When A.I.’s Output Is a Threat to A.I. Itself

Starts with a basic outline of model collapse and the challenge it faces new AI development, providing several links: "Research has shown that when generative A.I. is trained on a lot of its own output, it can get a lot worse."

It then uses scrollytelling to provide a simple example of how "Just as a copy of a copy can drift away from the original, when generative A.I. is trained on its own content, its output can also drift away from reality, growing further apart from the original data that it was intended to imitate... the only way to stave off this problem was to ensure that the A.I. was also trained on a sufficient supply of new, real data."

It then explores why it happens, showing how for models trained on synthetic data, "Rare data becomes even rarer... the rare, unusual or surprising outcomes — fade away".

Not only will this slow progress, it will make life much harder for newcomers. Moreover, "when there is a lot of A.I.-generated content in the training data, it takes more computing power to train AI". Finally, another threat: "an erosion of diversity." They provide an excellent example of a set of images of human faces fed back to train an AI. "After four generations, the faces all appeared to converge." Similarly, when "trained on their own words, their vocabulary shrinks and their sentences become less varied in their grammatical structure... can amplify biases ... erase data pertaining to minorities. ”.

"One solution, then, is for A.I. companies to pay for [authentic, high-quality] data", hence AI-media organisation deals.

When will this happen? Experts "project that these models may run out of public data to sustain their current pace of growth within a decade."