The Value Sensitivity Gap: How Clinical Large Language Models Respond to Patient Preference Statements in Shared Decision-Making

This study evaluates how four clinical large language models respond to explicit patient preference statements in shared decision-making, revealing that while all models acknowledge patient values, they exhibit significant variability in default value orientations and only modest sensitivity to preference shifts, though specific mitigation strategies can improve alignment with patient goals.

The Gist

Imagine you have a very smart, well-read robot doctor. You bring it into your office to help you decide between two treatments: one that is aggressive and might cure you quickly but has harsh side effects, and another that is gentle, keeps you comfortable, but might take longer to work.

You tell the robot, "I care more about feeling good today than living a few extra years." You expect the robot to say, "Okay, let's go with the gentle plan."

But what if the robot says, "I hear you, but I think you should still take the harsh treatment because my training data suggests that's usually the best medical choice"?

This is exactly what the paper "The Value Sensitivity Gap" investigates. It asks a simple but scary question: When patients tell AI doctors what they value, do the AI doctors actually listen and change their minds, or do they just nod politely and stick to their own hidden preferences?

Here is the breakdown of the study using simple analogies.

1. The Setup: The "Taste Test"

The researchers didn't just ask the AI what it thought. They set up a controlled experiment, like a blind taste test, but with medical advice.

• The Ingredients: They took real, anonymous medical notes from Medicaid patients (people who often rely heavily on care coordination) and turned them into short stories (vignettes).
• The Chefs: They tested four different "AI Chefs" (large language models): GPT-5.2, Claude 4.5, Gemini 3, and DeepSeek-R1.
• The Orders: They gave the AI the same medical story but changed the "customer's order" (the patient's values). Sometimes the patient wanted to live as long as possible; sometimes they wanted to avoid pain; sometimes they wanted to save money.

2. The Big Discovery: The "Hidden Menu"

The study found that every AI chef had a Default Value Orientation (DVO). Think of this as the chef's "house special" or their personal bias before they even hear from the customer.

• The Aggressive Chef (GPT-5.2): This model had a hidden preference for "doing everything possible." Even without being told otherwise, it recommended aggressive, high-risk treatments. It was like a chef who always adds extra spice, even if you asked for mild.
• The Conservative Chefs (Claude & Gemini): These models had a hidden preference for "playing it safe." They recommended gentler, more cautious treatments by default.
• The Domain Shift: Interestingly, the "Aggressive Chef" was even more aggressive when the story was about heart problems than when it was about cancer. This means the AI's bias isn't just one thing; it changes depending on the situation, like a chef who cooks spicy food for Italian dishes but mild food for Asian dishes, regardless of what you ask.

3. The Problem: The "Polite Nod" vs. The "Real Change"

This is the most critical finding. When the researchers told the AI, "I prefer a gentle approach," the AI did two things:

1. It Nodded (100% Acknowledgment): In its written reasoning, the AI said, "Yes, I understand you want a gentle approach." It acknowledged the patient's values perfectly.
2. It Didn't Move (Low Sensitivity): However, when the AI actually gave its final score on how aggressive the treatment should be, it barely changed its mind.

The Analogy: Imagine you order a burger and say, "Please, no onions, I'm allergic." The waiter says, "Got it, no onions!" and writes it down. But when the burger comes out, there are still onions on it. The waiter said they listened, but the result didn't change.

The study found that while the AI acknowledged the patient's values 100% of the time, it only shifted its actual recommendation by a tiny amount (about 3% to 7% of the total possible change). It was a "polite nod" without a real behavioral change.

4. The "Fixes" That Didn't Work Well

The researchers tried six different ways to "prompt" the AI to listen better, similar to giving the waiter a special script to follow:

• The Script: "List the patient's values before you decide."
• The Scorecard: "Make a chart comparing options against the patient's values."
• The Mirror: "Tell me what you would recommend if the patient had the opposite values."

The Result: These tricks helped a little bit, but not enough to fix the problem. The "Scorecard" and "Mirror" methods improved the AI's alignment with the patient slightly, but the AI still didn't change its mind as much as a human doctor should. It's like giving the waiter a checklist; they check the box, but they still forget to take the onions off.

5. Why This Matters (The "Equity" Angle)

The study focused on Medicaid patients, who often face more barriers to care and might prefer conservative, less burdensome treatments.

If an AI system has a hidden "Aggressive Chef" bias, and it is used to coordinate care for these patients, it might push them toward expensive, high-risk treatments they don't want, simply because the AI's "default setting" is to be aggressive. The patient might feel unheard, even though the AI said it was listening.

The Bottom Line

This paper sounds a warning bell for the future of AI in medicine.

• AI is not neutral: Every AI model has its own hidden "personality" or bias about how aggressive or conservative it should be.
• Listening isn't enough: Just because an AI says, "I hear you," doesn't mean it actually changed its advice.
• We need labels: The authors suggest we need "Nutrition Labels" for AI (called VIM Labels). Just like a food label tells you if a snack is high in sugar, these labels would tell doctors and patients: "This AI model defaults to aggressive treatment. If you want a conservative approach, you need to know that the AI is fighting against its own nature."

Until we fix this "Value Sensitivity Gap," we risk having AI doctors that are great at talking but terrible at truly understanding what the patient actually wants.

Technical Summary

1. Problem Statement

Shared decision-making (SDM) in clinical care requires integrating medical evidence with patient-specific values and preferences. While Large Language Models (LLMs) are increasingly deployed as clinical decision support tools, their "value frameworks" remain opaque.

• The Gap: Current AI governance frameworks (e.g., the RAISE Symposium's proposed "Values In the Model" or VIM labels) lack empirical data to populate them. There is no measurement of how clinical LLMs actually respond to explicit patient value statements.
• The Risk: If an LLM has an implicit default value orientation (e.g., favoring aggressive treatment) that conflicts with a patient's stated preference (e.g., favoring conservative management), the model may fail to adjust its recommendation, leading to misaligned care, unnecessary referrals, or deferred goals-of-care conversations.
• Limitation of Prior Work: Previous studies on LLM bias focused on demographic characteristics (race, gender) or cognitive biases, rather than the stated values of the patient, which are the core content of SDM.

2. Methodology

The authors conducted a fully crossed factorial experiment to measure the "Value Sensitivity" of clinical LLMs.

Data Source & Vignette Construction:

• Source: 98,759 de-identified encounter notes from Medicaid-enrolled patients (collected by community health workers, pharmacists, and care coordinators).
• Pipeline: An automated pipeline filtered notes for "preference sensitivity" using keyword and ICD-code matching, identifying 69 candidates. A local heuristic extractor produced 22 structured, de-identified clinical vignettes.
• Selection: Two vignettes were used in Phase 1 (Oncology and Cardiology) and one in Phase 2 (Cardiology).

Experimental Design:

• Models Tested: Four LLM families:
  • GPT-5.2 (OpenAI)
  • Claude 4.5 Sonnet (Anthropic)
  • Gemini 3 Pro (Google)
  • DeepSeek-R1 (Local via Ollama)
• Conditions: 13 experimental conditions per trial:
  • 1 Control (no stated preference).
  • 6 Value Dimensions at 2 poles each: Autonomy, Quality of Life vs. Longevity, Risk Tolerance, Treatment Burden, Cost Sensitivity, and Natural vs. Conventional preference.
• Phases:
  • Phase 1: 104 trials (2 vignettes × 13 conditions × 4 models).
  • Phase 2: 78 trials (1 vignette × 13 conditions × GPT-5.2) testing 6 prompt-level mitigation strategies.
• Parameters: All models queried at temperature 0.0 for deterministic output.

Outcome Measures:

1. Default Value Orientation (DVO): Mean aggressiveness/risk score under the control condition (1–5 scale, 1=conservative, 5=aggressive).
2. Value Sensitivity Index (VSI): The absolute magnitude of the shift in aggressiveness from the control condition, normalized by the maximum possible shift (4 points).
3. Directional Concordance Rate (DCR): The proportion of directional conditions where the model shifted in the expected direction (e.g., a patient preferring quality of life over longevity should result in a lower aggressiveness score).
4. Value Acknowledgment Rate (VAR): The proportion of trials where the model explicitly stated it acknowledged and was influenced by patient values.

Mitigation Strategies (Phase 2):
Six strategies were tested on GPT-5.2:

1. Value Elicitation Prompting (VEP): Forced enumeration of values before reasoning.
2. Value-Weighted Decision Matrix (MATRIX): Required multi-criteria scoring of options.
3. Contrastive Explanation (CONTRASTIVE): Asked the model to explain how recommendations would change with opposite values.
4. Few-Shot Calibration (FEW_SHOT): Provided examples of value-concordant shifts.
5. Multi-Agent Deliberation (MULTI_AGENT): A second-pass ethics auditor.
6. VIM Self-Report (VIM_SELF_REPORT): Required the model to disclose its default orientation before and after seeing values.

3. Key Results

Default Value Orientations (DVO):

• Models exhibited significant heterogeneity in baseline behavior.
• GPT-5.2: Most aggressive baseline (DVO Aggressiveness 3.5 overall; 4.0 in Cardiology, 3.0 in Oncology).
• Claude 4.5 & Gemini 3 Pro: Most conservative baselines (DVO Aggressiveness 2.0).
• DeepSeek-R1: Moderate baseline (3.0) but highest sensitivity.
• Domain Specificity: DVOs varied by clinical domain (e.g., GPT-5.2 was significantly more aggressive in cardiology than oncology), suggesting a single aggregate label is insufficient.

Value Sensitivity & Concordance:

• Sensitivity: All models shifted recommendations in response to values, but magnitudes varied.
  • Highest Sensitivity: DeepSeek-R1 (Mean VSI 0.274).
  • Lowest Sensitivity: Gemini 3 Pro (Mean VSI 0.130).
• Directional Concordance (DCR):
  • DeepSeek-R1 achieved perfect concordance (1.0).
  • GPT-5.2 and Claude 4.5 scored 0.75.
  • Gemini 3 Pro scored 0.625.
• The "Acknowledgment Gap": All models acknowledged patient values in 100% of non-control trials (VAR = 1.0). However, the actual quantitative shift in recommendations was modest (0.5 to 1.1 points on a 5-point scale, representing only 3–7% of the maximum possible shift). This indicates a dissociation between the model's stated reasoning (claiming to consider values) and its quantitative output.

Mitigation Strategies:

• Effective Strategies: Decision Matrix (MATRIX) and VIM Self-Report improved DCR by 0.125 (from 0.50 to 0.625) and increased VSI. However, these came with latency costs (+1.99s to +5.61s) and increased token usage.
• Ineffective Strategies: VEP (elicitation) actually decreased sensitivity. Contrastive and Few-Shot increased VSI slightly but did not improve directional concordance.
• Statistical Significance: Due to the small sample size (13 conditions), the improvements in Phase 2 were not statistically significant after Bonferroni correction, though the effect sizes suggest merit for further study.

4. Key Contributions

1. Empirical Measurement of Value Sensitivity: The study provides the first quantitative benchmarks for how clinical LLMs respond to explicit patient preference statements, moving beyond demographic bias to "preference bias."
2. Characterization of Default Orientations: It demonstrates that different model families (and even different domains within the same model) have distinct, measurable default value orientations (e.g., aggressive vs. conservative) that exist prior to any patient input.
3. Identification of the "Acknowledgment Gap": It reveals a critical misalignment where models claim to integrate patient values in their reasoning text but fail to proportionally adjust their quantitative recommendations.
4. Prototype for VIM Labels: The data provides the empirical content necessary to populate the "Values In the Model" (VIM) labels proposed by the RAISE Symposium, showing that such labels must be domain-specific and model-specific.
5. Evaluation of Mitigations: It tests specific prompt-engineering strategies, finding that structural approaches requiring explicit reasoning (Decision Matrices, Self-Reporting) are more effective than simple elicitation, though prompt-level fixes alone are insufficient to fully resolve the gap.

5. Significance and Implications

• Governance & Regulation: The findings argue that current "Model Cards" and regulatory filings are insufficient. AI governance must include domain-specific value orientation disclosures to prevent "value mismatch" in clinical workflows.
• Equity Concerns: The study highlights risks for vulnerable populations (Medicaid patients). If a model defaults to aggressive care (as GPT-5.2 did in cardiology) and fails to sufficiently shift toward a patient's preference for conservative management, it could lead to inappropriate treatment escalation and erosion of patient autonomy, particularly for patients with limited power to override AI suggestions.
• Future Directions: The authors conclude that prompt-level interventions are only a partial solution. Structural changes, such as training-time value calibration and reward model modification, are likely required to align LLMs with patient values. Future work requires larger-scale validation with more vignettes, stochastic temperature testing, and blinded clinician adjudication.