Why Creativity Cannot Be Interpolated

AI systems can produce individually novel outputs, but novelty alone is not creativity. We argue that genuine creativity requires respect for constraints -- the accumulated structure of prior discoveries -- and that current AI systems lack this capacity because their training takes greedy paths that preclude the right kind of representations.

Key claims

• Creativity requires deep understanding and constraint-respecting extension. "Slop" is the conjugate failure mode -- what you get when novelty is unconstrained by understanding.
• More intelligence and more agency can harm transformational creativity. As a corollary of Kenneth Stanley's Why Greatness Cannot Be Planned, objective-driven optimisation helps with exploratory creativity but inhibits -- perhaps completely -- the transformational kind. This matters for AI efforts trying to square knowledge synthesis with genuine "unknown unknown" discovery.
• Constraints operate at three levels: physical (baked into matter), concrete (instantiated in a fixed substrate), and modelled (represented so they can be manipulated, transferred, and counterfactually varied). Understanding is the cognitive form of this third level -- the capacity to navigate between constrained perspectives and integrate across them.
• Current AI is only meaningfully creative under human supervision. LLMs are coherent within any single frame, but they possess no trajectory of their own; their aggregated voice resolves into a coherent perspective only when a competent human supplies the grounding.
• Human-AI co-creativity is the most promising path forward, though the risk of underplaying human supervision is real -- as evidenced by wild extrapolations about near-term labour-market disruption.
• We leave the door open for better architectures. Any system -- biological or artificial -- whose learned, factored, path-dependent representations let it extend its own phylogeny could be genuinely creative, regardless of substrate.