I hate to say it, but the idea is doomed to failure from the start.
Both human writing and AI writing are moving targets, and you have no real visibility into the mechanics of either one. By the time you test and validate any detector, it will be obsolete. You will never have the opportunity to test an individual validator's effectiveness over time, because major new models are released every month or two. And the prevalence of AI writing is already influencing how real people write (especially young people who are only learning to write in the age of AI).
I'm not sure what the answer is here. But pouring time and money into a bad idea just because you don't have a good idea is not a winning strategy.