How Clinical Framing Influences Medical AI
Artificial intelligence is increasingly being used to support medical decision-making. Large Language Models (LLMs) such as ChatGPT are able to generate differential diagnoses, identify red flags, and recommend diagnostic steps within seconds. But an important question remains:
How objective are these recommendations?
Our project investigates whether the way a clinical case is presented influences the diagnostic reasoning of different AI models. In other words: Does changing the wording of a case change the diagnosis?
Our Research Question
We developed three different medical case vignettes. Each case was presented in three different versions:
- a neutral version,
- an anchoring version,
- and a framing version.
Although the medical core of each case remained identical, small contextual changes were introduced to investigate whether the AI’s diagnostic reasoning would change.
Each prompt was analyzed using four different Large Language Models, resulting in a total of 36 AI-generated responses.
To compare the answers systematically, we are conducting a structured qualitative content analysis based on Mayring’s methodology. We examine changes in differential diagnoses, red flags, diagnostic recommendations, and the prioritization of medical reasoning across the different prompt versions.
Why Does This Matter?
As AI becomes increasingly integrated into healthcare, understanding its strengths and limitations is essential.
Our findings suggest that AI models generally retain important differential diagnoses and patient safety considerations. However, the wording and context of a prompt can influence which diagnoses are prioritized first and how diagnostic pathways are structured.
Rather than changing what the models know, prompt design appears to influence how this knowledge is activated and prioritized.
Looking Ahead
Our goal is to contribute to a better understanding of clinical prompting and responsible AI use in medicine. By exploring the effects of anchoring and framing, we hope to support the development of more robust and transparent AI-assisted clinical decision-making.
This project combines behavioral science, medicine, and artificial intelligence demonstrating that sometimes, even small changes in wording can have a meaningful impact on decision-making.
