If you compose a text of enough references and (well-known) in-jokes, and get the perplexity/burstiness stats right, you too can reliably produce text that the AI-detectors think is inhuman. I suspect that doesn't work so well for the latest-generation AI-detection systems (e.g. https://doi.org/10.1016/j.inffus.2025.103465), but there's definitely a way to fool those, too.
