According to God of Prompt on X, recent tests show Google’s SynthID watermark can be partially evaded but not fully removed, highlighting limits of watermark-based authenticity strategies (source: God of Prompt, Apr 12, 2026). According to the thread, one researcher extracted SynthID’s spectral fingerprint using 200 black and white images and FFT analysis, confusing the decoder rather than deleting the watermark, with a reported 16% evasion rate for a V2 method and improved but still uncertain V3 results (source: God of Prompt). As reported by the same source, a peer-reviewed University of Waterloo tool, UnMarker, reduced SynthID detection from 100% to about 21% in black-box conditions but required a high-end Nvidia A100 GPU, limiting broad misuse (source: IEEE S&P via God of Prompt). According to the thread, Google’s 2025 paper states the decoder can be updated on the fly, suggesting versioned resilience; however, SynthID only tags Google-generated images, leaving non-Google model outputs undetected, which undermines policy assumptions in the White House commitments and EU AI Act watermarking requirements (source: God of Prompt citing Google and policy docs). Business takeaway: watermarking remains valuable for provenance in Google ecosystems, but enterprises and regulators should pair it with multi-signal provenance frameworks and cross-vendor standards to address gaps in open-source and non-participating platforms (source: God of Prompt).