When the opportunity arose within the Nudging module to write a scientific publication instead of a classic nudge project, it quickly became clear to us that we wanted to take this opportunity. What particularly appealed to us was the chance to gain insight into the scientific publication process and to experience for ourselves how it actually works. At the same time, this way of working was completely new to us. Unlike a classic term paper, there was neither a specific research question nor a fixed structure. The only thing that had been determined was the topic Human Biases in AI. We developed the specific focus and the direction in which our contribution should evolve ourselves in several brainstorming sessions.
We first considered which developments actually occupied us in our everyday lives. We quickly noticed that we personally had long since stopped using Large Language Models (LLMs) such as ChatGPT only for term papers or other study related tasks. Increasingly, we also turn to them for health related questions, whether to assess symptoms, obtain information about medical conditions, or prepare for a visit to the doctor. This led us to ask whether we were alone in using LLMs in this way or whether they are increasingly taking on the role of digital health advisors for many people.
A look at current surveys and studies quickly showed that this is by no means an individual phenomenon. According to a Bitkom survey, 73% of people already research health related questions online as of 2025, and almost every second person has already used an AI chatbot such as ChatGPT, Gemini, or Copilot for health related questions (n = 1145).
The problem, however, is that LLMs are not neutral. They can exhibit a wide range of different biases that systematically influence their responses. Studies show that responses can differ, for example, depending on gender or background (Levartovsky et al., 2025; Pendse et al., 2025; Bouguettaya et al., 2025). This is particularly relevant in the health context because people make decisions about their own health based on such information. This was precisely the issue we wanted to present in an accessible way and discuss in our first exposé. From the very beginning, our goal was to make the relevance of the topic tangible. We first wanted to show that LLMs can exhibit systematic biases in the health context and explain why these are becoming increasingly important in light of their growing use. Building on this, we examined how such biases can be identified, which strategies can promote a critical and reflective approach to LLM responses, and how potential biases can already be reduced through the deliberate design of prompts.
Once we had developed our basic idea, we focused intensively on the methodology and, in parallel, began searching for a suitable journal. In retrospect, however, this approach was not particularly effective. We were still too firmly rooted in the mindset of classic seminar papers and focused on content related details at an early stage, even though the most important step should initially have been the selection of a suitable journal. Only through the repeated guidance of Prof. Dr. Harff were we able to move away from this approach and initially shift our focus entirely to the search for a journal.
In the process, we became aware that the choice of a suitable journal significantly determines the entire subsequent publication process. Finding a journal that matched our contribution thematically and whose formal requirements we were able to meet proved to be considerably more difficult than expected. It was particularly challenging for us to consistently think from the perspective of the respective target group. We were not familiar with this way of working from our previous academic work.
At the beginning, we mainly contacted medical journals. Based on the feedback we received, however, we realized that this was probably not the right direction for our contribution. Unlike medical students, we do not have the necessary medical expertise to write a primarily medical academic article. Over time, we therefore broadened our scope and contacted not only medical journals but also general academic and education oriented journals. The idea of raising awareness of the responsible use of LLMs in the health context always remained at the center of our work.
We were therefore all the more pleased when we finally received positive feedback on our exposé from a journal in the field of continuing education. By this point at the latest, we realized how much time we could have saved at the beginning if we had prioritized the search for a journal earlier. Many of the content related details we had previously discussed now had to be adapted anyway to the editor’s requirements and the journal’s guidelines.
As a result, we engaged even more intensively with the target group and further developed our contribution accordingly. Our focus is now on making biases in LLM responses understandable and tangible while also presenting concrete courses of action. In particular, we aim to support educators in passing on this knowledge as multipliers and in promoting a reflective approach to AI supported health information.
