Launching today
ImageTruth AI
5 AI models analyze images for AI generation evidence
4 followers
5 AI models analyze images for AI generation evidence
4 followers
Most AI image detectors run one model and hand you a verdict. ImageTruth AI runs every image through 5 independent models — Winston AI, SightEngine, AI or Not, Hive, and Gemini — and shows you the exact findings from each one. No verdict. No single-model guess. Just the full evidence, yours to interpret. Hive alone checks 107 generator signatures. For professionals who need to present and defend their conclusions, that difference is significant.
I've spent my whole life creating art (drawing, painting, making volcano cakes, etc.) and believing that a lifelong conviction to Beauty, Goodness, and Truth is essential. I’ve worked 20+ years as a graphic designer in visual communication — often thinking about how images communicate, persuade, and sometimes deceive. So when some AI-generated images started becoming genuinely hard to distinguish from real photographs or human-created digital art, it was clear this would lead in some cases to significant problems. Visual truth sets a foundation that decisions are based on.
It occurred to me to use AI itself to help identify images it created.
As I began developing my app, I learned that AI image detection was much more difficult than expected, and that various AI models look for different indicators. A bit more alarming was all the models I came across were declaring very confident yes or no verdicts. They may often be correct, but sometimes they are not — it's one model's guess. And if you're a journalist about to publish a story, or an investigator, or someone in the legal system making a decision about a real person — acting on a single guess with little behind it felt genuinely risky to me. What happens when you need to explain your reasoning? What do you point to? Wouldn’t it be great to back up your experienced gut feeling on an image with some hard evidence?
This changed my original intent of giving a single yes/no verdict, and clarified for me the real value my app could provide. What if instead you just showed people the details of what each model actually found? Five independent models, full breakdown, no conclusion drawn for you. You decide what decision is warranted.
Building it taught me other things I didn't expect. At one point I had a heatmap feature — a colored overlay showing where manipulation might have occurred. It looked impressive. I pulled it before launch because it wasn't based on real pixel-level analysis. It would have implied something the data couldn't actually support. Removing it felt like the right call, even though it hurt a little.
It's live, it's free to try, and I'm genuinely curious what people think. Happy to answer anything in the comments.
... and a big thank you to Zach James who took the time and made the effort to review, critique and test my app during development.