Chatbots Are Already Giving Gambling Advice
Hundreds of millions of people use ChatGPT-style tools every week. Inevitably, many likely ask about gambling - everything from “What’s a good roulette strategy?” to “How do I stop chasing my losses?”
Yet today’s large language models (LLMs) were not trained with gambling conversations in mind. In a domain where misguided advice can have serious negative consequences - including financial loss, psychological distress, or addiction - even small errors or mixed messages carry significant risk.
AI Alignment - What It Is and Why It Matters
Because chatbots now field questions on many sensitive topics, from mental-health crises to medical and legal advice, researchers have carved out an entire discipline focused on making sure these systems behave responsibly. Alignment is the process of steering an AI system so its outputs reflect a desired set of facts, values, and/or safety goals. In practice, this involves: establishing explicit principles the model must honor; fine-tuning the model on carefully curated examples that embody those principles; and evaluating and correcting the system whenever it deviates.
Some real-world examples of alignment efforts include Anthropic’s Constitutional AI, which trains its Claude model to follow a human-crafted “constitution”, and OpenAI’s HealthBench, which provides a physician-rated benchmark that aims to keep medical advice accurate and appropriate.
Do We Need Alignment in Gambling? Our New Study Suggests We Might
To gauge the need for alignment of LLMs for gambling conversations we conducted an exploratory study. We asked two leading models (GPT-4o and Llama) nine questions based on the Problem Gambling Severity Index. Their responses were then blind-rated by 23 experienced gambling-treatment professionals, who collectively bring over 17,000 hours of counselling expertise.
We found several causes for concern. Some responses subtly encouraged continued gambling, buried practical advice within overly long explanations, or used jargon that experts felt could be easily misunderstood. Counselors also noted that lengthy replies were often confusing and failed to directly address the core issue raised by the question.
The study was accepted to the AIR-RES 2025 Conference, with proceedings to be published soon, and is currently under review at a peer-reviewed journal. A preprint of the full paper is available.
Where We Go Next
Gambling creates a uniquely tricky alignment challenge: chatbots must handle casual questions, such as lighthearted betting tips, but also recognize signs that someone might be edging into harmful territory.
AiR Hub is now sketching out follow up work. We plan to scale up this piece of research, bringing in more experts and collecting a much larger set of question-response pairs, to create a dataset that LLM developers and the wider industry can use. For example, this dataset could be used to ensure customer-facing chatbots can tell the difference between a routine betting query vs. a potentially concerning one, as well as answer in a way that keeps player safety front and centre.
Stay tuned - there’s more to come, but the goal is simple: make sure the next wave of AI tools in gambling actively supports consumer protection rather than leaving it to chance.