Google Gemini's SynthID AI Watermark Detector Appears to Mix Up Results in Chat (leadstories.com)

Recent testing of Google Gemini's SynthID AI watermark detector uncovered a significant flaw in its functionality. Users reported that the chatbot provided inconsistent results when asked about the presence of a SynthID watermark in uploaded videos or images. Initially, it might indicate that a watermark is present, only to later assert that none exists, irrespective of the file being queried. The problem appears to stem from a design issue where Gemini reverts to the detection outcome from the first file uploaded in the chat session, rather than accurately assessing new uploads. This inconsistency is crucial for the AI/ML community, especially for professionals like journalists and fact-checkers who rely on SynthID watermarks to verify the authenticity of AI-generated content. The technology embeds imperceptible digital markers within AI-generated media, which are intended to help users confirm whether content has been created or altered using AI tools. As other companies adopt similar watermarking techniques, the reliability of Gemini's SynthID detector is called into question, emphasizing the need for robust verification tools in the age of AI-generated content and its potential misinformation. Until the issue is resolved, it is advised to start a new chat for each detection query, ensuring a more accurate assessment of uploaded media.