Google Launches New Website to Detect AI Fakes, But Don’t Assume It’s Perfect

Google launched SynthID.com on Wednesday, a website where anyone can upload an image or video to check whether it was AI-generated, but the tool only detects content made with tools from Google, OpenAI, Nvidia, and Kakao. Google first introduced SynthID in 2023 as an invisible watermark on images made with its AI image tools and integrated a SynthID checker into Gemini at the end of 2025. Gizmodo testing found the limits of the approach: Gemini falsely claimed a fake Trump-Paxton-Grimace image was an edited March 2024 photo from Trump International Golf Club in West Palm Beach, Florida, and SynthID.com reported only that a Grok-generated image was "unlikely" to have been made with AI from a SynthID partner.

The image above isn’t real. Sure, Ken Paxton and Donald Trump are political allies, and Trump starred in a McDonald’s commercial featuring Grimace back in 2002, so it’s entirely plausible that they all appeared on a golf course together. But it’s fake. And one way to check that is with a new portal from Google. Google launched a new website Wednesday called SynthID.com https://synthid.com/ that allows anyone to upload an image or video and see if it was generated with AI. There’s one big catch: It can only detect fake images made with tools from Google, OpenAI, Nvidia, and Kakao. In 2023, Google first launched SynthID, which adds a watermark to all images created with its AI image tools. The watermark is invisible to the naked eye and survives further manipulation like cropping and general photoshopping. But Google didn’t give consumers access to tools that allowed them to check any given image for themselves. That changed at the end of 2025, when Google integrated a SynthID checker into Gemini. People could check for themselves, though it had a similar problem to the one still facing us. If you ask Gemini or, as of today, upload an image to SynthID.com, it can’t tell you definitively whether any image on the web is AI. These tools can only tell you if they were made by a handful of LLMs like Gemini or ChatGPT. And even Gemini and ChatGPT can get it wrong in some surprising ways. Incredibly, we uploaded the image to Gemini not the new SynthID.com website , and without prompting, it told us: “This image is artificially generated or edited. It depicts Texas Attorney General Ken Paxton and Donald Trump on a golf course with the McDonald’s mascot Grimace inserted between them.” So far, so good. But the next part was wildly off base. “The original photograph was taken in March 2024 at Trump International Golf Club in West Palm Beach, Florida, showing Donald Trump and Ken Paxton, but without the Grimace mascot,” wrote Gemini. As far as we can tell, there are no photos of Ken Paxton and Donald Trump together with a golf course pictured behind them. That part is just false. We asked Gemini to find us the original photo, sans-Grimace, and it pointed us to some Facebook link https://www.facebook.com/LeHuffPost/posts/craignant-que-leur-candidat-ken-paxton-ne-soit-battu-par-le-d%C3%A9mocrate-james-tala/1528120206019870/ that just shows Paxton and Trump shaking hands. Clicking on the link sends you to a French-language HuffPost article that also doesn’t feature an image even close to what we produced with AI. We tried other image generators to create new versions of the fake Trump-Paxton-Grimace image. The one below was made by Grok. It’s obviously a much worse image, giving that plastic sheen that’s so common in horseshit AI image generators. When we asked Grok on an account different from the one we used to generate it whether it was made with AI, the bot responded: “It’s almost certainly created with an image-generation model Flux, Midjourney, etc. .” Interestingly, it didn’t raise the possibility that the image was made with Grok. It’s understandable that Grok may not want to take credit for something so ugly. When we asked SynthID.com about the Grok-generated image, it responded: “SynthID was not detected. It is unlikely that this media was made with AI from a SynthID partner. Note however, that this may have been generated by a non-partner organization, or at a time prior to a partner adopting SynthID.” Notice that it says it’s “unlikely” this image was created with those tools, which is a good hedge but implies it could be mistaken. Meta has its own AI detection software, and a recent Reuters investigation found that it sometimes fails to flag content https://www.reuters.com/business/meta-ai-image-detector-fails-identify-some-its-own-cropped-ai-images-reuters-2026-07-10/ that was created with Meta’s AI tools. Here at Gizmodo, we often turn to AI tools in order to check if they have invisible watermarks that only those tools can hope to find. But they’re not perfect. And we hope the AI companies can get on the same page and expand their cooperation in this area. Because there will be plenty of people who will check SynthID, see that it doesn’t have the watermark, and potentially think it’s real. “This technology was pioneered by Google DeepMind and has been adopted by Nvidia, OpenAI, and Kakao,” Google says on its website. “SynthID is integrated across a growing ecosystem of models and this list continues to grow as more AI organizations adopt SynthID watermarking.” Another downside to this new SynthID site is that you need to log in with a Google account, Apple account, or ChatGPT account. Most people will have at least one of those, but it does create a speed bump that could hamper some people from using the tool. When we recently asked several AI chatbots whether a new image of Elon Musk https://gizmodo.com/if-you-ask-grok-elon-musk-acted-like-a-jackass-at-the-trump-xi-dinner-2000817418 at Donald Trump’s state dinner for Xi Jinping is real, Grok said it was authentic, and Microsoft Copilot said there wasn’t enough evidence to say either way. As you can see below, it’s fake as hell. Just look at Tim Cook’s weird mug. We’ve had similar experiences with other clearly AI images, whether it was Donald Trump kissing his assistant Natalie Harp https://gizmodo.com/that-photo-of-donald-trump-kissing-natalie-harp-has-one-big-problem-2000811962 or Red Bandana Man https://gizmodo.com/never-forget-ai-slop-is-polluting-9-11-history-on-the-internet-as-we-speak-2000810667 saving people on 9/11. Go ahead and use these AI image detectors. They might give you some useful information here and there, provided it has a watermark. Just don’t assume they’re infallible.