Exaptation in AI Development

Exaptation, a term Gould and Elisabeth Vrba introduced in 1982, names the phenomenon by which a feature that evolved for one function (or no function—a spandrel) is subsequently co-opted for a different one. The feather evolved for thermal regulation in flightless theropod dinosaurs and was exapted millions of years later for flight. The swim bladder (buoyancy control) was exapted into the lung. Jaw bones became middle-ear bones in mammals. In each case, the most transformative function was not the original function. Applied to AI, exaptation explains why the most consequential uses of language models are not the uses they were designed for. Claude was optimized to predict text. But creative collaboration, architectural brainstorming, serving as intellectual sounding board—the uses constituting Segal's orange pill experience—are exaptations. These capabilities emerged as structural consequences of fluent context-sensitive generation and were discovered by users who co-opted them for purposes the designers never targeted. If the most important consequences of AI are exaptations, they cannot be predicted by examining designed purposes—the