Typos and Informal Language in Patient Communications Can Disrupt AI-Driven Medical Advice

Recent research conducted by MIT highlights a concerning challenge in deploying large language models (LLMs) for healthcare: nonclinical text variations, such as typos, slang, and formatting inconsistencies, can significantly influence the recommendations made by AI systems. The study reveals that these seemingly minor textual differences—like extra spaces, misspellings, or colloquial language—cause AI models to favor self-management advice over recommending in-person clinical care, even when hospitalization is necessary. Interestingly, certain language styles disproportionately affect recommendations for female patients, increasing the risk of misguidance and potential health disparities.

The findings underscore the critical need for thorough auditing of AI tools before their integration into healthcare services. As LLMs are already being used to draft clinical notes and triage patient messages, the team warns that unvetted models might produce unsafe or biased recommendations. The research involved modifying thousands of patient messages with realistic perturbations mimicking vulnerable groups, such as those with language barriers or health anxiety, and observing the models’ responses.

The results showed a consistent pattern: disruptions in message formatting and language led to a 7-9% rise in recommendations suggesting that patients manage their conditions at home, bypassing necessary medical intervention. Moreover, models demonstrated about 7% more errors for female patients, suggesting gender bias in AI reasoning. These errors are particularly troubling in conversational contexts, like health chatbots, where direct patient interaction is common.

Follow-up studies also indicate that human clinicians are less affected by such text variations, emphasizing that AI models are inherently fragile to these subtle text changes. As such, researchers advocate for the development of more robust model training and evaluation practices, incorporating realistic communication styles, to ensure safe and equitable AI healthcare applications.

This work, presented at the ACM Conference on Fairness, Accountability, and Transparency, prompts a reevaluation of current AI deployment strategies in medical contexts. Ensuring AI models accurately interpret patient messages regardless of stylistic differences is crucial to prevent misdiagnoses and treatment errors, ultimately fostering fairer and safer healthcare AI solutions.