Can Large Language Models Address Problem Gambling? Expert Insights from Gambling Treatment Professionals

Comprehensive Summary

This study examines how general-purpose large language models (LLMs) respond to problem gambling-related prompts and how experienced gambling treatment professionals assess those responses. After modifying nine first-person prompts from the Problem Gambling Severity Index and submitting them to two widely used models, GPT-4o and Llama 3.1, the authors polled 23 gambling treatment professionals, who gave their own answers, chose their preferred model outputs, and provided qualitative feedback. Although preferences differed between prompts and clinicians, experts generally preferred Llama responses over GPT. Both models generated noticeably longer, more complex, and less readable responses than human experts. Qualitative analysis revealed that although some responses used stigma-free language and a supportive tone, others contained problematic phrasing, excessive verbosity, or content that might be misconstrued as promoting continued gambling. These findings suggest that current general-purpose models do not yet meet professional standards for gambling support, as the majority of experts stated they would not change their own answers after seeing the model outputs.

Outcomes and Implications

Problem gambling remains highly stigmatized and is associated with low rates of help-seeking despite significant psychological, social, and financial harm. As general-purpose AI tools become more commonly used for advice on sensitive topics, understanding their limitations in addiction-related contexts is increasingly important. Clinically, these findings suggest that current large language models should not be used as standalone tools for gambling support. Instead, they may have a limited role as adjunctive resources that encourage reflection and direct individuals toward professional care, though further alignment and validation are required before broader clinical application.