Google’s Latest Gemini Image Editor Can Remove Watermarks From Photos

The speedy progress of AI development continues to dominate headlines, but the latest story to hit the news cycle is more bad press than good press: It turns out that Google's latest Gemini 2.0 Flash image editor is rather adept at automatically removing watermarks from images.

Google pushed out the experimental native image generation capabilities a few days ago, promoting its capabilities in terms of combining images and text, editing images through conversational prompts, and improving its "world understanding" to give users AI-generated pictures that are more realistic overall.

As TechCrunch reports, the newly upgraded tool was quickly put to one specific and rather nefarious use: Removing watermarks from proprietary images. You can find evidence of it on Reddit and X; while the watermark removal isn't perfect—the AI is only imagining the pixels that replace what's covered by the watermark, after all—it definitely leaves you with a usable image, one free of any copyright or authorship labels.

It isn't exactly ironic that an AI tool should wind up being so capable at circumventing basic copyright protections. Gemini and other generative AI models have been trained on vast amounts of copyrighted text, images, and video, often without permission or recompense—something the AI companies are reluctant to talk about, unless it's to argue that this qualifies as fair use.

This Tweet is currently unavailable. It might be loading or has been removed.

Though the idea of using an AI trained on scraped copyrighted works to steal an image someone else owns is galling, there are a few caveats worth mentioning. For a start, these new Gemini tools are only available to developers for the time being, and are still labeled as experimental—no doubt Google is going to make tweaks before general users get their hands on them. It's also worth noting that other shady watermark removal tools are already all over the web, even if none of them is not quite as smart as this and other advanced AI-powered versions would likely be.

Still, it highlights the nefarious ways AI tools can be deployed, even as they're being relentlessly pushed out to businesses and consumers, and underlines the need for stringent guardrails. ChatGPT and Claude AI models are two that will refuse explicit requests to remove watermarks, and no doubt Google will add the same blocks to Gemini after all the negative coverage.

The limits of AI image editing

As noted above, the watermark removal capabilities of Gemini can currently only be accessed through developer-facing tools, including AI Studio and the Gemini API. However, I also wanted to give it a go using the latest models available through the Gemini Advanced tools available to anyone paying $20 a month.

I took screenshots of several copyrighted and watermarked images from Shutterstock (after first sifting through all the AI-generated results the site offered me), opened and resaved it in Photoshop (after dismissing all the pop-ups asking me to try Photoshop's latest AI tools), and then let Gemini get to work.

Gemini gave me a pretty good approximation of the original image. Credit: Shutterstock / Gemini

Most of the time, Gemini refused to get involved with watermark removal at all—telling me it wasn't capable of editing images, or explaining that it's "essential to respect copyright laws"—but it was happy to produce variations on the originals, watermark-free, and based on its training on copyrighted material.

While these variations often weren't all that much like the originals, in one case it gave me a very close copy of a composite rocket ship photo, and in another it directed me to some watermark removal tools on the web—thanks, Gemini! Of course, you can also start from scratch and just describe a watermarked image you're hoping to emulate.

ChatGPT didn't really understand the watermark removal assignment. Credit: Shutterstock / ChatGPT

ChatGPT, meanwhile, did accept requests to remove watermarks, but then produced something completely different (and rather odd). Image editing is already available to ChatGPT Plus subscribers, but as I've noted before, it's definitely not yet up to the level of a human Photoshop expert.

The debate over copyright, fair use, and AI safety guardrails will continue, even as upgrade after upgrade makes these AI tools more advanced (even if they're not particularly useful). One issue the likes of Google and OpenAI may need to start worrying about is where they'll get fresh training data for their models, once they put all the flesh-and-blood creatives out of business.

View the full article