Challenges of AI-generated stigmatizing language regarding substance use disorders

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Public Health

Background:

• Stigmatizing language in clinical settings negatively impacts patients with substance use disorders.
• Existing literature advises against using stigmatizing language in medical documentation and patient education.
• Large language models (LLMs) trained on diverse datasets, including electronic health records, risk generating output with stigmatizing language.

Purpose of the Study:

• To identify stigmatizing language related to substance use disorders in LLM output.
• To explore the potential for LLMs to perpetuate such language in clinical documentation.
• To discuss challenges clinicians face in mitigating this risk and propose solutions.

Main Methods:

• Literature review on stigmatizing language in substance use disorder contexts.
• Analysis of the potential for LLMs to incorporate and propagate stigmatizing language.
• Discussion of practical challenges and proposed strategies for clinicians.

Main Results:

• LLMs may inadvertently generate stigmatizing language due to their training data, posing a risk in clinical documentation.
• Challenges include inherent randomness in LLM output, balancing language use, and the need for human oversight.
• Diverging priorities between AI developers and clinical needs may hinder the elimination of stigmatizing language.

Conclusions:

• Clinicians must exercise caution with LLM-generated clinical documentation due to the risk of perpetuating stigmatizing language.
• Involving practicing clinicians in the iterative refinement of LLM output is crucial to reduce the perpetuation of stigmatizing language.
• Continued human review of AI-assisted medical records is essential to ensure patient safety and ethical care.