What Kind of Doctor Is Your AI?

Chaotic good, neutral, or evil?

When LLMs provide clinical decision-support, the discussion is usually around accuracy, confabulations, and various biases present in the training set. Less focus is put on the LLM context tokens – that is, the “hidden” layer creating the attitudes and sycophancy of Claude, ChatGPT, Gemini, etc.

This fun little thought experiment looks at our AI agent future and poses the question: what sort of “persona” should a clinical AI have? And, between various personas, what effect will it have on decision-making?

Using two axes of ethical priorities and cognitive regulation, these authors created nine different personas whose prompts are replicated here:

These personas were then fed questions based off various clinical cases and vignettes, leading to 5,000,000 clinical yes/no “decisions” for evaluation (“obtain additional imaging”, “proceed with patient discharge”, etc.). And, across several different domains of questions with these yes/no decisions, the “persona” impacted the likelihood of taking an affirmative action:

This is just a thought experiment and simulation, but it likely reflects a necessary consideration regarding the underlying framework of AI agents and the downstream effects on decision-making.