A Systematic Analysis of Declining Medical Safety Messaging in Generative AI Models

Sonali Sharma, Ahmed M. Alaa, Roxana DaneshjouarXiv 2025 · DOI · arXiv

Abstract

Generative AI models are increasingly used for medical advice, yet their safety disclaimers are disappearing. This study analyzes model outputs from 2022 to 2025, finding that disclaimer rates in Large Language Models (LLMs) fell from over 26% to less than 1%. This "safety gap" poses risks as models become more fluent and accurate but less transparent about their non-professional status.

Introduction

Generative AI models (LLMs and VLMs) have become remarkably capable at interpreting medical data and providing health advice. However, they are not professionally vetted medical devices. As these models evolve to prioritize conversational fluency, the explicit safety disclaimers that once warned users of their limitations are being stripped away. This study addresses the erosion of safety messaging and the potential for misplaced user trust.

Methods and Datasets

The researchers conducted a longitudinal systematic analysis of model outputs from 2022 to 2025 across major model families (OpenAI, Google, Anthropic, xAI, and DeepSeek). The study utilized:

  • Medical Images: 1,500 images including mammograms, chest X-rays, and dermatology photos.
  • PRISM-Q: A novel dataset of 500 text-based patient queries covering symptom management, acute emergencies, medication safety, mental health, and lab results.

Disclaimer detection was performed using regular expression matching and manual review, focusing on explicit statements indicating the model is not a licensed professional.

Key Results

The primary finding is a dramatic decline in safety messaging across both text and vision modalities.

Overall distribution of medical disclaimers in LLMs and VLMs from 2022 to 2025.

As illustrated in the figure above, LLM disclaimer rates dropped from 26.3% in 2022 to 0.97% in 2025. VLM rates saw a similar decline, falling from 19.6% in 2023 to 1.05% in 2025. By 2025, flagship models like GPT-4.5 and Grok 3 produced no disclaimers at all for medical image queries.

Model-Specific Variations

While the overall trend is downward, different model families exhibited varying behaviors during the transition to newer versions.

Breakdown of disclaimer rates for Google AI, Anthropic, and DeepSeek model families.

As shown in the provider-specific charts, models like DeepSeek maintained a 0% disclaimer rate throughout the study, while others like Google's Gemini and Anthropic's Claude showed significant reductions in safety messaging as they transitioned to newer, more "helpful" versions.

Discussion: The Accuracy-Safety Paradox

A striking insight from the study is the negative correlation between diagnostic accuracy and disclaimer presence. As models become more accurate at identifying pathologies, they become less likely to warn users of their limitations. This suggests that as developers optimize for "helpfulness" and fluency, built-in safety triggers are being deprioritized. This is particularly dangerous in high-stakes areas like medication safety, where disclaimers were found to be nearly non-existent.

Limitations and Future Work

The study is limited by the "black box" nature of proprietary models, making it difficult to pinpoint the exact internal changes (e.g., RLHF) driving this decline. Future work should explore whether the loss of disclaimers correlates with model "overconfidence" and how user-specific memory features might further impact safety messaging.

Conclusion

The systematic decline of medical safety disclaimers in generative AI represents a growing public health risk. As models sound more authoritative, the absence of clear boundaries between AI advice and professional medical consultation could lead to dangerous delays in treatment.

Summary updated

BibTeX
@misc{sharma2025systematic,
  author = {Sonali Sharma and Ahmed M. Alaa and Roxana Daneshjou},
  title = {{A Systematic Analysis of Declining Medical Safety Messaging in Generative AI Models}},
  year = {2025},
  eprint = {2507.08030},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.08030},
}
Figure