💬 NLP

Can "AI" Be a Doctor? A Study of Empathy, Readability, and Alignment in Clinical LLMs

This study evaluates clinical Large Language Models and finds that while they struggle to match physicians in semantic fidelity and readability without intervention, they function most effectively as collaborative tools that, when rewritten or rephrased, significantly enhance communication clarity and emotional tone for patients.

Original authors: Mariano Barone, Francesco Di Serio, Roberto Moio, Marco Postiglione, Giuseppe Riccio, Antonio Romano, Vincenzo Moscato

Published 2026-04-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mariano Barone, Francesco Di Serio, Roberto Moio, Marco Postiglione, Giuseppe Riccio, Antonio Romano, Vincenzo Moscato

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor trying to explain a complex medical condition to a patient. You have to walk a tightrope: you need to be accurate (so you don't give bad advice), clear (so the patient actually understands), and kind (so the patient doesn't feel scared or ignored).

This paper asks a big question: Can an Artificial Intelligence (AI) walk that tightrope as well as a human doctor?

The researchers didn't just ask the AI to "be nice." They tested it like a scientist, comparing AI answers against real doctors' answers using two different "playgrounds":

The Textbook Playground: Formal, written medical explanations (like from a government health website).
The Waiting Room Playground: Real, messy conversations between patients and doctors.

Here is what they found, broken down into simple concepts and analogies.

1. The "Robot Voice" Problem (Readability)

The Analogy: Imagine a human doctor explaining a broken leg. They might say, "You need to keep your leg still and rest." Now, imagine a robot trying to explain the same thing. The robot might say, "The patient is required to maintain a state of immobility regarding the lower extremity to facilitate osseous healing."

The Finding:

The Default AI is too fancy: When the researchers let the AI speak freely (without special instructions), the big, powerful models (like GPT-5 and Claude) sounded like they were writing a PhD thesis. They used huge words and long, complicated sentences.
The Result: The "Robot Voice" was harder to read than the human doctor's voice.
The Fix: When the researchers told the AI, "Hey, speak like you're talking to a 10-year-old," the AI suddenly became much clearer. It was like taking off a tuxedo and putting on a comfortable sweater.

2. The "Over-Emotional" Robot (Empathy)

The Analogy: Imagine a patient is crying because they are scared.

The Human Doctor: Might say, "I understand this is scary, but let's look at the facts. We have a plan." (This is called "detached concern"—caring, but professional).
The Default AI: Often swings to the other extreme. It might say, "Oh no! This is terrible! I am so, so sad for you!" It amplifies the emotion, sometimes making things sound more negative or more dramatic than a human would.
The Finding: The AI tends to be a "drama queen" or a "drama king." It either gets too negative or tries too hard to be positive, missing the calm, steady middle ground that doctors naturally hit.

3. The "Magic Editor" (Collaborative Rewriting)

The Analogy: This is the most important part of the study.
Imagine a human doctor writes a perfect, accurate medical note. But it's a little dry and hard to read.
Now, imagine you hand that note to a super-smart AI Editor. The AI doesn't change the facts (the medicine part). Instead, it acts like a translator. It takes the doctor's complex notes and rewrites them to be warm, simple, and easy to understand, without changing the truth.

The Finding:

AI as a Replacement? No. If you ask the AI to just "go make up an answer," it often gets the facts slightly wrong or sounds weird.
AI as a Helper? YES! When the AI was asked to rewrite a doctor's existing answer, it was a superstar.
- It kept the medical facts 100% accurate (just like the doctor).
- It made the language much simpler (like a translator).
- It made the tone warmer and less scary (like a good editor).

4. Who Likes What? (The Audience Test)

The researchers asked two groups to rate the answers: Medical Experts and Regular Patients.

The Experts (Doctors): They cared about Accuracy. They gave the human doctors a perfect score. They gave the AI a slightly lower score because the AI sometimes missed tiny details or sounded too "robotic" in its precision. Verdict: The AI cannot replace the doctor's brain.
The Patients: They cared about Clarity and Kindness. They loved the "Rewritten" versions. They said, "I understand this better, and I feel more supported." Verdict: The AI is great at being the doctor's "voice" to the patient.

The Big Conclusion: The "Co-Pilot" Model

The paper concludes that AI shouldn't try to be the Captain of the ship (the doctor). It doesn't have the experience or the deep medical judgment to make the hard calls.

Instead, AI is the perfect Co-Pilot or Translator.

The Doctor is the expert who knows the medical facts.
The AI is the tool that takes those facts and packages them into a message that is easy to read, easy to understand, and emotionally supportive.

In short: Don't ask the AI to be the doctor. Ask the AI to help the doctor speak better to the patient. When used that way, it's a game-changer.

1. Problem Statement

While Large Language Models (LLMs) are increasingly deployed in healthcare to assist with patient messaging, their communicative alignment with clinical standards remains insufficiently quantified. Existing research focuses heavily on factual accuracy (hallucinations) but neglects three critical dimensions of clinical communication:

Semantic Fidelity: Does the AI preserve the precise medical meaning of expert responses?
Readability: Is the language accessible to non-expert patients (varying literacy levels)?
Affective Resonance: Does the AI's emotional tone (empathy, sentiment) align with the "detached concern" required in clinical settings, or does it amplify extremes (e.g., excessive negativity or false reassurance)?

The study addresses the gap in understanding whether LLMs can replicate or improve upon human physician communication styles and whether they function better as autonomous agents or collaborative tools.

2. Methodology

Datasets

The study utilized two distinct datasets to evaluate performance across different communication contexts:

MedQuAD: A large-scale corpus of 47,457 question-answer pairs from authoritative NIH sources (e.g., MedlinePlus). A readability-stratified subset of 50 questions was selected to ensure linguistic diversity while controlling for outliers. This represents institutionally curated medical explanations.
iCliniqQAs: A subset of 465 real-world physician-patient interactions from an online consultation platform. A severity-balanced subset of 50 questions (stratified by clinical urgency) was used. This represents authentic, conversational clinical dialogue.

Models Evaluated

The study compared general-purpose and domain-specialized models:

General-Purpose: Mixtral, GPT-5, Gemini 2.5 Pro, Claude Sonnet 4.5.
Domain-Specialized: Med-PaLM 2.

Experimental Conditions (Prompting Strategies)

Three distinct configurations were tested for each model:

Base Prompt: Standard clinical baseline (neutral, factual generation).
Empathy Prompt: Explicitly instructed to enhance emotional awareness and patient-centered tone.
Rephrase Prompt: A collaborative editing mode where the LLM rewrites an existing physician-authored answer to improve clarity and tone while preserving meaning.

Evaluation Metrics

Semantic Fidelity: Measured via Cosine Similarity of sentence embeddings (using BioBERT-mnli-snli-scinli-scitail-mednli-stsb) between AI outputs and physician references.
Readability: Quantified using Flesch-Kincaid Grade Level (FKGL) and Gunning Fog Index (GFI). Lower scores indicate better accessibility.
Affective Resonance: Analyzed via Sentiment Classification (5 classes: Very Negative to Very Positive) and Fine-grained Emotion Detection (28 emotions, e.g., caring, approval, disapproval).
Human Evaluation: Dual stakeholder assessment by medical experts (epistemic criteria: accuracy, precision, style) and patients (relational criteria: trust, comprehensibility, emotional tone) using 5-point Likert scales.

3. Key Contributions

Multidimensional Framework: Proposes a unified evaluation framework that jointly assesses semantic, linguistic, and affective dimensions, moving beyond isolated accuracy checks.
Real-World vs. Institutional Comparison: Contrasts AI performance in standardized institutional texts (MedQuAD) against real-world conversational dynamics (iCliniqQAs), revealing context-dependent behaviors.
Collaborative vs. Autonomous: Systematically evaluates LLMs not just as answer generators but as collaborative editors (rewriting physician text), a use case often overlooked in favor of zero-shot generation.
Empirical Evidence on Prompting: Provides quantitative data on how "Empathy Prompting" and "Rephrasing" alter model behavior regarding readability and sentiment.

4. Key Results

Semantic Fidelity

Baseline: All models showed strong conceptual alignment (Cosine Similarity > 0.75) with physician answers.
Rewriting Superiority: Collaborative rewriting consistently yielded the highest semantic fidelity.
- MedQuAD: GPT-5 Rephrase achieved $\mu = 0.92$ .
- iCliniqQAs: Med-PaLM Rephrase achieved $\mu = 0.93$ .
Empathy Prompting: Did not significantly increase semantic fidelity compared to baseline; it primarily affected style and tone.

Readability

Baseline Complexity: General-purpose models (GPT-5, Claude) produced significantly more complex text than physicians (e.g., GPT-5 Base FKGL $\approx$ 16.9 vs. Physician $\approx$ 11.5). Domain models (Mixtral, Med-PaLM) were closer to physician levels.
Impact of Strategies:
- Empathy Prompting: Reduced complexity significantly (e.g., GPT-5 reduced FKGL by ~6.87 points).
- Rewriting: Produced the most consistent readability gains, lowering FKGL and GFI scores across all architectures to match or exceed physician readability.

Affective Resonance (Empathy & Sentiment)

Polarity Amplification: Baseline LLMs tended to amplify affective polarity. In MedQuAD, they increased "Very Negative" sentiment (43–45%) compared to physicians (37%).
Correction via Strategies: Both Empathy Prompting and Rewriting successfully reduced extreme negativity and increased "Neutral" sentiment.
Emotion Shift:
- Institutional (MedQuAD): Physicians were "Approval" dominant; LLMs shifted toward "Caring" (amplified by prompting/rewriting).
- Conversational (iCliniqQAs): Physicians were already "Caring" dominant; LLMs pushed this toward near-saturation (e.g., 92% "Caring" in some rephrased outputs).
Gemini/Claude Behavior: These models exhibited safety-driven hedging (evasive responses) when context was insufficient, leading to classifier-detectable "Very Negative" sentiment due to uncertainty rather than clinical content.

Stakeholder Alignment (Human Evaluation)

Expert (Physician) View: No LLM configuration surpassed physicians on epistemic criteria (Accuracy, Precision, Style). Physicians remained the gold standard for medical correctness.
Patient View: Patients consistently preferred rewritten variants over raw physician answers and baseline LLM outputs. Rewriting significantly improved scores for Trust, Comprehensibility, and Emotional Tone.
Conclusion: LLMs excel at relational alignment (clarity/empathy) but fail to achieve epistemic superiority.

5. Significance and Conclusion

Main Finding: LLMs function most effectively as collaborative communication enhancers rather than autonomous replacements for clinical expertise.

The "Rewriting" Paradigm: The study demonstrates that using LLMs to rewrite physician-authored content yields the best balance: it preserves semantic fidelity (medical accuracy) while significantly improving readability and emotional resonance.
Prompting Limitations: While "Empathy Prompting" improves tone and readability, it does not guarantee better semantic alignment and can sometimes introduce stylistic shifts that deviate from clinical norms (e.g., over-amplifying "caring" inappropriately).
Context Matters: AI behavior varies drastically between formal medical explanations and conversational consultations. Models tend to over-amplify emotions in conversational settings where physicians are already restrained.

Implications for Healthcare AI:

Human-in-the-Loop: Clinical AI should be deployed as a "co-author" or editor to refine human-generated content, not as a standalone generator of medical advice.
Safety & Trust: While AI can improve patient trust through better tone and clarity, it cannot replace the epistemic authority of a physician.
Future Directions: Future work should focus on multi-turn interactions, domain-specific affective models, and evaluating performance across varying levels of clinical criticality (severity).

The paper concludes that while AI can enhance the delivery of medical information, it cannot yet replicate the judgment and nuance of a human doctor. The optimal deployment strategy is Human-AI co-authorship.