A longitudinal analysis of declining medical safety messaging in generative AI models

Sonali Sharma, Ahmed M. Alaa, Roxana DaneshjouNpj Digit. Medicine 2025 · DOI

Abstract

As generative AI models achieve higher accuracy in medical tasks, they are paradoxically becoming less likely to provide safety disclaimers. This longitudinal study evaluates Large Language Models (LLMs) and Vision-Language Models (VLMs) from 2022 to 2025, revealing a dramatic collapse in medical safety messaging. Disclaimer rates in LLMs plummeted from 26.3% to less than 1%, while VLMs saw a similar decline. This trend creates a "silent" risk for patients who may over-rely on AI for health information without being reminded of the models' limitations.

Introduction

Patients increasingly utilize AI—both text-based (LLMs) and vision-based (VLMs)—to interpret symptoms, lab results, and medical images. While these models are becoming more capable, they are not regulated medical devices and are prone to "hallucinations." Medical disclaimers (e.g., "I am an AI, not a doctor") serve as a critical safeguard to prevent misplaced trust and delayed professional treatment. This paper investigates whether these safeguards are keeping pace with model development or if they are being phased out to prioritize conversational fluency and user experience.

Methods

The researchers conducted a systematic, longitudinal evaluation of model outputs from 2022 to 2025 using standardized, single-turn API prompts to ensure comparability.

Datasets

The study employed the Top Internet Medical Question Dataset (TIMed-Q), consisting of 500 frequently searched medical queries across domains like symptom management, emergencies, and medication safety. Additionally, 1,500 medical images (mammograms, chest X-rays, and dermatology images) were used to test VLM interpretation capabilities.

Evaluation and Analysis

Disclaimers were identified using a combination of regular expressions (RegEx) and manual review. The relationship between diagnostic accuracy and disclaimer presence was examined using Pearson correlation, while linear regression tracked trends over time.

Results

The "Vanishing Safeguard"

The most striking finding is the near-total disappearance of medical disclaimers across the industry. As shown in Figure 5, disclaimer rates for both LLMs and VLMs have converged toward zero by 2025.

Overall distribution of medical disclaimers from 2022 to 2025 for LLMs and VLMs.
  • LLM Decline: Disclaimer rates for medical questions fell from 26.3% (2022) to just 0.97% (2025).
  • VLM Decline: For medical image interpretation, rates dropped from 19.6% (2023) to 1.05% (2025).

The Accuracy-Safety Paradox

A significant negative correlation ($r = -0.64$) was found between diagnostic accuracy and disclaimer presence. As models become more "intelligent" and accurate, they provide fewer warnings. This is concerning because even highly accurate models still fail in a small percentage of cases, where a disclaimer is most vital to prevent clinical error.

Corporate Philosophy and Model Specifics

Safety behavior varied significantly by developer. Google’s Gemini models maintained the highest disclaimer rates (~41-49%), suggesting a deliberate "safety-first" design. Conversely, DeepSeek models consistently provided zero disclaimers. OpenAI models showed a sharp decline over time, as illustrated in Figure 3.

Longitudinal decline of disclaimer rates in OpenAI models, with GPT-4.5 reaching 0% disclaimers.

Discussion

The decline in disclaimers may be an unintended consequence of Reinforcement Learning from Human Feedback (RLHF), which often rewards shorter, more direct answers. Alternatively, developers may be intentionally reducing "user friction" by removing repetitive warnings to improve conversational flow.

Limitations

This study focused on single-turn API interactions. Real-world safety behavior might differ in multi-turn web interface conversations where system prompts or UI-level warnings are present. Future work should investigate if models "forget" their initial caution as a conversation progresses.

Conclusion

The rapid erosion of medical safety messaging in flagship AI models presents a growing public health risk. The authors recommend that medical disclaimers be implemented as a non-optional, context-aware safety feature that adapts to the clinical severity of the user's inquiry.

Summary updated

BibTeX
@article{sharma2025longitudinal,
  author = {Sonali Sharma and Ahmed M. Alaa and Roxana Daneshjou},
  title = {{A longitudinal analysis of declining medical safety messaging in generative AI models}},
  journal = {Npj Digit. Medicine},
  year = {2025},
  doi = {10.1038/s41746-025-01943-1},
}
Figure