[Atlas22]

Its potentially an even worse problem than overfitting because of the error accumulation and undermining of the fitness functions. Its similar to asking students to create test questions to evaluate themself. As the proportion of self output questions grows, the context eventually drifts to be meaningless. In other words it leads to test "questions" like "The answer is A".

One measurement of this would be to watch the growth of probability of new models to say similar things to older model filters like "As a large language model ..."