Google Gemini's SynthID Watermark Detector Shows Inconsistent Results in Chat Sessions
Key Takeaways
- ▸Gemini's SynthID detector appears to repeat the detection result from the first uploaded file in a chat session, even when new files are uploaded and the same question is repeated
- ▸The bug undermines the reliability of a critical tool used by fact-checkers and journalists to verify AI-generated content, creating potential for false positives or false negatives
- ▸Lead Stories recommends users exercise caution when using Gemini for SynthID watermark verification and suggests waiting for Google to address the issue before relying on detection results
Summary
Lead Stories discovered a significant reliability issue with Google Gemini's SynthID AI watermark detector during fact-checking work. The detector, which identifies content generated or edited with Google AI tools by recognizing imperceptible digital watermarks, was found to return inconsistent results for the same files within a single chat session.
Through systematic testing with multiple chats, researchers found that Gemini's SynthID detector appears to return the detection result from the first image or video uploaded in a chat session, regardless of whether new files are subsequently uploaded and analyzed. For example, the detector initially identified a video as having a SynthID watermark in one chat, but in separate chats it claimed the absence of a watermark for the same file—directly contradicting its earlier analysis.
The issue is particularly concerning because SynthID detection is widely relied upon by journalists and fact-checkers as a critical tool for verifying whether content has been generated or modified using AI. Google currently directs users to Gemini for watermark verification, with its professional verification portal limited to credentialed media professionals on a waitlist. Lead Stories has reported the issue through Gemini's feedback tool and is attempting to contact Google's engineering team directly to address the inconsistency.
