The study from detection firm AI or Not tested 205 images generated by Meta AI, comparing the effectiveness of native watermarking against its own content-based API. While both systems performed reliably on untouched files, the results diverged sharply once images were subjected to common edits like cropping or resizing. Meta AI’s native labeling, which relies on embedded signals, saw its detection rate collapse from 98.1% to 35.3% following such modifications. In contrast, the AI or Not model maintained 98% accuracy across the same tampered set.

These findings echo a recent Reuters analysis, which similarly found that Meta’s watermark-based system struggled to verify images after minor alterations. Experts suggest the issue lies in the structural nature of provenance signals: they act as a "birth certificate" for digital files but are easily stripped away by standard workflows like screenshotting or re-encoding. Anatoly Kvitnitsky, CEO of AI or Not, argues that relying solely on metadata creates a security gap that bad actors can easily exploit, particularly during high-stakes periods like election cycles. The consensus among researchers is that robust verification requires a dual approach, combining provenance standards for original content with forensic models that analyze the visual data itself.