Why ordinary watermarks fail against AI watermark removal
Three months in
Stopping AI watermark removal is two tracks. Local pixel displacement leaves obvious scars when removal algorithms run, so restoring a clean original stops being cheap. Faint characters and frequency-domain marks stay in files that were never deeply redrawn, so later attribution is still possible.
JWProtect has been live for three months. The backend has processed about 900 images.
Watching how people use it, and the feedback they leave, I keep coming back to one question. In front of today's AI tools, does protecting an image still do anything?
The internet usually offers two extremes. One side thinks an invisible watermark is enough and you can stop worrying. The other side thinks it is all wasted effort. Once a picture is on a screen, anyone can run it through AI and every mark will vanish.
Everyone is hunting for a universal lock. There is no such thing. Against theft and AI watermark removal as they actually happen, we ended up with two paths.
Path one. Displacement that breaks structure so AI cannot cheaply restore the original
An AI eraser can wipe a typical semi-transparent overlay in half a second.
So the visible marks on the Protect page, displacement especially, are not a stamp. They deliberately disturb local structure. Tiny shifts along line edges and light-dark boundaries.
To a person, it is just extra texture and the picture is still readable. Removal models expect smooth neighbors to inpaint from. When those neighbors are twisted on purpose, the fill fails. Force the wipe and you get dirty patches and odd warps.
To erase that fully, the thief has to crank redraw strength. Crank it high enough and the drawing turns into something else. Warped faces, extra limbs, a different picture.
That raises the price. A clean original is no longer cheap. For anyone washing images in bulk, one file that needs heavy manual cleanup is often not worth it.
Path two. Faint characters and invisible marks so the thief carries the evidence away
Plenty of illustrators and photographers do not want the surface disturbed at all.
That is the second path. Do not make it obvious.
Faint characters and invisible layers (frequency-domain watermarks and steganography) sit in the tones, or stay invisible. After Protect, the Verify page can read those marks.
The goal is a false all-clear. A casual look suggests raw, unprotected art, so the file gets posted without an AI wash step.
If it was never redrawn, the buried claim stays in the file. When that image shows up elsewhere, the evidence is still inside and can be extracted.
Why evidence on the file is not enough. Provenpix
Hardening the file, or letting someone walk off with a marked copy, is only the first half.
Most creators cannot search the whole web every day. Even when they catch a steal, platform complaints, support tickets, and evidence packs are tedious enough that many give up.
Protecting the image and keeping the evidence is the first half of the match.
Provenpix is the product we are building for the second half. Keep watching shops and commercial accounts for unauthorized use of those marks. When a hit appears, assemble a timestamped comparison pack and help organize what the platform asks for.
Defense is an ID for the work. Monitoring and complaints are how that ID gets used.
There is no absolute anti-theft. Keep the original.
Keep your source files. That is the honest part.
Anyone promising 100% theft-proof software that can never be beaten is selling a story. If someone is determined, they can even trace pixels off a screen. No tool stops that.
JWProtect and Provenpix exist for a simpler reason. Drawing and photography should stay a joy, not a panic before every post. Give serious creators tools that make a hard wipe expensive, and make attribution hard to deny when they need it.