Blog · Higher education

Using AI in higher education – practical experiences from scientific writing Master’s courses

· 7 min read
No AI was used to write this text.

AI is a new challenge

The rise of AI has certainly constituted a major challenge for my established Master’s courses in scientific writing. The course is designed to teach the students the process of academic publishing by starting from a results figure. The students can pick one of two provided figures and need to come up with an underlying experiment that they will use for writing the paper. It is a simple 2x2-factorial design, so one example could be that iron supplementation increases serum iron levels in males but not in females or in young but not in old participants. The students then write the results section, the introduction, the discussion and all the rest – section by section, one part every week with me providing short feedback to each participant for maximal learning. While this is quite a lot of work for courses with up to 50 Master students, it was very rewarding to see how the students improved and the end results were astonishingly good. However, when AI chatbots became readily available, the course went downhill fast – students skipped the weekly feedback sessions and simply prompted the chatbot to write a full paper at the end of the semester for assessment.

Detection difficult

While I became quite good at spotting AI written manuscripts, there was no way to reliably detect AI use. Of course, we had the obvious issues with hallucinated references, which was an immediate fail grade for the student, but even with the most recent AI detection tools, there was no certainty. And irrespective of the factual content, I also noticed that the language quality was much better when the students used AI, as most of us are non-native speakers. Consequently, I decided to embrace the use of AI instead of prohibiting it. But it required a fundamental change in the course.

Three issues students need to learn

If one encourages the use of AI by students, there are a few aims that you want to achieve:

  1. AI use should not be an autopilot. You don’t want the students to upload the figure and say “here is the figure, come up with an experiment and write me a paper on it”. Not only will the quality be usually less than good, but there is no learning.
  2. Students need to be aware of the classic issues with AI. This includes training bias, which means the AI has been trained on a defined data set, and if that data set e.g. contained only data on male health, the output will reflect this. Secondly, most AI chatbots have a prompt bias (“sycophancy”), meaning they try to anticipate an answer that matches the prompt style and makes the user happy. After all, the AI companies probably want to sell their chatbots and if they are rude and totally honest, that could trigger a switch to a competitor’s model instead. And finally, and most importantly, depending on the model, AI can make up convincing references and sources for statements that later turn out to be fake. And that is not an isolated incident, a Nature report from 2026 finds that the recent literature is full of thousands of these hallucinated references.
  3. Students should be aware of the legal issues and at least have pondered the ethical aspects, which include not only the lack of learning experience but also the question whether it is still their work if AI has written it. The legal issues are somewhat of a lost battle: The vast majority don’t know that uploading research data to servers outside the European Union may violate University policies and of those who do, most don’t care and keep doing it anyway – even those in leading positions. Unfortunately, the alternatives are scarce – while some universities have local LLMs for their employees, these are usually still a far cry from the current frontier models and often without internet access to the newest literature.

Two things I aim for

What I now aim for is to convince the students to use AI as some kind of PhD-level sparring partner. They should use it to discuss their ideas and ask plenty of questions until they are certain they understand the output and can defend it. I think there is actually some learning potential, as most people don’t have any inhibitions to ask seemingly stupid questions to a chatbot, while in human interaction they might consider how it would reflect on them when they ask a question.

A second aim is to get them to compare their prompting. The individual sections for the scientific writing course follow a certain guideline for how to write for instance, an introduction. In brief, it should have a funnel structure, starting with an important global issue and moving down to the specific knowledge gap that my study will fill later. So if I can get the students to transfer this concept into the prompt so that the output matches the guidelines, while they did not do the actual writing themselves, they have grasped the general concepts of how an introduction should be written. And if they now discuss how to improve the prompt to get the best possible outcome, I got them to reflect on the topic – sure, it might be a different way than it was before, but maybe prompting is the new skill to teach!

How to assess students in times of AI

The assessment is somewhat more difficult now, and depends on whether it is a graded course or a simple pass/fail course. With pass/fail a list of a few simple must-have criteria is often enough. In the case of scientific writing, the no-gos are hallucinated references, lack of transparent declaration of AI use, plagiarism and not following the provided task (e.g. by using a completely different random figure, not having all the parts of the manuscript). It becomes more challenging for the actual graded course, but in that course we have additional tasks such as peer review of another student’s work, revision of their own work after peer-review etc. From my own experience, while the previous courses had a Gaussian bell-shaped grade distribution with usually a very good peak, we have now two peaks: manuscripts that are extremely polished (as AI improves style and language) and those that fail the basic requirements (lack an AI declaration or contain unverified and fake sources).

Summary

While the availability of AI is certainly a challenge for classic teaching, not only in higher education but also in schools, I believe that it can be a opportunity as well. It most certainly requires a rethinking of our teaching aims and the way we design courses and how we assess student performance. More importantly, we need to provide guardrails for the use of AI, and in my case of scientific writing it even required me to develop a completely new tool, PublicationGod.com – which supports scientific writing with AI use but in a transparent and structured way, designed to prevent hallucinated references and hosted on EU servers.

Tips for students on using AI in university courses:

  • Ask for clear rules, what you are allowed to do and what is required when you use AI.
  • Do not turn off your brain - AI is not your autopilot. Treat it as your sparring buddy, ask it anything you don’t understand – it won’t judge you. Try to catch it in a lie or try to convince it of the opposite to see how firm the argument really is.
  • Own the output, make it yours, check every fact and be able to defend it.
  • If you used AI, be transparent about it. You don’t need to be afraid of AI detection tools if you are upfront about your use.

Tips for teachers on using AI in university courses:

  • Start with a task that requires scientific thinking, and suggest using AI as discussion partner e.g. to design experiments.
  • Divide the work into smaller tasks with more specific aims, try to get the students to work on the prompt to achieve a better result – this teaches them a new skill and they learn the intended outcome as well.
  • Provide short feedback from your expert point of view so they can see how far AI got them.
  • Teach transparency, provide clear rules and clear failing criteria (e.g. hallucinated sources).
  • Allow the students to make mistakes using AI. Don’t provide the best prompts or the most suitable tools at the beginning, but at the end tell them the solution.
  • If you cannot provide a proper tool, try to build one. Get a Pro account for a frontier AI model, tell it what you want and vibe-code it your way, it doesn’t need to be production-ready to be helpful for students.

Draft your manuscript from your own figures

PublicationGod drafts results from your confirmed figures and helps you write an introduction and discussion that fit your target journal.

Start drafting
⌂ Start Paper Start Grant Presubmission Account & Credits About Blog Contact Terms Privacy Impressum Revoke