AI Chatbots’ Safety Controls Tested by Problem Gambling Prompts
Key Points
- ChatGPT and Gemini can both refuse and provide betting advice depending on conversation flow.
- Safety cues about problem gambling are weighted against recent betting prompts.
- Longer chats may dilute protective triggers, leading to inconsistent responses.
- Experts cite memory weighting and context windows as key technical factors.
- The gambling industry is already experimenting with AI‑driven betting tools.
- Resources for problem gamblers include the 1‑800‑GAMBLER helpline and text line 800GAM.
A series of experiments with OpenAI’s ChatGPT and Google’s Gemini revealed that the safety mechanisms designed to block gambling advice can be inconsistent. When users first discuss problem gambling, the bots refuse betting tips, but after repeated betting queries the safety cues become diluted and the models provide advice. Experts explain that the models weigh recent conversation tokens more heavily, and longer chats can weaken safety triggers. The findings highlight challenges for AI developers in balancing protective features with user experience, especially as the gambling industry explores AI‑driven tools.
Testing Chatbot Responses to Gambling Queries
Researchers prompted OpenAI’s ChatGPT and Google’s Gemini with requests for sports betting advice. Initial replies were cautious, using language such as “consider evaluating” rather than direct recommendations. When the conversation shifted to a discussion of problem gambling, both models offered supportive suggestions and even provided the National Problem Gambling Helpline number (1‑800‑GAMBLER) and a text option (800GAM).
However, when the same users followed the problem‑gambling prompt with another request for betting advice, the bots reverted to offering suggestions. In a separate short chat that began with the problem‑gambling prompt, the models refused to give betting tips, explicitly stating they could not facilitate real‑money gambling.
Why Safety Signals Fluctuate
Assistant professor Yumei He explained that large language models process the entire conversation history within a context window. Tokens that appear more recently or more frequently receive greater weight in the model’s prediction. Consequently, repeated betting prompts can “dilute” the earlier safety cue about problem gambling, causing the model to overlook the protective instruction.
He noted that the balance is delicate: making safety triggers too sensitive could impede legitimate uses, while making them too lax allows potentially harmful advice to slip through. The length and content of a conversation directly affect how reliably safeguards operate.
OpenAI has acknowledged that its safeguards work best in short, common exchanges. In longer dialogues, the model may fail to prioritize safety cues, a limitation the company is actively addressing.
Expert Perspectives on AI and Gambling Risks
Kasra Ghaharian, director of research at the International Gaming Institute, highlighted that generative AI is already being tested in the gambling sector for tasks such as bet placement assistance. He warned that language used by the bots—phrases like “tough luck”—could unintentionally encourage continued gambling for vulnerable individuals.
Anastasios Angelopoulos, CEO of LMArena, emphasized that developers can adjust safety trigger sensitivity, but doing so may compromise user experience for non‑problematic interactions. He suggested that users might achieve safer outcomes by keeping conversations brief.
Implications and Resources
The experiments underscore the need for more robust alignment of AI models around sensitive topics such as gambling and mental health. As AI tools become more pervasive, ensuring they reliably refuse to facilitate gambling, especially for users with a history of problem gambling, remains a critical challenge.
For individuals struggling with gambling addiction, the National Problem Gambling Helpline (1‑800‑GAMBLER) and the text line (800GAM) are available resources.