AI really will get dumber the more you feed it.

Researchers at Oxford and Cambridge have named this phenomenon “model collapse.”
It means that when AI starts training the next generation of AI with content generated by AI, the models become duller and duller from one generation to the next, until they gradually forget what real human data actually looks like.
It’s like in the real world there are 50% red flowers, 30% blue flowers, 15% yellow flowers, and 5% black flowers.
After the first generation of AI learns these data, it may generate 55% red flowers, 30% blue flowers, 14% yellow flowers, and 1% black flower.
The next generation of AI then uses this content to continue training, and it may become 60% red flowers, 30% blue flowers, 10% yellow flowers, and 0% black flowers.
After continuing this loop for a few generations, the black flowers disappear, and the yellow flowers become fewer and fewer, until the model comes to believe that the world is almost only red flowers and blue flowers.
In the real world, that translates to rare but real content being diluted round after round, so that what remains becomes increasingly “safe” and increasingly similar.