An ‘uncertainty principle’ for generative AI: the more we measure a culture, the less we can predict where it is heading

Three researchers argue that AI finally makes it possible to measure how individual thought and culture shape one another, but that, by becoming embedded in social life, it makes that culture less predictable.

Cognition and culture have long been studied in parallel, linked by metaphors rather than mechanisms. Douglas Guilbeault, Bhargav Srinivasa Desikan and James Evans defend the idea that they constitute each other and that this becomes measurable: each subjectivity can be seen as a trajectory in a "cultural field" with a very large number of dimensions, and this field is only the sum of these trajectories.

The geometry of culture

Vector embeddings and large generative models would provide for the first time a common framework, from an individual's thought to mass ideologies, including communication within organizations and non-textual content. The authors point to the possibility of predicting simultaneous discoveries made by distant minds: if we know the shape of the field, we can anticipate the paths that research takes in it.

When the instrument becomes an actor

The text then focuses on generative AI agents, both simulators of human subjectivities and a new medium for social coordination. The authors formulate a "Heisenberg-like" principle of uncertainty: the more precise the tools for situating a culture become, the less we can predict its trajectory, since the machines that describe the culture now contribute to producing it. They conclude on the ethical questions of AI-induced cultural drift and synthetic culture that feeds on itself, and propose rules for safe and privacy-friendly research.