already here, and they have been for a while. ChatGPT, Claude, Perplexity, Gemini, whatever your tool of choice is, our students have been using these things for years now, often more skillfully than we have. And for a long time we, colleagues or the administration reacted to that the way people usually react to something they didn’t ask for: with rules. Bans. Clauses in exam regulations stating that the use of AI chatbots is forbidden.

It didn’t work. It was never going to work.

A few of our science colleagues said as much last year already, that these bans are basically futile, because students are simply further along in their AI skills than the people writing the rules. And now the state of Bavaria has made the discussion moot anyway: AI chatbots are explicitly allowed in studies, including in exams. So the question isn’t “should we allow this” anymore. The question is what we, as researchers and lecturers, actually do about it now.

We’ve talked about this a lot internally over the past months, and our take, after quite some back and forth, is this: AI chatbots are genuinely great. Not in a hand-wavy, marketing-brochure way, but in a very concrete sense. They take a lot of the grunt work off our plates, the debugging, the boilerplate code, the first draft of a paragraph that just won’t write itself, and that frees us up to focus on the thing we’re actually supposed to be good at as researchers, which is thinking. Asking the right questions. Judging whether an answer makes sense.

But that’s exactly where the catch is. An AI chatbot will happily write code that runs, or a paragraph that reads well, without anyone involved actually understanding what’s going on underneath. And that’s a real risk, maybe the central risk, when it comes to teaching and training the next generation of researchers, whether they’re MSc students or PhD candidates.

So instead of ignoring or silently accepting it, we’ve decided to actively teach AI chatbot usage in our courses. We will encourage its use, but we’ll also be upfront about where it falls apart. The non-negotiable part, the thing we keep coming back to in every discussion, is this: the code, the analysis, the scientific argument, all of it has to be understood by the person submitting it, regardless of whether an AI chatbot helped produce it. And the results have to be checked critically. Not trusted because the chatbot said so confidently, because confidence and correctness are two very different things with these tools.

What we still need to figure out, and we don’t pretend to have this solved, is how to actually teach that mindset. Not just “you’re allowed to use it,” but the awareness of where it tends to go wrong, the fallacies, the subtle errors that look completely plausible, the moments where critical evaluation matters more than ever. That’s a teaching skill in itself, and it’s a new one for most of us. And we have to adapt our exam or student project tasks so that we evaluate understanding not the coding or final product itself. Honestly, we do not know yet how and what, but we are working on it and will refine it in the next years.

And this isn’t some niche side topic that only concerns chatbots in a browser tab. AI chatbots are already wired directly into our workflows. There are R packages built around them now. QGIS has plugins for it. With the more advanced, paid versions, you can apparently control software like Blender through text prompts alone. This is moving into the tools we use every day, not staying contained to a chat window.

So yes, exciting times. But also genuinely challenging ones, both from a teaching perspective and a scientific one. We’re starting to adapt our courses and our research practices now, and we’re fairly sure we’ll have to keep adjusting as these tools keep improving. There’s no fixed endpoint here, no final policy we write once and forget about. It’s going to be an ongoing conversation, and we’d rather be part of shaping it than reacting to it after the fact.