You can introduce canaries, and ban auto-pwning in general. that's usually banned anyways. Some challenges just can't be solved by a human in under a certain period of time.

Another idea is deep red herrings. solves that lead to more solves, on and on, except only if the previous solves were solved quickly. The effect will be that participants who solve things quickly will keep finding things to solve. they can't know that the path they're on will lead to victory, even if they artificially slow down, unless they consistently slow down just as a human would. It will eliminate the speed advantage. For the skill advantage, other than having another LLM procedurally generate challenges, I don't know of a good solution.

There are always things like captchas. or the good 'ol honor system. A person can spend only so much for things that have no financial reward in the end, only clout.

---

Alright, all that said, i think i really do have a good solution for this, as well as academic exams. Or I think I do, because it's so simple, I've been scratching my head as to why everyone isn't doing it already.

Require screen sharing/recording. LLMs can't fake that well enough. Have another LLM audit the video for mouse, key stroke, window movement and other details to see if it looks human-generated or not.

If a student has an essay assignment, have them record their screen as they research, and actually type out the whole thing. In the extreme, require anti-cheat proctoring software installed, as is done in remote examination. In an even more high-stakes and extreme scenario, have them share their face. Their eye and face movement, correlated with the screen-share, and correlated with the activity observed on the server end, should be pretty hard to beat, even in the next ~5 years of LLM advances.