Google Gemini has a new feature: Letting users generate AI slop without visible watermarks to identify it.<\/p>\n
The official Twitter account<\/a> for the LLM announced on Friday (news dump, everyone!) that users will soon gain the ability to \u201cdecide if your image, video, and music creations made with Gemini will have visible watermarks.\u201d Now it\u2019s a toggle.<\/p>\n A few months ago, per Android Authority<\/a>, Google VP Josh Woodward polled Gemini users on their most-requested features and found removing the visible watermarks on images generated by Google\u2019s Nano Banana, Gemini\u2019s image-generation model, was in the top 10. It\u2019s hardly a shocker that many AI image generation enthusiasts\u2019 top priority is removing visible evidence of AI usage.<\/p>\n So what, if any, safeguards against generative AI abuse remain in Gemini? Google points to two technologies: C2PA (content credential) metadata and SynthID, an invisible watermark that encodes itself into images, video, and audio.<\/p>\n It\u2019s not like it takes much effort to remove a visible watermark from an image. Just crop or obfuscate it. C2PA, being metadata, is similarly trivial to remove. What\u2019s more, most social media services strip some or all metadata from uploaded content anyways, meaning that anyone looking to spread disinformation doesn\u2019t necessarily even have to take additional steps.<\/p>\n SynthID is more robust. It works by encoding itself into the pixels of an image, and is designed to be resilient enough that it registers even after an image has been cropped, resized, and reduced in quality until it looks like a 1990s AOL Mail attachment. Ars Technica<\/a> tested it, finding that while SynthID becomes less detectable with iterations, the image is \u201cuselessly blurry by the time that happens.\u201d<\/p>\n Still, Ars noted that SynthID (by Gemini\u2019s own admission) is not designed to withstand adversarial attacks. At least two commercial services, Rephrasy and SynthID Bypass, already claim to beat the technology. There\u2019s also open-source tools and how-to guides. One author of one Medium post<\/a> on the subject admitted it took \u201cweeks, 123k image pairs, 200 plain Gemini images and a very specific understanding of spread-spectrum encoding\u201d just to make changes to images that sometimes fooled decoders into missing confidence thresholds.<\/p>\n Any way you slice it, this decision is another little landmark on the road towards total erosion of everyone\u2019s ability to recognize AI-generated images at face value. That\u2019s probably good for people profiting from AI slop, and bad for everyone else.<\/p>\n