Why AI Watermark Removal is Not a Game Changer

On May 12th, Google announced the launch of its Gemini 2.0 Flash image generation model. According to the release, the goal was to enable “conversational image editing,” where users could provide images and have Gemini edit them.

However, testers quickly discovered that it was very talented at one unseemly task: removing watermarks.

In the tests, the system could (largely) successfully remove watermarks from various stock photo providers, including Getty, iStock, and Shutterstock. Other AI systems either struggle with this task or refuse to complete it.

This sent shockwaves through tech news sites, photographers and visual artists alike. Many expressed concern that the system could aid in copyright infringement, plagiarism and other misuse.

But, as impressive as this tech is, it is not a game changer. Of all the copyright, authorship and other concerns raised by AI systems, this one is relatively minor. In fact, it’s not even a new concern.

An Old Concern, Made New

The idea of automatically removing watermarks is not new. It likely dates back to digital images themselves. However, the topic has been particularly active in the past 10-15 years.

In May 2012, Getty Images announced they were changing their iconic watermark. Though there are many reasons for the change, one key benefit was that it made removing the watermark more difficult.

By then, tools like Snapheal and Photo Stamp Remover were already doing a decent job of removing watermarks. The new Getty watermark was more resistant to those techniques, but technology continued to improve.

What followed was a cat-and-mouse game between artists and tools to remove watermarks. However, artists were always hamstrung by a simple problem; they needed to ensure their images were viewable. Too large or distracting of a watermark might thwart bots but would also deter human visitors.

Watermarks have a hard ceiling with their effectiveness. Fortunately, much of the industry realized that some time ago.

A Change in Tactic

Even if one can’t remove a watermark, the game is over the first time an unwatermarked version of the image is posted. Finding similar photos is as simple as a Google Search; no watermark removal is necessary.

Between that and the cat-and-mouse game with watermark-removing tools, much of the industry has begun to rely less on watermarks and more on detection and enforcement. Companies such as Pixsy, Copytrack, Image Match, Lenstracer, and PhotoClaim provide these services.

One of the industry’s biggest trends over the past 15 years has been the pivot to invisible watermarking (fingerprinting). It’s a technique that’s also being explored as a way to prevent AI systems from training on your images.

This is not to say that watermarking is worthless. A watermark is Copyright Management Information (CMI). In the United States, removing CMI is a separate infringement from the copying. In short, if the alleged infringer removed the watermark, they may be subject to higher damages.

However, artists should not rely on a visible watermark as a first, last or primary line of defense. But that has been true for a very long time.

Where Google is Still Wrong

None of this is to let Google off the hook. Google can and should prevent this model from being used to remove watermarks.

Similar AI tools either fail at the task or refuse to do it. Clearly, Google could create similar guardrails to prevent this use.

While it would not make a massive difference for artists and photographers, it’s still an important step. Google is one of the largest tech companies and one of the largest AI companies. For better or worse, it plays a key role in setting ethical standards in this space.

To be clear, Google’s taking this step would be almost purely symbolic. There are much more significant issues regarding AI, ethics and copyright, but it would at least be a simple consideration.

Google does have some history here. In August 2017, Google researchers devised a new algorithm to remove watermarks en masse. Though they published the equations they used, the researchers never released the software. Furthermore, they explained ways to defeat the algorithm, such as altering the mark from image to image.

However, it’s pretty clear that the Google of 2025 is very different from the Google of 2017.

It’s artists that will bear much of that brunt.

Bottom Line

As impressive as the technology is, it doesn’t change the calculus for visual artists.

Watermarking has been waning in usefulness for well over a decade. Though it is still a useful tool, especially for increasing potential damages, its effectiveness has diminished dramatically.

To be clear, Google is still wrong here. They can and should add guardrails to prevent their system from being used this way. But, given the number of copyright and ethical issues around AI training, this is a relatively minor issue.

If you are a photographer or other visual artist, you should continue to watermark your images and follow the guidelines Google offered in 2017. However, know that it’s possible to defeat even the best watermarks and the only realistic counter is to focus more on detection and enforcement.