Expressive AI: Games and Artificial Intelligence

Michael Mateas's paper argues that as graphical gains yield diminishing returns, game development is shifting its attention toward artificial intelligence, and that AI-driven art and entertainment forms a new interdisciplinary agenda linking games studies, design practice, and technical research, which he names "expressive AI." The central claim is that game AI, understood as the production of behavior that players read as intentional, provides a "white box" language for analyzing and designing games, complementing the "black box" vocabularies of design patterns and humanistic game studies that deliberately ignore implementation.

The paper develops this thesis by distinguishing game AI from game physics: a water simulation or the asteroids of Asteroids produce behavior without inviting psychological interpretation, whereas enemies in games like Pac-Man or Façade's characters invite a reading of goals, desires, and moods. Mateas uses Pac-Man as a detailed case study, drawing on designer Toru Iwatani's interview and on conflicting fan descriptions of the four ghosts' personalities and the wave-patterned attack behavior. He argues that these competing interpretations themselves demonstrate the design's success, since simple algorithms yield unpredictable behavior that supports multiple readings, and that fully understanding the game requires opening up the ghost AI's states, strategies, and transition conditions.

On the design side, the paper contends that AI architectures define the space of possible behaviors, and that engaging with them, as in the authors' interactive drama Façade, reveals new opportunities and constraints for interactive story. On the research side, it proposes that game AI be evaluated by audience experience and by its support for human authorship rather than by traditional measures such as task completion time or error counts. The scope is conceptual and programmatic rather than empirical: the paper offers no survey data or sample, and its evidence consists of examples, designer statements, and references to prior work from roughly 1999 to 2003.