Can consciousness emerge when AI agents invent their own language?

Studying whether AI can be conscious is hard because most approaches either compare systems to theory checklists or build consciousness-inspired modules directly—both risk seeing what humans expect to see. Wu and Xiao propose letting agents invent language and concepts through multi-agent reinforcement learning with minimal starting conditions, ensuring any structure reflects actual task demands. In a simple environment, agents spontaneously developed self-referential communication and an echo-mismatch detection mechanism not explained by task structure or architecture alone.