Imagine you’re scrolling through social media and you see a photo of a politician smiling with a group of supporters, a guide dog sitting in the sunshine, or a candidate holding posters on the campaign trail. The images look completely ordinary. But then an AI detector tells you they were “made or edited with AI.” Would you believe the tool or your own eyes? This is exactly the kind of dilemma that Google’s new public AI detection tool was supposed to help solve. When Google launched its SynthID Detector to the public earlier this month, it seemed like a breakthrough. Finally, anyone could upload an image and find out whether it was authentic or the product of artificial intelligence. It promised to make the internet a little more trustworthy at a time when fake images are everywhere. But our fact-checking team at Full Fact quickly discovered that the tool wasn’t as reliable as it sounded. In fact, we found five examples of images shared by politicians that were flagged as AI-generated even though the people who posted them insisted they were real and unaltered. When we presented our findings to Google, the company fast-tracked an update to the tool. That update changed the verdicts on those five images from “made or edited with AI” to something much more cautious: “Unsure.” It was a quiet but significant admission that even the best AI detection technology can get things wrong.
This is not to say that tools like SynthID aren’t valuable. They are. SynthID is Google’s invisible digital watermarking system, embedded into AI-generated images, audio, text, and video in ways that humans cannot perceive. It works in partnership with other major AI companies, including OpenAI, Nvidia, and Kakao, and it is designed to help people identify synthetic content before it spreads. At Full Fact, we have been using an early version of this detector for nearly a year, and it has genuinely helped us spot real cases of AI-manipulated media. It has allowed us to catch politicians sharing AI-generated images and to flag content that might otherwise have fooled the public. The version available to media professionals is even more detailed than the public one, offering heatmaps that show exactly where a watermark was detected, along with a confidence rating. In many ways, it felt like having a superpower. But with great power comes great responsibility, and also great scrutiny. As we used the tool more and more, we began to notice something troubling. Occasionally, it would flag an image as AI-generated when there was absolutely no visible sign of manipulation, and when the person who posted the image insisted it was genuine. We didn’t want to jump to conclusions, but the pattern was worth investigating.
So we went back through our recent monitoring of politicians and election candidates and identified five images that raised red flags. The first was an April photo posted by Niamh Salkeld, a Plaid Cymru candidate in Wales who went on to be elected as a member of the Senedd. The image showed her holding posters alongside supporters on the campaign trail. There were no obvious signs of editing, no strange fingers, no distorted faces, no warped backgrounds. Yet the detector said it contained a SynthID watermark. When we spoke to her team, they were adamant: “There is no AI used in this image.” The second was a photo posted by Stephen Thompson, an SNP councillor in Dumfries and Galloway, showing a group of supporters in a car park. Again, the detector flagged it as AI-generated. But Thompson told us he didn’t edit it at all, knowingly, and had simply posted it from his iPhone. The third was an image posted by the account for Liberal Democrat MP Steve Darling’s guide dog, Jennie. Yes, a dog’s Twitter account. The tool claimed the picture had been made or edited with AI, but there was nothing in the image to suggest it was fake. A staff member who ran the account told us they hadn’t used AI or any other editing software to alter any images on Jennie’s Twitter. The fourth was a photo posted in August on the Facebook page of Labour MP Kirsteen Sullivan, depicting an event in a church. The tool flagged it, and Sullivan said no AI was used. The fifth was a picture of Labour MP Nadia Whittome on a tram, also from August, which the tool said was AI-generated. Her team told us absolutely no AI was used to create that image.
We described these as “likely false positives” because we couldn’t completely rule out the possibility that AI had been used inadvertently to make minor edits, such as brightening or cropping. But we had no evidence to suggest that was the case. The images looked real, the people who posted them said they were real, and yet the detector kept insisting otherwise. When we checked the media version of the tool, the results were no more helpful. With Thompson’s image, for example, the tool said it had “high confidence” that AI was used, and it highlighted the trees in the background. With the image of Jennie the guide dog, the watermark was localised to what appeared to be a random patch of grass. That didn’t exactly inspire confidence. So we contacted Google with our findings. A spokesperson told us that using the SynthID Detector without any prior suspicion of AI editing naturally increases the chance of encountering an occasional false positive. That made sense, but it also raised a bigger question: if the tool is meant for public use, shouldn’t it be reliable even when people use it spontaneously, without suspicion? The spokesperson also recommended interpreting the detector’s decision alongside other external signals, such as traditional fact-checking methods. That is exactly what we do, but we worried that everyday users might not know to take that extra step. Google initially told us that our examples had pushed it to fast-track an update to help mitigate false positives and more clearly communicate confidence intervals. On 15 October, just a few days after we first contacted the company, that update was rolled out.
When we checked the five images again after the update, the tool’s classification had changed. Instead of saying the images were made or edited with AI, it now said “Unsure,” with a message explaining that it could not tell whether the image had been made with AI from a SynthID partner. The message even noted that AI might have been used for a very small edit, or that the image might have been tampered with. Google described this “Unsure” classification as a public-facing tier designed to help communicate confidence levels more clearly, signaling when the tool detects ambiguous or faint indicators. It was a welcome improvement, but it didn’t solve everything. Even after the update, we continued to find examples that raised questions. In one case, an image posted by Sinn Féin MP Órfhlaith Begley in September, showing her at a care home, was flagged as “made or edited with Google AI” by the public detector. Begley told us there was no AI used on the image. Google’s spokesperson reiterated the same point: using the tool at scale without prior suspicion increases the likelihood of finding false positives. They also explained the underlying challenge. The threshold for detecting a watermark has to balance two competing risks. If the threshold is too low, the tool will flag things that aren’t actually watermarked, creating panic and misinformation. If the threshold is too high, the tool will miss images that really were made with AI, allowing fake content to spread undetected. It’s a difficult line to walk, and no algorithm can get it right every time.
So what does this mean for the average person? It means AI detection tools are useful, but they are not magic. They should be part of a larger toolkit, not the final word. When we at Full Fact check whether an image is fake, we never rely on SynthID alone. We look for visual clues, we talk to the people involved, we check metadata, and we consider the context in which the image was shared. The same approach should apply to anyone using the public detector. If you upload an image and the tool says it was made with AI, don’t immediately share that conclusion. Ask questions. Does the image show any telltale signs of manipulation? Has the person who posted it responded to concerns? Is there an official source or original version available? Sometimes an AI detector is right, and sometimes it is wrong. The fact that Google moved quickly to update its tool after our investigation is a good sign. It shows that the company is listening and willing to improve. But it also shows that we need to keep pushing for transparency and accountability in AI technology. As more images, videos, and audio clips are generated by artificial intelligence, the line between real and fake will only become harder to draw. Tools like SynthID help us draw that line, but they are not infallible. The next time you see an image that sparks outrage, joy, or disbelief, pause for a moment. Look closely. Think critically. And remember that behind every verdict from an AI detector, there is still a need for human judgment. That is what makes us different from the machines. And that is exactly why this work matters.