AI Optimises Paradigms. Humans Break Them.

In 2020, DeepMind's AlphaFold solved one of biology's hardest problems — predicting the three-dimensional shape of proteins from their amino acid sequence. It was a genuine scientific breakthrough, described by Nobel committee members as a revolution in structural biology. AlphaFold had done something that would have taken human researchers decades. It is the most compelling demonstration of what AI can do. It is also the clearest illustration of what AI cannot do. AlphaFold worked within the existing framework of biochemistry. It didn't discover that proteins exist. It optimised within a paradigm that humans built.

This distinction — between optimising a paradigm and breaking one — is the most important idea in the entire AI debate, and almost nobody is talking about it.

Every major leap in human knowledge was not a better answer to an existing question. It was a completely new question that made the old framework obsolete. Copernicus didn't calculate planetary positions better than Ptolemy — he discarded the earth-centred framework entirely. Germ theory didn't improve the miasma theory of disease — it replaced it. Einstein didn't refine Newtonian mechanics — he showed that Newtonian mechanics was a special case of something more fundamental. These paradigm breaks share a common feature: they required a human mind to look at existing evidence and see it completely differently. Not more efficiently. Differently.

AI, trained on the totality of existing human knowledge, is the most powerful optimisation engine ever built. Given a paradigm, it can exhaust its possibilities faster than any human. AlphaFold proved this. But AI cannot stand outside a paradigm and see it for what it is — a framework, not a truth. That cognitive act — the one that precedes every genuine scientific revolution — requires a mind that can entertain the possibility that everything it knows is wrong. That is not a limitation of current AI architecture. It is a deeper question about the nature of discovery itself.

Sam Altman captured this carefully in his 2025 essay "The Gentle Singularity": he predicted that 2026 would likely see AI systems capable of generating novel insights — but framed this as an AI research intern, not an AI research director (Altman, 2025). The distinction matters. An intern can contribute within a project. The director changes what the project is.

The implication is not that AI is less transformative than advertised. It is that the transformation is different from what most people fear. The jobs most at risk are not the ones requiring the deepest human judgment — they are the ones doing the most repetitive optimisation within existing frameworks. The jobs most resilient are those that question the framework itself. That's not comfort for everyone. But it is a map.