It’s getting harder to trust your own eyes online, and that’s not just a feeling—it’s becoming a measurable fact. In 2025, researchers at Microsoft’s AI for Good Lab ran a large-scale experiment where more than 12,500 people were shown a mix of real photographs and images generated by artificial intelligence. The participants were asked to decide which was which, and they managed to get it right only 62 percent of the time. That might sound like a passing grade, but in a world where the truth of a photograph can determine whether a story gets published, whether a student gets credit for original work, or whether someone gets scammed out of money, 62 percent is barely better than a coin flip. And the odds are only going to get worse. AI image generators are improving by the minute, and the gap between what a camera captures and what an algorithm invents is closing fast. The point isn’t that you’re naive if you can’t tell the difference. The point is that no one really can anymore, and the old assumption that a photograph must be real because it looks real has quietly stopped holding. That’s why tools that can read the image itself, rather than relying on human perception or fragile file metadata, are starting to feel less like a luxury and more like a basic safety net for living in the modern information ecosystem.
The industry’s first instinct was to solve this problem at the source. Tech companies started embedding invisible watermarks and signing records into the files that their AI tools produce, so that anyone who checked the metadata could confirm whether a machine was involved. That’s a clever idea in theory, but it falls apart the moment someone takes a screenshot. A screenshot strips away the hidden layers of a file; so does re-uploading an image to a different platform, or saving a copy from a social feed. Even when the watermark survives, many generators simply don’t add one at all. The images that truly do the most harm tend to arrive with no identifying information attached, looking just like any other jpeg. This is the problem that Phrasly’s new AI image detector was built to solve. It is a free, no-account-needed tool that accepts JPG, PNG, WebP, HEIC, AVIF, and GIF files up to 20 megabytes, and you can get it to analyze an image by uploading it, dragging it into a browser window, pasting it, or providing a direct link. Once you do, the detector gives you three things: a verdict on whether AI was involved, a confidence score that tells you how sure it is about that verdict, and a heatmap that highlights exactly which parts of the image it is reacting to. It checks every image in memory rather than storing it, which is an important privacy detail—you don’t have to worry about your suspicious screenshot being saved on someone’s server forever.
The reason Phrasly took a different approach from watermark-reliant systems is simple: watermarks and metadata come off by accident as often as they are removed on purpose. A screenshot removes them, a re-upload removes them, and a copy saved from a social feed removes them. Many AI generators never add them in the first place, either because the user has turned them off or because the platform that created the image hasn’t bothered to build that feature in. Phrasly decided to solve the problem by doing what humans used to be able to do, just far better: it looks at the picture itself. The company trained its own model on millions of real photographs and AI-generated images from every major generator, including ChatGPT, DALL-E, Gemini, Midjourney, Stable Diffusion, FLUX, and Adobe Firefly. The model learns to recognize the subtle statistical fingerprints that machines leave behind in pixel patterns, lighting, texture gradients, and structural details that the human eye almost never notices consciously. In Phrasly’s own testing, the detector is over 99 percent accurate with or without a watermark, on images from all of those major tools. That accuracy rate is important, because it isolates the model’s performance from the fragile metadata that can be stripped away so easily. As Victor Rijo, the CEO of Phrasly, puts it: “A year ago most people could still spot an AI image if they looked closely enough. That is no longer true, and by the time an image reaches the person who has to make a decision about it, it has been screenshotted and reposted enough times that whatever proof was in the file is gone. We built the detector to read the picture itself, because that is the part that survives.”
When you actually run a scan on Phrasly’s detector, it’s not relying on a single method. It runs three separate checks, each with its own strengths, and then presents the verdict alongside all of them so you can see exactly why it came to the conclusion it did. The first check is Phrasly’s own proprietary model, which reads the image directly. This is the check that still works once everything else has been stripped away, and it is the foundation of the tool’s reliability. The second check looks for SynthID, an invisible watermark that was developed by Google DeepMind and is applied by both Google and OpenAI to images produced by their tools. When SynthID is present, the verdict is essentially conclusive, because it’s a cryptographic marker that was embedded at the moment of generation and can’t be perceived by the naked eye. The third check examines AI signatures. Certain generators, including ChatGPT, Gemini, Adobe Firefly, Copilot, and FLUX, sign the images they produce with a digital certificate, and Phrasly verifies that signature against the official C2PA trust list. C2PA is a widely recognized standard for verifying the origin and history of digital content, and if the signature matches a known AI generator, that’s another strong signal. If the file carries no watermark and no signature, the result rests on Phrasly’s own model alone, and the confidence score reflects how certain the model is with the limited information it has. Every check appears beside the verdict, so whether you’re a journalist vetting a photo from a breaking news wire or a teacher looking at a suspiciously polished art assignment, you can see what kind of answer you’re getting and decide how much weight to give it.
The harder case—and honestly, the more common one in real life—is the image that is only partly generated. Most pictures circulating online today weren’t created from nothing. They’re real photographs that have been altered in one specific area: a background swapped from a beach to a war zone, an object added to a scene that was never there, a face retouched to make someone look younger, thinner, or entirely different. In these situations, a simple verdict on the whole image is almost useless, because the honest answer is both. The image is a real photo, and it’s also partly fake. Phrasly’s detector handles this problem by returning a heatmap that marks the exact area that was generated, so you can look straight at the part that needs scrutiny. This is incredibly valuable in practice. A photo of two people at a protest might be completely real except for one face in the crowd that was added by AI; a product listing on a marketplace might show a real box of electronics with an entirely fabricated logo. The heatmap allows a human reviewer to focus their attention where it matters, instead of having to make an all-or-nothing judgment call. The same model also handles screenshots and re-uploads well, which are the exact cases that defeat traditional watermark detection. These files have lost their metadata entirely, so there’s no chance of finding a C2PA signature or a SynthID watermark inside them—and yet the underlying visual evidence is still there, waiting to be read.
Ultimately, Phrasly built this tool for anyone who needs to know an image is real before they act on it. Journalists checking a photograph before it runs in a story. Teachers reviewing artwork that was submitted as original work. Moderators clearing listings and profile pictures on platforms that are flooded with fake content. Buyers trying to decide whether the product in a photo actually exists before they hand over their payment details. These people all share the same kind of risk, and a wrong call is not just a minor embarrassment. It means publishing something false that could damage a reputation, grading work that was never actually done, letting a scam pass through a trusted platform, or paying for something that was never there in the first place. At the far end of the scale, it means allowing a fabricated image to spread as evidence during a breaking news event, standing behind an accusation against someone who was never at the scene, or powering a financial scam that works precisely because a photograph used to be the way you settled a question. In a world where a single manipulated image can go viral worldwide within minutes, the ability to check provenance isn’t just a feature for tech enthusiasts. It’s becoming a necessary layer of trust between what we see and what we decide to believe. Tools like Phrasly’s detector don’t promise to end misinformation overnight, or to catch every single fake that crosses their path. But they do something essential: they restore a little bit of confidence that even when the file loses its invisible markers, and even when your eyes can’t tell the difference anymore, there’s still a way to ask the image itself whether it’s really what it appears to be.