Work · 01
untrainable.art
For artists who post their work online and don't want it used as AI training data.
What it does for you
You drop in an image. You tell it what the image shows and what you'd like a model to mistake it for: a painting for a photograph, say, or a watercolour for an oil painting. untrainable.art then adds noise to the image that's meant to be invisible to people, and gives you three views to compare: the original, the poisoned version, and the noise on its own, amplified ten times so you can see it. Then you download the poisoned copy and post that instead.
Turning the strength up makes the noise more visible. Raising the number of iterations makes it take longer. There's also an option to write a "not for AI training" flag into the file's metadata, so the image carries that request in a form software can read.
Illustration only. Nothing here changes your image.
Reading in this tab…
Dots that moved stand in for the noise. On the real image, it's meant to be invisible.
To protect the image for real, take it to untrainable.art. Open untrainable.art ↗
That doesn't look like an image. Try a JPEG, PNG, WebP or GIF.
An image redrawn in dots. The offset dots stand in for the noise untrainable.art adds.
What it claims, and what it doesn't
In the tool's own words: "When AI models train on poisoned images, they learn corrupted associations." It's built on Nightshade, research from the Glaze Project at the University of Chicago, and the tool links to the paper.
We don't claim more than that. Nobody can promise what a particular model, trained by someone else, will do with a particular image.
Runs in your tab
The first time you visit, your browser downloads the model it needs. After that, your image is read, perturbed and handed back inside the browser tab. It isn't uploaded. On Chrome or Edge with WebGPU it runs at full quality. Other browsers fall back to the CPU and use fewer iterations, so it's quicker but the noise is weaker.