July 4, 2026
Marked safe from the watermark
Meta's Un-Stable Signature
Meta’s secret AI stamp is getting roasted as flimsy, easy to dodge, and maybe pointless
TLDR: Testing suggests Meta’s hidden AI image label is far less reliable than advertised, adding to wider doubts about similar tools from Google and Adobe. In the comments, people immediately zeroed in on the real drama: if these marks are easy to dodge but laws may still require them, everyone could end up trusting a broken system.
Meta’s so-called invisible AI signature just got dragged into the same mess as Google and Adobe, and the crowd is not being polite. The investigator behind the write-up says these hidden “proof marks” don’t work anywhere near as well as the companies claim. In plain English: the systems are supposed to quietly tag AI-made images so they can be identified later, but testing suggests they can miss real tags, find fake ones in random noise, and generally act a lot less magical than the marketing promised.
That alone would be juicy, but the comments turned it into a full-blown tech popcorn moment. One of the loudest reactions was basically, “So this can already be dodged?” with a commenter dropping a link and calling it “easily bypassed now.” Ouch. Another hot thread pulled in the EU AI Act, which pushes for labels on AI-made content that can’t be removed — or makes removing them illegal. That sparked the bigger drama: if lawmakers want permanent labels, what happens when the labels are unreliable, easy to strip out, or possibly mixed with other hidden data like user IDs?
The mood was a mix of cynicism, dread, and dark comedy. Readers seemed less shocked that the tech failed than that big companies kept selling it as solid. The unspoken meme running through the whole discussion: the watermark is invisible, and apparently the reliability is too.
Key Points
- •The article compares published claims for Google SynthID, Adobe TrustMark, and Meta Stable Signature with the author’s own empirical testing.
- •It cites Google’s paper as claiming SynthID has a true positive rate above 99.97%, while the author reports much weaker results in testing.
- •It cites Adobe’s materials as claiming TrustMark can exceed 96% bit accuracy under severe noise degradation, while the author reports a 10%–20% false positive rate.
- •The article describes Stable Signature as encoding a 48-bit watermark into image content using AI-based embedding and decoding.
- •The article explains that traditional watermarking methods use techniques such as least significant bits, brightness changes, DCT, and FFT, while newer systems use AI and different noise-handling strategies such as repetition, BCH, and Hamming distance.