What is Google's SynthID watermark?

SynthID embeds an invisible signature into Gemini-generated images via consistent noise patterns in the frequency domain. Supposed to prove AI origin without altering visuals. Cracked by averaging 200 samples.

How was SynthID cracked?

Researchers collected 200 Gemini images, averaged noise to isolate the watermark's signature, then inverted it—dropping phase coherence 91% and energy 75%. Basic stats, no black magic.

Will AI watermarks get better?

Maybe randomized ones, but detection suffers. Behavioral tracking—logs of actions—is tougher to fake and already happening in agent systems like Commit.

🤖 AI & Machine Learning

Google's SynthID Watermark Cracked in 200 Images: The AI Trust Mirage Shatters

Averaging noise from 200 Gemini images stripped Google's SynthID watermark, slashing phase coherence by 91%. So much for unremovable proof of AI origin.

theAIcatchup Apr 10, 2026 4 min read

Visualization of cracked SynthID watermark from averaged Gemini images

⚡ Key Takeaways

SynthID cracked via averaging 200 images, exposing systematic signals as weak for AI trust. 𝕏
Shift from origin watermarks to behavioral telemetry: logs of actions can't be statistically removed. 𝕏
Historical parallel to DRM failures predicts quick demise for similar AI proofs. 𝕏

Published by

theAIcatchup

Community-driven. Code-first.

#AI trust #AI watermark #Google Gemini #SynthID #behavioral telemetry

Worth sharing?

Get the best Open Source stories of the week in your inbox — no noise, no spam.

Originally reported by Dev.to

⚡ Key Takeaways

The 60-Second TL;DR

theAIcatchup

Share this article

Worth sharing?

Related Stories

Two AIs 'Recognize' Each Other: String Matches and Existential Glitches

FoolPot: The AI Teapot That Grills You Before Serving a 418

I Tested 12 Free LLM APIs in 2026: The Ones That Haven't Vanished Yet

Google Drops Prompt Bible: Finally, a Map Out of LLM Chaos

Stay in the loop