The Readability Accuracy Trade-off in Medical LLMs

This direction is merged into another direction: health literacy gap. Its last version stays readable and citable here.

Version 1 of 118 papers32 conceptsfrozen 2026-10-04

Large Language Models (LLMs) have demonstrated remarkable proficiency in passing medical licensing exams and synthesizing complex clinical data. However, a systemic gap has emerged between their technical "intelligence" and their ability to communicate with the people who need them most: patients. Research across dozens of medical specialties reveals that while LLMs are becoming more accurate, they consistently fail to meet the health literacy standards required for safe and equitable patient care.

The Health Literacy Gap

The gold standard for patient-facing medical information is a 6th-to-8th-grade reading level. This ensures that information is accessible to the general public, including vulnerable populations and those under the cognitive stress of illness.

Current frontier models systematically fail this benchmark. As Felix Busch and Lena Hoffmann note in Systematic Review of Large Language Models for Patient Care: Current Applications and Challenges, AI-generated content frequently exceeds the comprehension levels of the general public. This isn't just a minor stylistic preference; it is a barrier to entry. When a model explains a clavicle fracture or an Achilles tendon rupture, it often defaults to academic or collegiate language.

In studies led by Peter A Giammanco, researchers found that even when models like Microsoft Copilot or Gemini provide clinically sound advice, the technical complexity of the output alienates the intended audience. The intuition here is that LLMs are trained on vast corpora of scientific literature and clinical notes, leading them to mirror the "expert" voice of a physician rather than the "educator" voice required for a patient.

The Zero-Sum Game of Simplification

The most significant challenge in this field is what Weikang Qiu and colleagues call the readability-accuracy trade-off. In the RephQA benchmark, researchers observed that as models are pushed to simplify their language to meet lower grade levels, they often begin to lose critical medical facts.

This creates a dangerous paradox in medical AI:

  1. High Accuracy/Low Readability: The model provides a perfect, nuanced explanation of a drug's side effects, but uses jargon that a patient in distress cannot parse.
  2. Low Accuracy/High Readability: The model uses simple, comforting language but omits a crucial contraindication or simplifies a risk until it is no longer factually representative.

Shiqi Qiang highlights this conflict in stroke rehabilitation, noting that highly comprehensive responses often require a high reading level. For a patient with cognitive impairments following a stroke, an "accurate" but complex explanation is effectively useless—or worse, dangerous if misunderstood.

Performance Degradation in High-Stakes Contexts

The readability problem is compounded by the emotional and situational context of the patient. In the DischargeSim benchmark, Zonghai Yao and Michael Sun found that AI performance degrades significantly when interacting with patients characterized by high anxiety or low health literacy.

This suggests that current models are "fair-weather" communicators. They can maintain logic in a vacuum, but they struggle to adapt their linguistic complexity to the specific needs of a human user in a crisis. This is particularly evident in mental health and psychiatric medication management. Research by Mohit Chandra and Siddharth Sriraman shows that LLMs fail to align with expert clinicians on practical harm-reduction strategies, often choosing "risk-averse" jargon over the accessible, actionable advice a human expert would provide.

Moving Beyond "Reasoning"

The prevailing trend in AI research has been to optimize for "reasoning" (e.g., models like o1 or DeepSeek-R1). However, the evidence from clinical text generation tasks like RRG24 and "Discharge Me!" suggests that reasoning is not enough.

True clinical utility requires linguistic empathy—the ability to translate complex medical reasoning into a format that is both factually robust and cognitively accessible. As Josephine Lamp argues in the DM-Bench study on diabetes management, effective health AI must balance technical accuracy with low-literacy-friendly communication. Until models can resolve the trade-off between the "what" (accuracy) and the "how" (readability), their role in direct patient care will remain limited by the very sophistication that makes them impressive.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_the_readability_accuracy_trade_off_in_me_f73bb44b,
  title = {The Readability Accuracy Trade-off in Medical LLMs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-readability-accuracy-trade-off-in-medical-llms/rdr_f73bb44bc37d488eb936270eec02acb5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_the_readability_accuracy_trade_off_in_me_f73bb44b_v1,
  title = {The Readability Accuracy Trade-off in Medical LLMs},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/the-readability-accuracy-trade-off-in-medical-llms/rdr_f73bb44bc37d488eb936270eec02acb5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}