The Future of AI Cannibalism

Chat GPT isn't seeing a decrease in quality right now, but so-called AI cannibalism might lead to diminished performance. This could come as the result of more and more text from large language models being included in the training sets of future training models. Excluding this content is actually really hard right now because LLM generated text detectors aren't reliable. Ars Technica's a Benj Edwards found it looks for degrees of perplexity and burstiness. If it's all uniform or you're going run on sentence and then staccato bang bang, the more perplexe and burstiness theoretically, the more human. But people can be boring and AI is getting more complex.