How to Protect Your Art From AI Training: What Actually Works
No single tool keeps your art out of AI training, but four layers together make it much harder: cloak the images you post (Glaze, Nightshade), block AI crawlers from your own site (robots.txt), state an opt-out from text and data mining, which AI companies must respect in the EU, and keep a dated record of your originals. The first three reduce how much of your work is collected. The fourth matters when something has already happened: when an AI-generated image copies your work, or someone claims your art is AI-made. None of them can pull your work back out of a model that has already been trained.
This is general information, not legal advice for a specific dispute.
How your art ends up in AI training
Image models are trained on billions of pictures collected by crawlers from public websites: portfolios, social media, shops, blogs. Some datasets are built by third parties (for example the Common Crawl archive or LAION), others directly by AI companies. Once a model is trained, individual images cannot practically be removed from it. So protection is mostly about what is collected next, and about the evidence you keep.
What you can do, compared
| Measure | What it does | Limit |
|---|---|---|
| Glaze | Adds changes almost invisible to people that make a model misread your style | Free; can leave visible artefacts on flat colours; not guaranteed against future models |
| Nightshade | Makes images teach a model the wrong thing if they are trained on | Free; works against training, not against someone copying one image |
| robots.txt on your own site | Tells AI crawlers such as GPTBot, Google-Extended, CCBot and ClaudeBot not to collect your pages | Only for sites you control, and only for crawlers that honour it |
| Platform opt-out settings | Some platforms let you mark work as not for AI training | Depends on the platform; does not cover reposts elsewhere |
| A text and data mining opt-out (EU) | Under Article 4(3) of the EU copyright directive, a rights reservation removes the automatic permission to mine your work; the AI Act requires model providers to respect it | Must be expressed in an appropriate, preferably machine-readable way; enforcement is still developing |
| Dataset opt-out registries (e.g. Spawning's "Do Not Train") | Lets you check if your images are in known datasets and record an opt-out | Voluntary for the companies that use them |
| A qualified timestamp on your originals | Shows your work existed on a date, unchanged, presumed accurate across the EU | Does not stop training; it is evidence for when something goes wrong |
A practical routine
- Before posting: timestamp the original file, or a ZIP of the piece with its sketches and layered file. Keep the full-resolution original offline.
- Post a prepared version: smaller size, run through Glaze or Nightshade if you accept the visual trade-off, with your name or a watermark.
- On your own site: add AI crawlers to robots.txt and put a clear sentence in your terms or footer reserving your works from text and data mining.
- On platforms: turn on any AI-training opt-out and use its "no AI" tags.
- Once a quarter: check dataset search tools and reverse image search for your best-known pieces.
When something has already happened
- An AI image reproduces your actual artwork (not just a similar style): that can be infringement like any other copy. Save it, note the tool and the prompt if you know them, and show your dated original.
- Someone sells AI images "in the style of" you under your name: using your name to sell can be a problem even where style is free to imitate. Report it to the platform.
- You are accused of using AI: layered files, sketches, timelapses and a dated original from before the accusation answer it.
What a StampR certificate adds
A StampR certificate binds the SHA-256 fingerprint of your artwork, or of a ZIP of the piece and its working files, to a qualified electronic timestamp from BORICA B-Trust, together with your declaration that you made it. Under Article 41(2) of the EU eIDAS Regulation, its date and integrity are presumed accurate in every EU member state. That gives you an independent date that predates any copy, AI-generated or not. Only the fingerprint is sent, so the original is never uploaded and can never end up in a dataset through us.
Timestamp your artwork before you post it – the first certificate on an account is €4.90.
What it does not do
- It does not stop crawlers or training. Glaze, Nightshade, robots.txt and opt-outs work on that.
- It does not verify that a work is human-made. It dates a file and your declaration; process files show how it was made.
- It does not cover a style. Style is not protected by copyright for anyone; your actual works are.
- It does not create or record any rights. Copyright arises when you create the work.
The wider context: copyright protection – how the right arises, and where it is weak in practice.
Related reading
- AI copied my art style – what can I do? – where style ends and copying begins.
- How to show my art is not AI – the process evidence that answers the accusation.
Frequently asked questions
Can I stop AI from training on my art?
Not completely. You can reduce it: block AI crawlers on your site, opt out where platforms allow it, state a text and data mining reservation, and cloak posted images with Glaze or Nightshade.
Do Glaze and Nightshade really work?
They make your posted images harder to learn from for current models, but they are not a guarantee, and future models may get around them. Use them as one layer, not the only one.
Is it legal for AI companies to train on my art in the EU?
EU law allows text and data mining of lawfully accessible works unless the rights holder has reserved it in an appropriate way, such as machine-readable means for online content. The AI Act requires general-purpose model providers to respect those reservations.
Can I get my art removed from a model that was already trained?
In practice, no. You can opt out of future datasets and act against specific outputs that copy your work.
Can a certificate prove my art is not AI?
No. It shows that a specific file existed on a date and that you declared yourself its author. Sketches, layered files and timelapses show the human process; the two complement each other.
Sources
- Directive (EU) 2019/790, Article 4 – text and data mining and the rights reservation.
- Regulation (EU) 2024/1689 (AI Act), Article 53(1)(c) – model providers must respect reservations.
- Glaze and Nightshade, University of Chicago – what the tools do and their limits.
- Regulation (EU) No 910/2014 (eIDAS), Article 41 – legal effect of qualified electronic timestamps.
Sources checked 2026-11-05.
Certify your work with a trusted date.
Create a certificate →Your first certificate is €4.90 with an account.