A developer has published a free, open-source tool on GitHub that can partially bypass Google’s SynthID, the invisible watermark embedded in every image generated by Gemini. The tool has garnered over 1,600 stars on GitHub. India’s Information Technology Amendment Rules, 2026 (IT Amendment Rules 2026), which came into force on February 20, require platforms to deploy reasonable and appropriate technical measures to proactively detect and label synthetic content. This compliance framework assumes that detection works reliably. However, this tool raises questions about whether it does.

What is SynthID?: When Gemini generates an image, Google invisibly embeds a hidden pattern into every pixel. This pattern is too subtle for the human eye but detectable by machines. This technology, known as SynthID, was developed by Google DeepMind. According to its research paper, Google has embedded SynthID in over 10 billion images.

The idea is that Google’s detection system can scan an image at any point and confirm that it was AI-generated, even after cropping, screenshotting, or compressing.

What the developer did: Developer aloshdenny noticed that Google uses the same hidden pattern across every single Gemini image. He then:

• Generated 200 plain black and white images from Gemini and averaged them together. Because the images are pure black or white, nothing else interferes. The only element that survived the averaging process was the watermark, the one feature consistent across all images, allowing it to reveal itself.
• Built his own detection tool that identifies SynthID watermarks with 90% accuracy.
• Developed a bypass tool that strips enough of the signal for Google’s own detector to fail to flag the image as AI-generated. To the human eye, the image appears completely identical afterward.
• Published the entire project, including code, methodology, and dataset, in a GitHub repository along with a detailed write-up.

The developer has clarified that he did not fully destroy the watermark. The best he could do was confuse the decoder enough for it to fail. Because the watermark is woven into the image during the generation process itself, complete removal is difficult. However, partial removal is sufficient to defeat detection.

Why this matters: India’s Synthetically Generated Information (SGI) rules, notified under the IT Amendment Rules 2026, require platforms to deploy reasonable and appropriate technical measures to proactively detect synthetic content before it spreads. This obligation kicks in independently of whether a government or court order has been issued.

Platforms that fail to meet this proactive detection standard risk losing safe harbour protection, the legal shield that protects them from liability for user-generated content.

Watermark-based detection is one of the primary technical mechanisms platforms use to meet this obligation. A publicly available bypass, open source and requiring no special technical access, allows anyone to strip the watermark signal from a Gemini image before sharing it. As a result, a platform’s detection system may perceive the image as clean, enabling synthetic content to circulate undetected.

While MeitY’s rules mandate the deployment of such technical measures, they do not specify how platforms should respond when a watermark has been deliberately removed. As of now, MeitY has not publicly acknowledged or addressed this gap.

Also read:

• YouTube’s AI Avatar Tool for Shorts Raises Questions Around India’s Deepfake Rules, Personality Rights
• How Should the Government have Gone About Deepfake Regulation?
• Explainer: Why detecting deepfakes is a challenging problem