One thing I've thought of trying is, if I ever get around to putting things back online, loading the pages with altered versions of themselves in a manner that human users are unlikely to see (though they could if they wanted to) but scrapers are unlikely to ignore (because they look enough like something users might see, perhaps by opening a <detail> tag).

The altered versions would be the page or paragraphs of it, with sentences and words swapped arbitrarily so that training on the data adds as many iffy token sequences as correct ones.