A recent study from Mass General Brigham has examined how large language models (LLMs), a type of artificial intelligence used in healthcare communication, respond to questions about addiction and substance use. The research found that many AI-generated answers contain language that could be harmful or stigmatizing toward people dealing with alcohol or drug-related conditions. This raises concerns about how these systems might influence patient care and public perceptions.
The researchers tested 14 different language models by giving them prompts related to alcohol use disorder, alcohol-associated liver disease, and substance use disorder. Medical experts then reviewed the responses, looking for stigmatizing words or phrases based on official guidelines from major health organizations. Results showed that over one-third of the answers from models without special adjustments included language that could make patients feel judged or unwelcome.
This study also explored how modifying the input given to the AI models, known as prompt engineering, could change the outcome. By carefully crafting the instructions for the models, the researchers were able to reduce the use of stigmatizing language by nearly 90%. This suggests that these tools can be improved to communicate in a more supportive and respectful way.
The use of patient-centered language matters because it helps build trust between healthcare providers and patients. When people feel respected and understood, they are more likely to engage openly in their care. On the other hand, language that carries stigma may cause patients to hide information or avoid seeking help altogether.
Another finding showed that longer AI-generated answers were more likely to contain stigmatizing language. This pattern held true across all tested models, though some models performed better than others in avoiding harmful terms. The research highlights a clear need for ongoing improvements to these systems to ensure they promote positive communication.
The authors recommend that clinicians review any AI-generated content carefully before sharing it with patients. They suggest offering alternative wording that is more patient-friendly and free of stigma. This approach can help prevent unintentional harm and support better outcomes.
Future efforts should include people with personal experience of addiction in developing and refining the language used by AI tools. This will ensure that the technology reflects the needs and preferences of those most affected by substance use disorders. Involving patients and families can provide valuable insights into what kinds of words and expressions feel respectful and helpful.
As AI becomes a more common part of healthcare, attention to how these systems use language will be essential. This study shows that without careful oversight, large language models might spread harmful stereotypes even as they assist with information and advice. However, it also demonstrates that with the right guidance and design, these tools have the potential to support more compassionate communication.
By addressing these concerns, healthcare providers and AI developers can work together to create systems that help patients feel valued and understood. This could lead to improved trust and better engagement with treatment. As research continues, it will be important to balance the benefits of AI with careful consideration of its impact on language and stigma.
The study underscores the responsibility of those using and creating AI tools in medicine to prioritize respectful, non-stigmatizing communication. This is a key step toward building a healthcare environment that supports all patients, especially those facing addiction and related challenges.
Sources:
Study reveals stigmatizing responses in LLMs for addiction-related queries
Join the conversation!