The usual question about language models is whether a system that predicts the next token can ever produce something truly new. I think that question points at the wrong constraint. Generating novel output is easy for these models, and they do it constantly. What they lack on their own is a way to tell a good new idea from a bad one, and that gap, not the token-by-token mechanism, is what limits them.
Margaret Boden’s categories help here. Models are strong at combinatorial novelty, joining familiar ideas in unfamiliar ways, and at exploratory novelty, pushing to the edges of an established style or rule set. They struggle with transformational novelty, the kind that changes the rules, like Cubism or quantum mechanics. The reason is not that they can’t produce unlikely ideas. Raising the sampling temperature produces plenty of those. The reason is that a transformational idea looks, from the inside, exactly like a mistake, and a model trained to match existing text has no basis for telling the two apart. People who make breakthroughs generate a lot of unlikely ideas too, but they test them against evidence and discard almost all of them. The selection is where the value comes from.
This is why the recent gains in AI discovery have come from domains with a cheap, reliable judge. Code either runs or it doesn’t, a proof either checks or it doesn’t, and a protein structure either matches the measurement or it doesn’t. In those settings you can let a model generate thousands of candidates, keep the ones that pass, and train it on what worked, which rewards being right rather than sounding human. That loop is how systems have found better sorting algorithms and new mathematical results.
The practical implication is that you can predict where AI will produce genuinely new work by asking where verification is cheap. In math, software, and parts of the physical sciences, expect real invention. In strategy, policy, and most organizational decisions, where the only check on an idea is years of consequences and a lot of argument about what caused them, expect excellent recombination of existing thinking and very little that changes the frame. For those problems the scarce resource is still human judgment about which ideas deserve to survive, and better models don’t change that until someone builds a better judge.