Assessing demographic variation in large language model outputs for patient education materials in cataract surgery
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
• Demographic factors influence LLM readability and word count in cataract education • Indigenous and Black prompts elicited the most complex LLM-generated responses • Insured patients consistently received longer, more complex content across models • Geographic disparities across Canadian provinces were observed, with Nunavut receiving the least readable outputs. • No LLM met AMA readability standards; all exceeded sixth grade reading level To evaluate whether large language model (LLM)-generated patient education materials for cataract surgery vary in readability, length, and accuracy based on demographic modifiers including race, gender, geography, and insurance status. Cross-sectional study This study analyzed 7,000 responses from five LLMs (ChatGPT, Claude, Copilot, DeepSeek, and Gemini) between March-May 2025 using 280 standardized prompts that varied by race, gender, province/territory, and insurance coverage. Each prompt was submitted five times. Readability was assessed using the Flesch-Kincaid Grade Level (FKGL), Flesch Reading Ease (FRE), and SMOG index. Accuracy was assessed by dual blinded reviewers against AAO clinical guidelines. ANOVA was performed (α = 0.05). LLM outputs differed significantly across all metrics (p < 0.001). Gemini generated the longest (876 ± 143 words) and most complex text (FKGL 15.2 ± 0.8). Race, insurance status, and geography significantly impacted readability. Prompts referencing Indigenous patients were the most complex (FKGL 11.1 ± 1.8, FRE 36.5 ± 7.9). Insured prompts were longer and more complex (11.0 ± 1.7 vs. 10.8 ± 1.7; 429 vs. 399 words; p < 0.001). Prompts from Nunavut and Manitoba were the least readable (FKGL ≥ 11.1), while Quebec and PEI were most readable. Gender had minimal impact. No outputs contained clinically unsafe information, but most lacked sufficient depth. None of the responses met the AMA’s sixth-grade readability recommendation. LLM-generated patient education for cataract surgery varies by patient demographics. These disparities may hinder equitable access to health information and highlight the need for bias-aware development of AI tools in healthcare.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.002 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.001 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.002 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it