Challenges of AI-generated stigmatizing language regarding substance use disorders.

Abstract

Stigmatizing language describing substance use behaviors in clinical documentation and in patient education materials can harm patients and their families. Recent literature has discouraged the use of stigmatizing language in treatment settings and medical documentation. However, large language models (LLMs) generate output using previously generated text, including sources such as electronic health records which may include stigmatizing language. While artificial intelligence (AI) developers continue to update and create LLM products, these developers might prioritize other issues over reducing stigmatizing language surrounding substance use disorders. These diverging priorities make it possible that updated and new LLMs might not eliminate the problem of stigmatizing language in LLM output, even as researchers explore methods to reduce the impact of stigmatizing language in AI models. The potential for LLM output to include stigmatizing language warrants caution on the part of clinicians as AI developers release products intended to facilitate clinical documentation. This article identifies stigmatizing language regarding substance use disorders and explores the potential for LLM output to include stigmatizing language. The author also discusses several challenges that clinicians may face in preventing stigmatizing language from transferring from LLM output into medical records, including randomness inherent in LLM output, balancing appropriate versus inappropriate uses of potentially stigmatizing language, and the need for continued human review of medical documentation even in the face of potentially increased patient volumes. Finally, the author proposes that practicing clinicians be included in an ongoing, active process of iterative refinement of LLM output for clinical use to reduce the risk of perpetuating stigmatizing language regarding substance use disorders.