Towards AI-Assisted Board Game-Based Learning: Assessing LLMs in Game Personalisation

This study evaluates whether general-purpose large language models can generate personalisation suggestions for board game-based learning comparable to those produced by human educators. Working within a structured instructional framework based on the ADDIE model, the authors tested ChatGPT (GPT-4 and GPT-4o), Microsoft Copilot, and Anthropic Claude against three expert educators. The task centred on two classroom activities drawn from BoardGameGeek, one for primary and one for secondary school, with all models receiving identical, standardised prompts covering persona, context, learning goals, scenarios, and game rules. The work is positioned as a follow-up to earlier pilot studies and addresses two questions: whether performance differs among AI tools, and whether general AI can match expert helpfulness.

In the pilot study, four AI tools and three experts each produced in-game and around-game personalisation suggestions, which were then blindly rated on constructive alignment, feasibility, and inclusivity. Evaluators identified the source of suggestions at near-chance rates of 38 to 45 percent. No statistically significant difference emerged among the AI tools (Kruskal-Wallis p=0.64). AI suggestions scored significantly higher than human suggestions on constructive alignment for in-game personalisation (p=0.004), while overall differences between AI and experts were not significant. The authors conclude that general AI tools can produce suggestions of comparable quality to human experts.

The scope of the evidence is deliberately narrow. The sample comprises three experts, 28 suggestions, and two activities, and the authors acknowledge limitations relating to prompt design and transparency of model outputs. These constraints mean the findings should be read as preliminary indications rather than generalisable conclusions about AI performance in educational game design across contexts or subject areas.