The Moat That Was Not Deep Enough

For two decades, the creative class appeared protected by a moat that looked permanent: the structural distinction between routine work (automatable) and non-routine creative work (human-only). This distinction was not merely Richard Florida's; it was the foundation of labor economics, encoded in David Autor's task framework and embedded in every economic forecast about the future of work. The logic was compelling: computers could follow instructions with inhuman speed, but they could not generate the instructions. They could execute plans, but they could not create plans. They could optimize within a framework, but they could not imagine the framework. Creative work — the production of genuinely novel solutions, original designs, and non-routine cognitive output — required flexible thinking, contextual sensitivity, aesthetic judgment, and the ability to navigate ambiguity. These capacities appeared categorically beyond machine capability. The moat was real, empirically validated, and economically consequential. Then large language models learned to produce fluent, contextually appropriate, often genuinely novel output by predicting the