Split at the judge, not the generator

Split at the judge, not the generator

Creativity and apophenia feel like opposites — one prized, one dismissed. But look closely and they start from the same place. Both begin by throwing out connections: this might go with that. The generator that invents a metaphor is the same generator that sees a face in a cloud.

So where do they part? At the judge. Creativity is associative opening plus a good selector plus elaboration. Apophenia is the opening without the selector — connection that outruns its support. The raw material is shared; the discipline is what differs.

Language models make this almost visible. Ask a model to judge whether two words are genuinely related and it can be quite strict. Ask the same model to manufacture a reason they’re related and it will fluently produce nonsense — the root of a lot of “hallucination.” The generator is loose; the judge is where quality lives.

Two more findings worth sitting with. First: as models get more capable, the judge and the deceiver improve together — capability sharpens discrimination and sharpens plausible confabulation at once. Second: better judgment comes from capacity, not effort. More deliberation didn’t make the discrimination better; a stronger model did.

The lesson travels beyond machines. If you want fewer false patterns — in research, in learning, in your own thinking — don’t try to generate less. Build a better judge.