Playing Telephone with AI: How Topology Helps Us See Patterns in Generative Systems

Do you remember that childhood game telephone?

The one where a whispered message circles a room, passed on player to player, and by the time it reaches the last person it has often transformed into something hilariously different.

The process is simple yet profound: as information passes through the multiple ‘media’ of human voices, expressions, and memories - it mutates, drifts, and sometimes even forms entirely new meanings.

This same idea of ‘drift’ in a message can be taken into the world of artificial intelligence (AI). Instead of children passing along a message, imagine a network of generative AI models, each transforming information from form to form - text to image, image back to text, and on and on. This is the core concept behind PANIC (Playground AI Network Interactive Creativity), our experimental multimodal AI system.

In PANIC, each output is fed back in as the next input, creating an iterative loop, generating sequences of evolving images and texts. Just like in the game of telephone, we observe semantic drift: meanings shift, details morph, and surprising new patterns emerge.

And here’s the intriguing question: amid the apparent chaos, is there hidden structure?

Can we identify patterns in how AI ‘conversations’ evolve? And if so, how might we measure them?