How to Place Watermarks When AI Can Remove Almost Anything

What still helps – and what no longer does

AI tools for removing watermarks have changed a lot.
Things that worked well only a few years ago now fail in seconds.

This does not mean watermarking is useless.
But it does mean that the goal has changed.

Today, watermarking is no longer about hiding marks or making them invisible.
It is about making reconstruction costly, messy, and visibly damaging.

This article explains what still matters, what no longer does, and why placement decisions need a rethink.

The uncomfortable truth about watermark removal

Modern AI removers can now:

detect repetitive patterns
smooth noise-based structures
rebuild textures like grass, walls, fabric and wood
recreate hair, hands and even faces (sometimes inaccurately, but usable)

There is no perfect protection anymore.

A watermark is not a lock.
It is a disturbance.

A good watermark does not aim to politely disappear for the eye of the customer while still protecting the image (spoiler: it can’t).
It aims to leave visible damage when removed.

Why classic watermark advice no longer works

Many older recommendations were based on ideas like:

“Use complex textures”
“Repeat the watermark many times”
“Noise patterns confuse AI”

These approaches worked — until AI learned to recognise repetition and regularity.

Today:

repetitive grids are detected and wiped
uniform noise gets averaged out
simple texture patterns are reconstructed surprisingly well

That is why many older noise-based watermark designs, including my own early versions, no longer hold up. Single-color watermarks had to go even before that, as they were a mere calculation away from being wiped.

Surfaces that look safe – but are actually easy to rebuild

This is where many people still get caught out. AI-watermark removers and AI image reconstruction are getting crazy good.

Some surfaces look complex to the human eye, but are statistically simple for AI models.

Flat backgrounds (no surprise)

blue sky
studio walls
blurred backgrounds
snow, sand, fog

These are rebuilt almost perfectly.

Uniform grey (often underestimated)

Mid-grey areas are especially risky.
Many reconstruction models use grey as a neutral base, making removal cleaner than expected.

Regular architecture

This surprises many people.

Examples:

brick walls
tiled floors
window grids
modern buildings

Even though they look detailed, the repetition makes them easy to recreate.

Grass, vegetation and foliage

Once considered “safe”, now much less so.

AI models have learned:

grass textures
leaf patterns
bushes and forest floors
basic tree patterns

Results may not be botanically correct — but they are visually convincing.

Hair, hands and faces

Another common misconception.

AI can now:

rebuild hair structure
correct hands
replace faces

Even when the face changes slightly, the image may still be usable — especially outside portrait contexts.

So what still matters about placement?

Placement no longer guarantees protection.
But it does influence how broken the image becomes after removal.

The goal is no longer:

“Can the watermark be removed?”

The real question is:

“How damaged does the image look after removal?”

Placement that still increases resistance

1. Transitions, not textures

AI struggles most at boundaries.

Good examples:

light ↔ shadow
object ↔ background
skin ↔ clothing
sharp contrast edges

Here, reconstruction often leaves halos, blur or warped edges.

2. Real information inside the image

This still works better than many expect.

Placing a watermark over:

shop or restaurant signs
menus
street signs
trail markers
printed posters
text on T-shirts
unusual folds in clothing (especially in combination with text and images on shirts)

…forces AI to reconstruct meaning, not just texture.
Generated text is often wrong, inconsistent or obviously artificial.

3. Irregular, non-repeatable coverage

Any pattern, any grid follows logic. And logic is what AI excels at.

Watermarks that:

have no symmetry
no repetition
no predictable spacing

are harder to fully neutralise.

This is where random-layered cloud structures outperform older pattern-based designs.

Where watermarks are often a bad idea

There are two areas in which watermarks are no longer of much use or accepted by the audience or platform.

Social media

visible watermarks reduce reach
platforms discourage overlays
compression destroys subtle structures

Many creators accept limited protection here and choose their images accordingly.

Product images

marketplaces often dislike watermarks
trust and clarity matter more
visual noise can reduce conversions

In these cases, watermarking may not be the right tool. You could (and should) include methods of branding the images, but watermarks are no longer the tool of choice here.

A realistic goal: visible damage, not invisibility

Modern watermarking is about raising the cost of theft.

A strong watermark:

does not disappear cleanly
leaves visible artefacts
makes reuse unattractive

If removal ruins the image just enough, the watermark has done its job.

Why I moved to random-layered cloud watermarks

Because repetition became the weakness.

My current approach avoids:

regular grids
predictable noise
symmetric layouts

Instead, it uses:

random-layered structures
non-repeating geometry
additional random noise passes

Not to look elegant, though I loved the elegant structures, - but to interfere with reconstruction. Automated removal should leave botched up images behind and dedicated thieves should have to do some real work for their imperfect results.