A researcher evaluated Meta's Stable Signature watermarking system and found that its claimed false‑positive rate of less than one in a million is not supported by empirical data, which indicate a collision probability of roughly one in four. The system embeds a 48‑bit sequence using a neural network that selects locations in an image to hide data with minimal visible distortion and uses Hamming distance to identify matches. The researcher collected 10,000 uncurated images from a public source and applied statistical tests from the NIST SP 800‑22 suite. Analysis showed that the 48‑bit outputs were not uniformly distributed, with about 56 % ones and 44 % zeros, producing a chi‑square statistic far above the threshold for randomness. Serial dependence tests also indicated strong correlation between adjacent bits. Clusters of images sharing identical or near‑identical bit sequences were observed, including 25 images with exact matches and a larger cluster of 450 images within six bit differences, representing about 4.5 % of the sample. These patterns contradict the assumption of independent bits used in Meta’s theoretical calculations, leading to a higher false‑positive estimate than claimed. Similar statistical issues were identified in Google’s SynthID and Adobe’s TrustMark watermarks, which also assumed independent bits despite evidence of bias and dependence. The findings suggest that current watermarking algorithms cannot reliably distinguish genuine watermarks from random noise, which could limit their usefulness for fraud detection and compliance with upcoming AI‑content regulations in the EU, China and California. The researcher called for independent verification and noted that the statistical mismatch undermines the credibility of the claimed accuracy rates across all three systems.