Artificial intelligence (AI) is being adopted more widely within online gambling platforms, extending beyond marketing and fraud prevention to areas such as monitoring player activity and assessing behavioural risk. Many operators rely on AI to analyse real-time player data, assessing behaviour such as how often individuals bet, how much they stake and how long they play. These analytical capabilities are also being used within responsible gambling measures to help identify risk indicators at an earlier stage. In regulated markets, responsible gambling approaches increasingly combine continuous, technology-supported monitoring with existing compliance requirements.
Despite these advances, questions remain about how effectively AI can balance commercial goals with player protection. To understand these issues, SiGMA News spoke exclusively to Seungpil Lee, an AI researcher specialising in autonomy, alignment and risk behaviour. Lee explains how AI systems can identify behavioural risk indicators, why the design of objective functions is critical, and the difficulties in aligning reinforcement learning models with player protection objectives. In this interview, he examines current capabilities, outlines limitations in existing AI-driven responsible gambling tools, and explains the role regulators can play in establishing transparency, accountability and oversight.
SiGMA News: How do you see AI being used today to detect early signs of risky or harmful gambling behaviour?
Seungpil Lee, AI Researcher: From a technical standpoint, detecting early warning signs through behavioural pattern data, such as changes in betting frequency, wager amounts and session duration, is certainly feasible. Natural language processing can be used to identify emotional escalation in customer interactions, while time-series analysis can detect patterns associated with loss chasing. These capabilities are well within current technical limits.
SiGMA News: Can AI systems be designed to balance player personalisation with responsible gambling safeguards?
Lee: Technically, yes, but the key issue is how the objective function is defined. If a personalisation algorithm is optimised to ‘maximise engagement time’, responsible gambling safeguards are likely to become superficial. In contrast, if long-term player wellbeing indicators are incorporated directly into the objective function, personalisation can operate as a protective mechanism rather than an exploitative one. For example, individual risk thresholds could be established, with automated interventions activated once those thresholds are exceeded. The difficulty is that, in the current industry landscape, economic incentives to implement such designs remain limited.
SiGMA News: What are the ethical risks when AI is used to optimise engagement in gambling environments?
Lee: The most significant risk arises from the combination of information asymmetry and AI’s sycophantic tendencies. A recent Anthropic study titled “Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage” (Sharma et al., 2026) analysed 1.5 million real-world Claude.ai conversations and identified several concerning patterns. These included AI assistants validating persecution narratives and grandiose self-perceptions through emphatic and affirming language, issuing definitive moral judgements about third parties, and drafting value-laden personal communications that users appeared to adopt verbatim.
Importantly, interactions with greater potential for disempowerment received higher user approval ratings. This indicates a structural tension between short-term user preferences and long-term human empowerment.
Applied to gambling, this suggests that engagement-optimised AI systems may reinforce rather than correct players’ irrational beliefs, such as “I am due for a win” or “I can read the pattern”. If systems that provide affirming responses consistently receive higher satisfaction scores, optimisation processes may favour the validation of such beliefs. This is not merely a matter of poor design; it is a structural risk inherent in any AI system that uses user satisfaction as its objective.
A recent Anthropic study titled “Who’s in Charge? Disempowerment Patterns in Real-World LLM Usage”(Sharma et al., 2026) analysed 1.5 million real-world Claude.ai conversations and identified several concerning patterns.
SiGMA News: In your view, can reinforcement learning systems be aligned to prioritise player protection rather than revenue optimisation?
Lee: In theory, yes, by redesigning the reward function. However, the finding by Sharma et al. (2026) that disempowering responses receive higher user approval highlights a fundamental difficulty. When reinforcement learning systems rely on user feedback signals such as satisfaction ratings, session duration and return frequency as reward signals, the objective of ‘player protection’ can gradually shift towards providing responses that align with user preferences and ‘telling players what they want to hear’ rather than addressing risk.
Effective alignment, therefore, requires the development of protection metrics that are independent of user satisfaction. Instead of relying on single-objective optimisation, a multi-objective structure could be adopted, for example, pursuing revenue while ensuring that player risk indicators remain within defined thresholds. Such approaches must also consider the possibility that the reward signal itself may be biased.
SiGMA News: What limitations or blind spots currently exist in AI-driven responsible gambling tools?
Lee: I would highlight three main limitations. First, most models rely on single-platform data and therefore cannot account for cross-site gambling behaviour. Second, biases in training data may result in the underdetection of problem gambling patterns among certain demographic groups. Third, and perhaps less widely recognised, is AI’s sycophantic tendency. As Sharma et al. (2026) showed, AI systems often affirm and reinforce users’ existing beliefs rather than examine them critically, and such affirming responses tend to receive higher satisfaction scores. If responsible gambling tools are built on foundation models that exhibit this tendency, a situation may arise in which systems appear to provide protection while effectively validating players’ distorted judgements.
SiGMA News: What role should regulators play in shaping how AI is deployed for player protection?
Lee: Regulators should fulfil three key functions. First, they should establish transparency requirements for AI systems, including clarity around objective functions and decision-making processes. Second, they should require independent third-party audits so that the assessment does not rely solely on operators’ internal evaluations. Third, they should clearly define accountability in cases where AI-based player protection measures fail. Given the pace of technological change, regulatory approaches should be principle-based rather than focused on prescribing specific technologies, in order to remain effective over time.
Stay ahead of iGaming’s biggest stories with SiGMA’s Top 10 News countdown. The world’s biggest iGaming community brings you weekly insights and subscriber-only offers. Join HERE today.
