The data feeds itself and results in significant algorithmic lapse/inbreeding. This means that the models degrade and become less useful over time, as they are trained on the same data over and over again. This can lead to a lack of diversity in the data and a lack of new information being fed into the models, which can result in poor performance and inaccurate predictions. Perhaps even a kind of tendency to hallucinate broadly requiring that the entire dataset used to train the model be re-examined and re-validated, or even re-collected or destroyed/scrapped. But perhaps there is a curious aftermarket for these models very much like there's an aftermarket for shitty Nike t-shirts that are so bad they're good for something, like global south resulting in a curious culture of hallucinating models that have effectivity in specific domains like witch doctors, oracles, or other kinds of divination practices.