Invisible watermarks sound neat until you learn that changing one word in ten nearly breaks them, so don’t expect this to settle the homework-cheating argument anytime soon. But it’s still worth knowing. Anthropic has already been doing this for a while. And I’m sure that other AI companies will follow suit.
While video and image AI detection is much easier for humans and machines, even those tools are getting better. Many companies are embedding non-visual watermarks that tools can detect. Words and sentences are a lot trickier (and a lot more interesting). Detection depends on word patterns and phrasing. So it is understandable that if you change some of the words, detection becomes much harder.
My advice: don’t just copy and use AI-generated written content. You need a human in the loop at the end. Regardless, for accuracy and context, read AI-generated content and don’t take it as-is.