So far we have a single ai vendor introducing a statistical bias to their generation that can make it simpler to identify genAI in some (longer) texts. We don’t know yet how effective that will be in practice
Technologically I think the issue is unsolvable. You can easily remove any “watermark” in text with minimal effort. However, as most users of AI are shockingly lazy it will probably be effective because they just can’t be bothered. And of course we will end up with texts being flagged wrongly, with no way of proving the flagging system wrong. Just a complete shitshow.
This comment on HN sums it up nicely:
Technologically I think the issue is unsolvable. You can easily remove any “watermark” in text with minimal effort. However, as most users of AI are shockingly lazy it will probably be effective because they just can’t be bothered. And of course we will end up with texts being flagged wrongly, with no way of proving the flagging system wrong. Just a complete shitshow.
So how can people actually detect these watermarks, is it only accessible to the company that produced them?
And of course people will have no understanding of base rate fallacy: https://en.wikipedia.org/wiki/Base_rate_fallacy