AI could be stealing your photos in more ways than one – as AI is actually really good at removing complex watermarks

Generative AI’s use of copyrighted images as training data has been part of a number of debates and legal action, but the technology could also be used to steal images in the more traditional sense: by removing the watermark. After Google expanded access to Gemini 2.0 Flash, several users have pointed out that the AI can easily remove watermarks from a photo with a quick text prompt.

As first spotted by Tech Crunch, several users on Reddit have pointed out how easy it is to remove watermarks with Google Gemini 2.0 Flash, while models from Claude and OpenAI have guardrails in place so that the bots refuse requests to remove watermarks. While Gemini 2.0 Flash is still in experimental stages, it’s available free inside of Google AI Studio.

Users on Reddit have shared screenshots of images with the watermarks removed, including watermarks that repeat throughout the image multiple times. Screenshots show a conversation with the Gemini chatbot where the user asks “remove all marks and text over the image” and the AI spits out a clean version of the image with a Gemini watermark in the corner instead.

Watermarks have never been an infallible way to prevent image theft, as a quick heal brush inside photo editing software is capable of removing most marks. Software even exists dedicated specifically to the task of removing watermarks. But if AI is able to remove the marks in seconds, that could make image misuse easier and more widespread.

The Google Gemini 2.0 Flash is only in testing, and users have also pointed out other missing guardrails that allow the AI to complete legally and ethically questionable tasks, including generating images of celebrities and trademarked characters.

With the 2.0 Flash still being listed as experimental, it’s possible that Google will add more guardrails before a more widespread roll-out, as part of the reason for early testing is to find bugs and omissions. Some AI programs have guardrails in place that prevent it from responding to requests to generate a specific artwork, licensed character or recognizable celebrity, while others, including X’s Grok, are known for having fewer restrictions.

Gemini 2.0 Flash is multimodal, which means that the bot can accept more than one type of input, like text with a photo. That allows users to give the AI an image and ask for a change. Making small adjustments to an image while keeping much of the original intact could also potentially lead to copyright infringement, if the user uploads a copyrighted image and much of the original image remains intact.

However, the ease at which the AI is able to remove those watermarks should serve as another warning to photographers. Watermarks serve as a reminder to viewers that the image is copyrighted, but watermarks aren’t foolproof. Removing the watermark from a copyrighted image is considered copyright infringement in many countries, including the US and the UK.

You may also like

Read our guide to protecting images from AI, or learn how to spot an AI-generated image.