Do proof-of-work pages actually stop AI bots? Big AI companies have enough compute to solve these challenges at scale. And if their bots are already doing much heavier work to fetch, read and process each page, then solving a small challenge first seems unlikely to be a serious barrier. Who are these proof-of-work challenges actually helping?
I believe poisonous loops of non-sense text are the best choice in terms of LLM capabilities and human distinguishing potential; the next iteration could be non-sense with reasonably intact grammar and content. At the very least, show some content whilst doing the PoW (isn't the point rising computational costs? Give me something like YouTube video decryption instead of having me wait at least) OR use the PoW for some useful protein-folding, finding the next prime, or an alternative monetization scheme.
Anubis does this: https://anubis.techaro.lol/docs/admin/honeypot/overview
Getting in the maze influences your client's challenge difficulty.