Google's Gemini SynthID Watermark Detector Is Mixing Up Results Mid-Chat

Google's SynthID detector, marketed as a defense against undetected AI content, has been returning inconsistent results within the same Gemini chat session, testers found. Uploaded images sometimes flip between confidently-AI and confidently-human depending on when in the conversation they are checked. It is an ugly moment for a system positioned as ground truth for content authenticity.

What is going wrong

SynthID embeds and detects statistical watermarks in images, text, and audio produced by Google's models. Testers report that when the same file is submitted twice in a row, the detector sometimes returns different confidence scores and even opposite classifications. Google has acknowledged inconsistency in Gemini's chat surface and is investigating whether the issue is in the detector, the chat pipeline, or both.

Why this matters for provenance policy

Watermarking is a central pillar of most current AI content-authenticity proposals, including drafts circulating in the US and EU. If the reference implementation of the leading watermark is unstable inside its own vendor's product, policymakers will (rightly) push back on any regime that leans on it as a compliance mechanism.

What to use instead, for now

Provenance metadata standards like C2PA remain the more durable answer, because they attach signed content credentials at capture and edit time rather than relying on statistical detection after the fact. Adobe, Nikon, and Sony's C2PA-enabled workflows do not fix watermarking, but they do give newsrooms and platforms a defensible chain of custody.

The bottom line

Watermarking is not dead, but SynthID's stumble is a reminder that content provenance needs layered defenses — capture-time credentials, watermark detection, and forensic analysis — not one silver bullet.