The tool also recognises watermarks from Google's SynthID partners, so it can flag media made with AI models from OpenAI, NVIDIA and Kakao. Apple participates in SynthID too, but detector support for its creations is still in the works. That is a step up from the old options: previously you could only ask Gemini itself whether something was AI generated, which was never going to satisfy the sceptical, and partner watermarks were out of reach entirely.

Because it looks for a watermark rather than analysing patterns, the detector should avoid false positives on human-made work. The trade-off is that it cannot spot content from tools that skip SynthID, such as Meta's, and it only says whether a label was found, not which parts of a file were AI generated. Google says the signatures survive edits, though there is no guarantee a determined faker will never find a way around that.

It is not the only game in town. OpenAI runs its own detector, Meta offers one for its Content Seal watermark, though a 2026 Reuters report found it sometimes missed them, and Anthropic and Microsoft watermark media without offering consumer detection. No reliable all-purpose AI detector exists, and the general-purpose ones that are public are unreliable and prone to flagging human work. Within its limits, SynthID Detector is likely one of the more dependable options.