How to protect your art from AI scrapers

What AI scrapers actually do

"AI scraper" is a loose label for a few very different things. Understanding which one is touching your work changes what defenses make sense.

• Declared training crawlers like GPTBot, ClaudeBot, CCBot, Google-Extended, and PerplexityBot identify themselves with a user-agent and (mostly) honor robots.txt.
• Retrieval bots fetch pages on-demand to answer a user's query. They often ignore training-only opt-outs because, technically, they aren't training.
• Anonymous and rotating scrapers rent residential IPs, spoof user-agents, and ignore every signal you send. These are the ones you actually have to defend against.

The first two groups respond to polite signals. The third group only responds to friction — the cost of grabbing your file has to be higher than the value of the file itself.

What actually works (and what doesn't)

No tool blocks 100% of AI scrapers. Anyone promising that is lying. Realistic protection stacks small frictions until your work is a worse target than the next thing in the queue.

Protecting images

The state of the art for images is adversarial perturbation: tiny, mostly imperceptible changes to pixel values that confuse the feature extractors most image models rely on. The two most-discussed academic tools are:

• Glaze (University of Chicago) — shifts the style representation a model sees, so training on your work pulls the model toward a misleading style cluster. Best for visual artists worried about style mimicry.
• Nightshade (same group) — a more aggressive "poison" that pushes the model to misclassify subjects. Designed to be used at scale across many artists' work.
• CloakMyWork — a lightweight, browser-side speed bump. It adds imperceptible high-frequency noise to your previews so casual scraping and embedding extraction degrade. Use it for fast, ad-hoc protection on portfolio pieces, social previews, and client mocks where the heavy Glaze pipeline is overkill.

A reasonable stack: Glaze your hero pieces, run everything else through CloakMyWork before upload, and never post a print-quality master where a thumbnail will do.

Protecting writing

Text is harder. Scrapers strip HTML, normalize whitespace, and feed the rest into a tokenizer. The defenses that survive that pipeline are:

• Invisible zero-width characters woven into your prose. Readers and screen readers ignore them; tokenizers treat them as garbage tokens that pollute embeddings and make the passage less useful as training data. CloakMyWork's text cloak does this.
• Homoglyph substitution — swapping a few Latin letters for visually-identical Cyrillic or Greek glyphs. Aggressive; can break copy-paste search.
• Posting full text behind a login — boring but effective. Most scrapers won't authenticate.
• Watermark canaries — unique invented phrases per published copy so you can later prove a model trained on your text.

Site-level defenses

These don't replace per-file protection — they raise the floor for the polite half of the bot population.

User-agent: GPTBot
Disallow: /

User-agent: ClaudeBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: CCBot
Disallow: /

User-agent: PerplexityBot
Disallow: /

• · Turn on Cloudflare's "Block AI bots" toggle if you're behind their proxy. It's free and catches a wider net than robots.txt.
• · Add an ai.txt file mirroring your robots policy — some emerging standards check it.
• · Serve hero images from a separate subdomain with stricter rules.
• · Rate-limit by IP at your CDN; most scrapers fingerprint as bursty.

The legal layer

Technical protection buys you time; legal protection lets you act after the fact. Both matter.

• · Add a visible copyright notice and terms of use stating that automated training is not licensed.
• · In the EU, file a TDM opt-out in a machine-readable form (robots.txt and image metadata) — the EU AI Act gives this real weight.
• · Register key works with your national copyright office where applicable — it's what makes statutory damages possible later.

A realistic checklist

1. Update robots.txt with the AI-bot block above.
2. Turn on your CDN's AI-bot blocking if available.
3. Run every public image through Glaze (hero pieces) or CloakMyWork (everything else).
4. Cloak long-form writing with zero-width characters before publishing.
5. Add a TDM opt-out notice in your site footer and image metadata.
6. Keep print-quality masters off the public web entirely.