WhatsApp Removes AI Watermark: SynthID Flaw Explained

WhatsApp removes AI watermark information unintentionally when users forward AI-generated photos, according to an investigation by Dutch news broadcaster BNR. The messaging app's image compression strips Google's SynthID invisible watermark, leaving AI-made images undetectable and raising fresh concerns about EU transparency rules that took effect on August 2, 2026.

Under the EU AI Act transparency requirements, developers such as Google, OpenAI, and Nvidia must label AI-generated content with at least an invisible watermark. SynthID, developed by Google DeepMind, is the most widely used detection system and is embedded in Gemini, ChatGPT, and other models. However, BNR's testing shows that sending an AI image through WhatsApp removes the very signal that SynthID relies on.

How WhatsApp compression breaks invisible AI watermarks

BNR journalist Niels Kooloos created two nearly identical portraits: one original photograph and one reproduced by Google's Gemini AI model. When asked directly, Gemini correctly identified the AI-edited image. But after the same AI image was sent through WhatsApp and resubmitted, Gemini responded: "There are no reliable signals or data found that indicate how this image was created or whether it was made with Google AI."

The cause appears to be WhatsApp's compression algorithm, which reduces file size by discarding pixel data. Those discarded pixels contain the subtle frequency patterns that SynthID uses to verify AI content. As a result, detection tools can no longer tell whether an image is authentic or AI-generated.

WhatsApp vs Signal: watermark detection comparison

BNR found that Signal's image compression does not break the watermark: photos sent through Signal could still be recognized by SynthID afterwards. WhatsApp users can send images as HD files to reduce compression, but AI images from ChatGPT and Gemini often cannot be sent as HD by default because of their standard resolutions. Users must explicitly request a higher-resolution AI image and manually select HD before sending for SynthID to still work.

Deepfake risk grows as AI detection fails

The timing is significant. Deepfake fraud attempts at financial institutions rose by 2,137% over three years, according to identity solutions provider Signicat. Separate research by Dutch programme Pointer found that more than 40 doctors and celebrities were victims of deepfake videos promoting supplements on TikTok. As deepfake fraud detection tools struggle to keep pace, the loss of invisible watermarks in everyday messaging could make verification even harder.

Google said last week that users can now choose to remove visible watermarks from AI-generated images, videos, and songs. Josh Woodward, Google's VP for Gemini, wrote on X: "We are seeking a balance between creative freedom and safety: while visible watermarks are now optional, invisible SynthID watermarks and C2PA metadata are still used for transparency." Yet BNR's test suggests that invisible transparency can be undone by simply forwarding an image on WhatsApp.

FAQ: WhatsApp AI watermark removal

Why does WhatsApp remove AI watermarks?

WhatsApp compresses photos to save storage and bandwidth. This compression discards pixel-level changes that SynthID uses to embed and detect invisible AI watermarks.

Can I prevent WhatsApp from removing the SynthID watermark?

You can try sending images as HD files, which reduces compression. However, AI-generated images from ChatGPT and Gemini often need to be generated at a higher resolution first, and you must manually select HD before sending.

Does Signal preserve AI watermarks?

Yes. BNR's testing showed that photos sent through Signal could still be detected by SynthID, even though Signal also compresses images.

Is Google aware of the WhatsApp watermark problem?

BNR contacted Meta, WhatsApp's parent company, for comment but had not received a response at the time of publication. Google has not publicly commented on WhatsApp-specific watermark removal.