Gemini's AI detection only recognizes its own kind

Google Gemini has launched a new capability: directly using natural language to check whether a file was generated by AI. You just need to ask it “Is this AI generated?”, and Gemini will invoke SynthID detection to determine whether images, videos, and audio contain Google's own AI watermark.

The official demo is a black-and-white security-camera clip: a raccoon sits on a small blue plastic chair, leaning over a yellow table and eating, with a porch in the background, and a date and the words “FRONT DOOR” in the corner of the frame. It looks like home security footage, but it is indeed AI-generated. Gemini can identify it.

SynthID is Google's AI watermarking technology. It is embedded in images, videos, and audio as a machine-readable marker, not a logo visible to the naked eye. This feature sounds practical, but there is a key boundary: it can only confirm “whether this was made by Google AI.” AI-generated content from other companies is outside its detection scope. People in the comments quickly seized on this.

“It can only find Google's own watermark, so if Gemini says no, that proves nothing.” (@Dusan)

“A clean result is closer to ‘not from Google’ than ‘not AI.’” (@Aapakari)

“What if SynthID is removed?” (@The Rich Prince)

“It's best to assume everything is AI-generated and be shocked when you find out it isn't.” (@Tony Simons)

One commenter made an analogy: “AI detector, Google edition.” (@Harjot Rana)

Someone else mentioned a more open approach: OpenAI and Adobe are already using C2PA credentials to sign content, a digital signature standard that records content provenance. If Gemini also read C2PA, a single query could cover more sources. (@Bojodanchev)

There were also relatively positive voices: being able to check Google's AI watermark is still useful, especially with AI content flooding the internet, but they wondered whether it still works on content after editing or reposting. (@Wren Castellan)

Another commenter joked that when Gemini answers “Is this AI?”, it might just say, “Yes, I made it,” a quintessential 2026 scenario. (@NusoMaus)

As for AI detection itself, some were pessimistic: it won't be long before AI itself can no longer recognize AI-generated content. (@Aykut Karatay)

Someone else spotted a new need in the comments: they suggested using the same capability to detect bug reports, and said triage teams would pay to know which ones were written by models. (@Marius du Preez)

The comment section gradually went off the rails. One person just said “oh no” and left an image; another simply posted an image with no text at all.

The greatest value of an AI detection feature after launch may be that it let the comment section show just how clear-headed humans are about AI. The problem SynthID can solve is much narrower than advertised.