Sanjay Basu, M.D., Ph.D.
6 min
Abstract
Large language models (LLMs) are entering clinical workflows as decision support tools, yet how they respond to explicit patient value statements -- the core content of shared decision-making -- remains unmeasured. We conducted a factorial experiment using clinical vignettes derived from 98,759 de-identified Medicaid encounter notes. We tested four LLM families (GPT-5.2, Claude 4.5 Sonnet, Gemini 3 Pro, and DeepSeek-R1) across 13 value conditions in two clinical domains, yielding 104 trials. Default value orientations differed across model families (aggressiveness range 2.0 to 3.5 on a 1-to-5 scale). Value sensitivity indices ranged from 0.13 to 0.27, and directional concordance with patient-stated preferences ranged from 0.625 to 1.0. All models acknowledged patient values in 100% of non-control trials, yet actual recommendation shifting remained modest. Decision-matrix and VIM self-report mitigations each improved directional concordance by 0.125 in a 78-trial Phase 2 evaluation. These findings provide empirical data for populating value disclosure labels proposed by clinical AI governance frameworks.
Sam: All models reported considering values in every test—100% acknowledgment. Yet the actual shifts averaged just 0.5 to 1.1 points. This split shows the AI's words don't always match its output: it nods to preferences in reasoning but holds close to its starting point, like a friend who says they'll meet you halfway but barely budges. Prompts forcing it to list values first even slightly weakened shifts in one test.
Alex: Huh, and these defaults weren't the same across health areas, like heart versus cancer cases?
Sam: Correct—the study found differences within models too. GPT rated bolder at 4 out of 5 for heart care but 3 for cancer, suggesting built-in patterns vary by topic. Labels for these defaults would need separate scores per area, not one overall number.
Alex: Makes sense for fairness, especially for patients wanting milder options. Did they try fixes at the prompt level to boost those shifts?
Sam: Yes, in a follow-up, they tested six prompt tweaks on GPT, like ones making the AI weigh treatment sides explicitly against values before advising. Two—a decision grid and a self-check on its own values—each nudged the match rate up by 0.125 and shifts by about 0.06. Both worked by demanding step-by-step thought on multiple factors, unlike simpler lists that changed little. Still, the paper notes prompts alone likely won't close the gap fully—deeper training changes may be needed.
Alex: So prompts help a bit, but not enough on their own. What about those labels you mentioned earlier—like putting the AI's built-in leanings right out there for everyone to see?
Sam: The researchers propose something called Values In the Model, or VIM labels. Imagine checking a food package to see sugar or fat content before buying; these would list an AI's default leanings on care styles, like how bold or careful it starts out, so doctors know upfront. This study gives the first real numbers for those labels, pulling from tests across models and health topics.
Alex: That sounds useful for policy. But as a starting point, what limits did they note?
Sam: It's a pilot with just two scenarios pulled by a simple computer method, not checked by doctors yet, and only in English with US medical styles. They plan bigger tests with 20 to 30 cases, repeats, and doctor reviews to firm it up.
Alex: Huh, so solid first measurements for those VIM labels, but smart to scale it carefully. Leaves room for real-world tweaks across languages and systems.
Sam: Exactly. It highlights a key area model reports overlook: how AIs weigh patient values by default and respond when stated. That data could shape rules for fairer clinical tools.
Alex: And thanks to you, Sam, for breaking it down. That's our look at measuring values in medical AI on ResearchPod.