Oops! Gemini 2.0 Flash Accidentally Erases Watermarks

This is going to cause Google so much trouble! They faced severe criticism last year when their AI Overviews in Search were giving insane pieces of advice to users. Now, there is a new model that can remove watermarks from copyrighted images.

Google Gemini AI Model can Remove Watermarks

Things started with X (Twitter) where users shared this discovery. They asked Gemini 2.0 Flash model to remove watermarks from images. This includes images from paid sources like Getty Images, iStock or Adobe Stock. Here’s what they got:

As you can see, Gemini AI intelligently fills the gaps in the image and edits in a way as though the watermark was never there. This boasts impressive image generation and editing capabilities also.

One of the ironies is that Gemini added its own watermark in the bottom-left corner.

Check out some more examples:

In the above image, there are a lot of watermarks, including texts & icons, and they are all over the image. Still Gemini did the work in seconds.

The image here has so many objects but a lot of small watermarks, but it was still an easy job for Gemini. Almost perfect!

I also tried to test this myself and here’s how it went:

Note that I had intentionally blurred the output image because we don’t own the rights to it and we don’t support such type of manipulation to the copyright images. This news article is for educational purposes only.

I must also mention that this ability is only available in Gemini 2.0 Flash’s image generation feature, which is labeled as “experimental” and “not for production use”.

How did users remove Watermark from Images?

Here’s how users are doing it: Go to Google AI Studio, select “Gemini 2.0 Flash (Image Generation) Experimental” in the model picker. Then, in the prompt box, attach an image and add the prompt to remove the watermark.

That’s it, that’s how easy it is!

This Raises Legal Concerns

This feature highlights how powerful AI models have become, especially in image processing. But it also raises ethical and legal questions. Removing watermarks without the consent of the copyright holders is illegal in almost all countries and undermines the intellectual property rights.

That’s why other AI models, like OpenAI’s ChatGPT and Anthropic’s Claude, have implemented in-built restrictions to prevent such action.

We also want to remind you again that this ability to remove watermarks is available only to the “experimental” model.

If you try to erase the watermark texts using Google Gemini 2.0 Flash model (that is available for the public):

It didn’t allow it even after trying multiple times.

Still, some copyright holders might take notice of this “experimental” model and take Google to court if they want. This is definitely illegal according to the US copyright law.

However, the model’s text-to-image generation feature also appears to have some guardrails: it cannot be used to generate images depicting real-world celebrities.

Takeaways

AI models are already receiving backlash in such a matter. The way they have trained the AI models using images and videos from the internet is under scrutiny from day one. OpenAI’s SORA AI video generation model is one of such cases.

Overall, as AI technology continues to advance, the balance between innovation and ethical use has become increasingly important.