Are you in the "it's just like with calculators" camp? I think there's a fundamental difference between a tool that just simplifies some very specific solution steps vs a tool that takes the literal assignment question, regardless of how "updated" or "more difficult" it is, and outputs an answer that you can probably submit without even checking for a passing grade. Watermarking, guardrailing etc. won't fundamentally fix this, there will always be access to more unfiltered models.
> Cheaters, by definition, avoid real work, so they won't be using custom models unless the stakes are sufficiently high.
I think it's a non-sequitur, some might get more enjoyment out of tinkering with LLMs than out of whatever the assignment is asking you to do. But even if you don't believe that, it should be obvious that the ones using the tool needn't be the same as the ones putting in the work.
