Subtle changes in the words used by generative AI tools could act as a virtual watermark for the text they produce to tip off readers, research has found.