2025 Volume 2026 Pages 41-46
We present a stochastic AI framework for sketch generation in a multiplayer social deduction game, Drawing Werewolf. In the game, Human-role players receive a specific theme, while Werewolf-role players know only a broader category. To enable AI participation, we trained 345 theme-specific SketchRNN models using a probabilistic encoder-decoder architecture. The AI generates theme-aligned sketches by sampling from a latent space, supporting stylistic variation and ambiguity important for deception-based gameplay. Preliminary evaluations show that the AI’s output blends naturally with human input. Our system illustrates how generative modeling under uncertainty can support interaction in stochastic, partially observable environments. However, limitations remain, including the lack of sketch recognition and drawing dynamics, which are critical for future socially intelligent AI. This study serves as an exploratory step toward bridging stochastic systems theory and creative AI in multi-agent settings involving incomplete information and social reasoning.
