The solution to this is subtle and gentle amounts of 2D Perlin Noise, as well as a touch of Gaussian noise on and around the watermark as well.
The more you can cloak the area around the watermark with subtly increasing amounts of noise; the harder it is for AI to manipulate it without mangling the image in general. (or leaving the watermark behind)
Similarly; leaving smaller artifacts like small signatures or wordmarks embedded in the image also makes sense, particularly small signatures hugging things like lineart in inconspicuous places or hidden in places with intricate detail.
Tools like Glaze and Nightshade also exist to "Poison" images at creation-time such that, if they go viral and get re-shared and AI remixed heavily, they won't be as easily usable by AI models to knock off your works. Yes, this technique is ineffective for existing works, as an ounce of prevention is worth a pound of cure. These tools use many different techniques beyond just subtly masking the entire image with multiple layers of imperceptible perlin noise. Which is a task that could take you several hours to get to looking right in your favorite image editing suite, as you'll be poking and prodding and tweaking that slider to maximize protection while minimizing it's visual perturbations.