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

Area of Science:

• Artificial Intelligence
• Behavioral Science
• Public Health

Background:

• Large Language Models (LLMs) are increasingly used for information retrieval, including sensitive topics like mental health and addiction.
• The application of LLMs to problem gambling advice requires careful evaluation due to potential risks.

Purpose of the Study:

• To investigate how Large Language Models (LLMs) respond to prompts concerning problem gambling.
• To assess expert interpretations of LLM-generated content for gambling harm interventions.

Main Methods:

• Nine prompts based on the Problem Gambling Severity Index were submitted to GPT-4o and Llama 3.1.
• Responses were evaluated by 23 experienced gambling treatment professionals via an online survey.
• Qualitative analysis of expert feedback and preferred LLM response selection was conducted.

Main Results:

• Llama 3.1 received a slight preference over GPT-4o, winning 7 out of 9 prompt evaluations.
• Experts identified LLM response strengths and weaknesses, including potential encouragement of gambling and misconstrued language.
• Concerns were raised about overly verbose messaging and safety guardrails.

Conclusions:

• LLM responses to problem gambling prompts require careful scrutiny by human experts.
• Future development of LLMs for gambling harm interventions must incorporate robust safety standards and expert-informed guardrails.
• Further research is needed to refine LLM outputs for sensitive health-related advice.