"Even if conflicts like this don’t impede the advance of A.I., it might be stymied in other ways. At the end of May, several A.I. researchers collaborated on a paper that examined whether new A.I. systems could be developed from knowledge generated by existing A.I. models, rather than by human-generated databases. They discovered a systemic breakdown — a failure they called “model collapse.” The authors saw that using data from an A.I. to train new versions of A.I.s leads to chaos. Synthetic data, they wrote, ends up “polluting the training set of the next generation of models; being trained on polluted data, they then misperceive reality.” The lesson here is that it will prove challenging to build new models from old models. And with chat-bots, Ilia Shumailov, an Oxford University researcher and the paper’s primary author, told me, the downward spiral looks similar. Without human data to train on, Shumailov said, “your language model starts being completely oblivious to what you ask it to solve, and it starts just talking in circles about whatever it wants, as if it went into this madman mode.” Wouldn’t a plug-in from, say, Wikipedia, avert that problem, I asked? It could, Shumailov said. But if in the future Wikipedia were to become clogged with articles generated by A.I., the same cycle — essentially, the computer feeding on content it created itself — would be perpetuated."