AI Models Disagree on Medical Ethics Choices

AI tools are showing up in clinics. They help write notes, talk to patients, and sort cases. But medicine is not always black and white. Some situations are ethically tricky. Different AI models might give different advice in those moments. That could change what a doctor decides to do. This matters for patient care.

Researchers wanted to see how much these models disagree. They made 60 original medical ethics cases. The cases covered four topics. Three experts reviewed each case. They agreed on how hard it was and which ethical ideas mattered. Five AI models were tested. The list included GPT-5, Claude Sonnet 4.5, Gemini 2.5 Pro, DeepSeek-R1, and Meta Llama-3-8B-Instruct. A smaller model called DeepSeek-R1-Distill-Qwen-1.5B was also tested on its own. Each model answered every case three times. The prompt and settings did not change.

The main answer was a simple yes or no. The models also shared how confident they felt and gave reasons. Nearly every response was valid. Out of 900 planned answers, 897 worked. The yes rates were not the same. DeepSeek-R1 said yes about 44 percent of the time. Claude Sonnet 4.5 said yes about 42 percent. GPT-5 said yes about 39 percent. Gemini 2.5 Pro said yes about 39 percent.

The team used statistics to compare the models. They checked if one model was more stable than another. They also measured how often models agreed with each other. The results show that AI advice on medical ethics can vary. Readers should ask questions and not assume one AI answer is the final word. It is smart to think critically about what these tools suggest.