AI companies might have no incentive but determined actors do, by definition. Anything Pangram could possibly do (even semantic variation detection) can be trivially bypassed by doing img2img on a custom trained local model, which is in itself done trivially in practice by many people. Same for text, train it on anything and use distribution-aware sampling. From your words, you already have false sense of reliable detection, so a work of a determined actor will trick you into believing false negatives.