AI Plays Dungeons & Dragons to Learn Human Teamwork

Researchers have discovered an unexpected tool for advancing artificial intelligence: the beloved tabletop game Dungeons & Dragons.

Scientists at UC San Diego built a gaming framework where AI models play D&D alongside each other and with human players. The goal is to test how well AI can make long-term plans, follow rules, and work as part of a team.

The research was presented at the NeurIPS 2025 conference in San Diego this December. Raj Ammanabrolu, the study's senior author and assistant professor at UC San Diego, explained that D&D creates a perfect testing environment because it requires both creativity and strict rule-following.

In the experiments, AI models took on roles as both the Dungeon Master (who creates the story and controls monsters) and as heroes fighting those monsters. Human players could join in any role, creating mixed teams of AI and people working together.

The researchers tested three major AI models: DeepSeek-V3, Claude Haiku 3.5, and GPT-4. They focused on combat scenarios from a pre-written adventure called "Lost Mine of Phandelver," with each test lasting 10 turns.

Claude Haiku 3.5 performed best overall, especially in difficult scenarios. It showed smart decision-making by using more resources when challenges got harder, leading to better outcomes for its team.

Why This Inspires

This research goes far beyond gaming. The skills AI learns through D&D translate directly to real-world applications that can help people.

Models that can plan ahead and coordinate with teams could revolutionize disaster response systems. Imagine AI assistants that truly understand how to work alongside human rescuers during emergencies, anticipating needs and coordinating complex operations.

The same capabilities could optimize supply chains, improve manufacturing efficiency, and enhance search-and-rescue missions. By learning to communicate and strategize through gameplay, AI is developing the social intelligence needed to become a genuine collaborator rather than just a tool.

The beauty of this approach is that it teaches AI through natural dialogue and decision-making. Instead of rigid programming, the models learn to understand context, remember past actions, and predict what teammates (human or AI) might need next.

These gaming experiments are building AI systems that can truly partner with people in meaningful ways.