Security researcher Dr. Neal Krawetz has published findings showing Meta's Stable Signature invisible watermark algorithm fails to meet its claimed accuracy standards, marking the third major AI watermarking system to underperform in independent testing.

Stable Signature encodes a 48-bit sequence into image content, allowing creators to embed unique watermarks that should survive image compression and editing. The system is open-source and available on GitHub, unlike Google's proprietary SynthID.

Krawetz's analysis follows his previous evaluations of Google's SynthID and Adobe's TrustMark systems, both of which showed significant gaps between claimed and actual performance. SynthID claimed a 99.97% true positive rate but achieved closer to 95% in real-world testing. Adobe's TrustMark showed 10-20% false positive rates despite claims of 96% bit accuracy.

The fundamental flaw

According to Krawetz, all three systems make the same "fundamental mistake" in their approach to invisible watermarking. Traditional watermarks hide in subtle locations like least significant bits or frequency spectrums, but AI-based systems appear to have inherent reliability issues.

The researcher noted that modern AI watermarking algorithms "are all making the same fundamental mistake," though he did not specify the exact technical flaw in the published excerpt.

Meta's system differs from competitors by being fully open-source, allowing independent verification of its claims. The algorithm embeds watermarks in visual content rather than metadata, making them theoretically more resistant to removal.

Krawetz's investigation into invisible watermark algorithms highlights a broader problem with AI-generated content detection. As deepfakes and AI-generated media become more sophisticated, reliable watermarking becomes crucial for content authenticity.

The findings suggest that current AI watermarking technology may not be ready for widespread deployment in content verification systems, despite claims from major technology companies about their effectiveness.