How to protect your art from AI training

No single tool can guarantee your art will never end up in a training set. The honest framing is harm reduction: a stack of defences that make your work harder to scrape, signal that you have opted out, and corrupt the value of any image taken without permission. Used together, these measures meaningfully shift the odds in your favour, even though none is absolute.

Glaze: cloaking your style

Glaze, built by Ben Zhao’s team at the University of Chicago, makes tiny changes to an image that are barely perceptible to a human viewer but that confuse AI models trying to learn your style. To a model, a Glazed painting looks like a different style than the one your eyes see, which disrupts attempts at style mimicry. It has been downloaded several million times since its release and is free for non-commercial use.

Glaze is a protective cloak, not an attack. It is best applied to every image you post publicly, before upload, so that the version circulating online is the protected one.

Nightshade: poisoning the well

Nightshade, from the same lab, takes the offensive approach. It alters pixels so that a model training on the image learns the wrong association, treating a cow as a handbag, for example. At scale, shaded images degrade the model’s reliability, which raises the cost of scraping unlicensed work. Glaze defends your individual style; Nightshade is a collective deterrent that punishes indiscriminate scraping.

Worth saying plainly: Realform never trains models on your art and never generates from it. Its agents compose your existing, human-made work onto products and run the shop around it. Compose, never generate. The protections in this guide are about defending your work out in the open web, not about anything Realform does with it.

Opting out with Spawning and Have I Been Trained

Spawning AI runs Have I Been Trained, a free tool that lets you search large public datasets like LAION-5B to see whether your images were included. If you find your work, or simply want to declare it off-limits, you can register it in Spawning’s Do Not Train registry. Spawning has said media in that registry were excluded from the training of newer Stable Diffusion versions, so the opt-out has had real effect with cooperating model makers.

• Search your work on Have I Been Trained to understand your current exposure.
• Register images in the Do Not Train registry to record a clear opt-out.
• Use Spawning’s ai.txt protocol on your own site to declare AI training permissions in a machine-readable way.

The important caveat: opt-outs only bind the companies that choose to honour them. They are a request, not a lock. But a documented, dated opt-out also strengthens your position if you ever need to show you never consented.

Blocking AI crawlers with robots.txt and ai.txt

If you publish art on your own website, you control a file called robots.txt that tells crawlers what they may access. You can add directives that ask known AI training bots to stay away. Commonly targeted crawlers include GPTBot, Google-Extended, CCBot, anthropic-ai, and ClaudeBot, among others.

• Add Disallow rules for the AI crawler user-agents you want to exclude in robots.txt.
• Add an ai.txt or llms.txt file, a machine-readable convention for declaring AI permissions separately from search indexing.
• Remember these are voluntary standards. Well-behaved crawlers respect them; others can ignore them, so pair this with the cloaking tools above.

Crucially, blocking AI crawlers in robots.txt is separate from blocking search engines. You can keep your work discoverable on Google while asking AI training bots to keep out, because they use different user-agent names.

Watermarking and posting habits

Watermarking will not stop a determined scraper, but a visible watermark across the composition makes scraped copies less usable and asserts authorship. Combine it with sensible posting habits: upload lower-resolution versions for the public web, keep your high-resolution masters private, and apply Glaze before posting rather than after.

A layered routine that works

• Glaze every public image before upload; add Nightshade where you want an active deterrent.
• Register your work in Spawning’s Do Not Train registry and audit your exposure on Have I Been Trained.
• Configure robots.txt and ai.txt on your own site to ask AI crawlers to stay out.
• Watermark public images and keep masters private and offline.

Treat protection as a routine, not a one-off. New crawlers appear, datasets refresh, and tools update, so revisit your defences every few months. None of this is legal advice, but a consistent, layered approach gives you the strongest practical position, and it pairs naturally with platforms like Realform that are built to keep the art human and untouched by generation in the first place.