As GPT-powered chatbots become a go-to tool for quick answers and emotional support, a growing number of problem gamblers are turning to artificial intelligence (AI) instead of therapists or helplines. But a new study from AiR Hub warns that these large language models (LLMs) may be providing misleading, and potentially harmful advice, on gambling addiction.
Why chatbots and gambling?
The research, led by Kasra Ghaharian, Ph.D., Director of Research at UNLVs International Gaming Institute and Co-Founder of AiR Hub, tested how two of the most prominent LLMsGPT-4o (ChatGPT) and Metas Llamarespond to questions based on the Problem Gambling Severity Index (PGSI). The was to evaluate whether these models offer advice that supports responsible gambling or inadvertently encourages risky behaviour. People are using LLM-based chatbots to ask questions about gambling, we thought it would be interesting to empirically investigate how they respond, specifically to questions related to problem gambling,” Ghaharian told SiGMA News.
The AI-generated responses were reviewed by 23 seasoned gambling-treatment professionalswith over 17,000 hours of combined experience. The feedback highlighted both models inability to consistently provide clear, responsible guidance, especially when the queries hinted at problematic gambling behaviour. The study finds that, “Some responses subtly encouraged continued gambling, buried practical advice within overly long explanations, or used jargon that experts felt could be easily misunderstood.“
The AI alignment problem
AI alignmentthe process of ensuring systems behave in line with human values and safety standardsis critical in sensitive domains. But gambling presents a unique challenge. Chatbots must not only provide reliable responses to casual betting queries but also detect when a user may be at risk of harm. Ghaharian said casual questions can be effectively answered by the assistant. Brian Pempus, founder of GamblingHarm.org, echoes this thought. He believes that general-purpose AI chatbots like ChatGPT are technically capable of recognising problem gambling signsbut only to a point.
ChatGPT appears equipped to recognise the defining signs of problem gambling established by the DSM-5, he explains. But whether it can recognise signs unique to the individual depends on how deeply and thoroughly they describe what is happening related to their gambling.
Complexities: AI’s foe?
However, the chatbot struggles when complexity rises. Ghaharian said, “If questions became more complex or showed signs of potential harm, there might not be any underlying/reliable source data or guardrails in place for the assistant to craft an adequate response or maybe even understand whether they should or should not answer it in the first place.”
This thought is seconded by Keith Scott Whyte, Founder & President of Safer Gambling Strategies LLC, who believes problem gambling is incredibly complex. Speaking to SiGMA News, he said, “Problem gambling is still heavily stigmatised and in some ways it’s still poorly understood. So, it’s not surprising that chat bots would have a hard time really picking up on some of these nuances and language.” He advices:
“From a player safety perspective, chatbots should almost assume a user has a problem until proven otherwise. The concept of ‘first do no harm’ applies here more than ever.
This becomes especially pressing as failure to properly distinguish between someone who has a gambling problem and someone who does not can be “life threatening,” Whyte warned. This isnt like getting your takeout order wrong. The consequences of failure here are catastrophic. In a candid conversation with SiGMA, Pempus highlighted one of the gravest concerns is how chatbots assess suicide risk, especially in the context of gambling addiction. Gambling addiction involves a heightened risk of suicidal ideation, a sign of a severe gambling problem, Pempus notes.
Research has shown that ChatGPT consistently rated suicide risk lower than human experts.
This disconnect, he says, suggests that chatbots are inadequate when it comes to recognising and responding to the most dangerous and urgent signs of gambling-related harm. Pempus also flags a subtler danger: the micro-messages that LLMs can send when replying to users unsure about their gambling behaviour. Some people can receive dangerous suggestions to continue betting in situations where a human would help them explore whether abstaining from gambling is the better idea, he warns.
Operators using AI for assistance, is it helping?
It becomes all the more important as gambling companies increasingly deploy conversational AI in customer service and betting assistance roles. To build safer and more aligned AI systems in gambling contexts, incorporating lived experience is non-negotiable, says Whyte. He also flagged the challenge of detecting harm from dialogue alone.
Gambling addiction is so much more complex than trying to solve chess. This might be the ultimate test of an AI systemto try and distinguish recreational gambling behaviour from at-risk behaviour purely based on chat.
Pempus has a different lens. While AI may have a place in customer support, Pempus draws a clear ethical line: Operators should not be the ones providing responsible gambling support via AI. Operators should and do offer some tools to help users potentially control their play, he said. However, they should not use AI to become directly involved in providing support or guidance to people dealing with problem gambling.
Meanwhile, Michael Pollock, senior policy advisor at Spectrum Gaming Group told SiGMA News, “Appropriate testing, monitoring, and licensure standards are required that would prohibit operators from engaging in AI-related efforts that are not in line with a firm and full commitment to responsible gaming.“
AI should be fed first-party data: expert
To improve accuracy and safety, Whyte suggested integrating first-party datasuch as transaction history or recent lossesinto AI models. Same language, same words, can take on a completely different tone depending on whether the user just lost 500,000 or just won big and is heading on holiday, he explained. Context is everything.
Pollock said, “The continued evolution of AI must be matched by appropriate regulatory oversight. That means that regulators need to come up to speed and stay abreast of all developments.“
Why the AiR Hub study, and why now?
The AiR Hub teams interest in examining AI-generated gambling advice stemmed from a noticeable shift in how people access information. Information retrieval has fundamentally changed, Ghaharian told SiGMA News. With chatbots, we retrieve an instantaneous answer, and we no longer have to sift through a list of search engine results. This shift carries serious implications in high-risk areas such as gambling.
The team crafted nine prompts, each corresponding to a PGSI item, and submitted them to two leading LLMs: GPT-4o and Llama, chosen for being the most popular proprietary and open-source models at the time of the study. Rather than relying on rigid scores, the researchers opted for open-ended qualitative evaluations from 23 seasoned gambling counselors with over 17,000 combined hours of treatment experience. We didnt want to impose a rigid scoring system that might overlook nuances or force consensus where none existed, Ghaharian explained. Instead, the counselors were asked to elaborate on whether they found the AI responses helpful, misleading, or potentially harmful.
While neither model convincingly outperformed the other, the expert reviewers found that both had a tendency to generate responses that could be misinterpretedor worse, subtly encourage continued gambling. Probably the biggest red flag was that both models had the tendency to provide advice that could be misconstrued, Ghaharian noted. Thats especially concerning when users may be in a vulnerable state.
Although the pilot study revealed concerning trends, it also laid the groundwork for solutions. Ghaharian said, “The study lays important groundwork by highlighting the potential risks and limitations of LLMs in gambling-related contexts. But its really just the beginning and theres much more we can do.” AiR Hub and UNLV researchers are now working on a benchmark dataset of gambling-related prompts and ideal responsesintended for use by both AI developers and gambling platforms. The benchmark could also serve a regulatory function. A regulator could require that any customer-facing LLM deployed by a licenced operator be evaluated against the benchmark, Ghaharian proposed. That would help ensure it can both recognise potentially harmful prompts and generate appropriate, responsible responses.
