I would actually say the opposite. We know that the biological brain does a lot of iterative refinement via recurrent processing (attractor dynamics), which is

I would actually say the opposite. We know that the biological brain does a lot of iterative refinement via recurrent processing (attractor dynamics), which is very similar to how diffusion works. However, while prediction is also a core functionality of the brain, it's not really that you always auto-regressively generate word-by-word.