Playing with Dezgo

Play is an unusually productive context for studying mental models of AI. Games make the interaction explicit, create natural opportunities for think-aloud data, and motivate users to form and test beliefs about the system rather than simply using it. This paper adapted Stable Diffusion for use in a social game — Supe's Terrible Clones, played on Reddit — and analysed how players reasoned about the AI's capabilities.

The findings extend and confirm what the Automation Confusion study found in a controlled game setting. Players did not form models of AI capability by observing the system — they inferred capability from task demand. When the task required the AI to know something, players assumed it did. When a prompt failed, players frequently attributed the failure to the AI rather than to their own prompt, even when their prompt was the cause. Players were unaware of outputs the AI generated without their input, filling in details they had not specified.

These findings were published in 2023. The user behaviours they describe — confident misattribution, unawareness of AI-generated content, capability inference from need rather than observation — have since become widely recognised problems with deployed generative AI products. The playful context made them visible and legible before they became a mainstream concern.

The paper also proposes design considerations for games that use generative AI: structures that make AI failures enjoyable rather than frustrating, constraints that reveal the limits of AI capability through play, and social arrangements that prompt players to discuss and revise their mental models with each other.

Read the paper →