Summary

  • Google’s Gemini 2.0 Flash AI model efficiently removes watermarks from images, including those created by humans.
  • While watermarking helps establish ownership and prevent copyright infringement, Gemini can easily circumvent these protections, raising concerns.
  • Despite ethical concerns, Gemini is currently free to use and lacks restrictions on watermark removal, potentially enabling widespread misuse.

Respect for intellectual property rights has been a bone of contention with AI training, where several creators are accusing tech firms of pirating data or acting without consent. Similar concerns remain about the media generated by these AI models and the implications of how they can be used to manipulate other watermarked content too. Google’s Gemini 2.0 Flash is one of the company’s latest lightweight AI releases, but it is stirring up trouble because several people have found it adept at removing watermarks from images.


Related


13 simple Google Gemini tips and tricks for maximum productivity

Summarize, simplify, streamline, and more



2

The efficacy of Google watermarking AI-generated images with telltale EXIF data is a relatively minor concern when real imagery created by humans is stripped of watermarks. Watermarking helps establish ownership, dissuades copyright infringement, and allows stock media outfits to help creators make a living off their work. As such, US copyright law reportedly deems watermark removal illegal if done without the original owner’s consent.

However, TechCrunch reports Google’s Gemini 2.0 Flash model is adept at stripping images of watermarks. It goes a step beyond and even fills in larger areas where telltale signs of removal may survive (via AndroidAuthority). The results seem comparable to Adobe Photoshop’s Generative Fill tool, but some Reddit users note that Gemini seems to re-draw the entire image instead of painting in the details, evident from the slightly crisp outputs.

That said, Gemini is currently free to use while Adobe’s suite needs an active subscription. With Gemini, you just supply the watermarked image paired with an instruction for removal, and you receive the result with minimal manual effort (albeit with a Gemini watermark now inserted in the corner). Other tests show Gemini generating images of celebrities and copyrighted fictional characters without hesitation.

Other AI models simply refuse to edit images like this

There is a silver lining to the cloud

It’s easy to see how Google’s AI can be misused to use stock images for free, potentially without repercussions, and at scale due to its sheer simplicity. Perhaps the only saving grace at this time is that Google says Gemini 2.0 Flash is still an experimental model, and you must use a developer-facing tool like AI Studio to upload images. Nonetheless, rival AIs like OpenAI’s GPT-4o and Claude 3.7 Sonnet straight up refuse to remove watermarks from images, cautioning users about the ethics and legality.

Google didn’t respond to TechCrunch’s request for a comment on the matter, but it could land in hot water if watermark removal isn’t restricted or deterred.