Use the model to generate some AI output. Then use that output to train a new instance of the model and use the resulting output to train a third version, and so forth. With each iteration, errors build atop one another. The 10th model, prompted to write about historical English architecture, spews out gibberish about jackrabbits.

That this happens doesn't surprise me, but I'd love to see a curve of how each organic vs machine content mixe ratio results in model collapse over N generations.