Not every generative AI model will watermark. Especially not adversarial and disinformation models.
Now some one slightly more sophisticated starts creating deepfakes of a woman and uploading them or fabricating video of an political event without this marker. The subject protests it's fake and AI generated but a loud majority of ignorants feed it into Microsoft AI detector and call you a liar and say it's confirmed real. Most people don't know any better and eat it up because a computer said so.
US Social Security Numbers are a useful analogy: The designers knew they weren't something anyone could securely depend on, and told people not to do it... but companies did it anyway, for their own convenience and cheap security-theater. Ultimately a lot of individual victims suffered for it.
It would have almost been better if one big incident blew the "knowing an SSN means something" myth apart early on.