Possibility Architectures: Exploring Human Communication with Generative AI

Tune in for the live stream on YouTube or Twitter.

Abstract: What we mean is given not just by what we say, but the things we could have said but didn't. More than that: as we navigate the possibilities before us, we conjure, at a distance, the ones we will later encounter. These emergent architectures, crucial to understanding human communication, appear as everything from verbal tics and cliché to higher-level figurative constructs such as irony, metaphor, and style. They sculpt possibility space on different timescales in ways that answer to cognitive bottlenecks and resource constraints, and social demands ranging from empathetic collaboration to self-censorship. Applying information-theoretic tools to the outputs of GPT-2, ChatGPT, and BERT, we reveal basic patterns in these possibility architectures. Functional, "Bauhaus" prose from the New York Times, for example, arranges possibilities in dense but predictable ways that are very different from the ill-defined Levy-flights of dream journal and the Gothic structure of Cormac McCarthy's *Stella Maris*. Our work reveals new scientific possibilities for Generative AI, and new insights into what — for now — makes us uniquely human.