Blog · Next steps · Updated 27 October 2026
AI training and your art: what the evidence says about opting out
Want to check your own artwork? Register it free and scan for copies. No credit card needed.
“How do I stop AI from training on my art?” gets a lot of confident answers online. Most of them are one tool or one setting presented as the solution. When you look at what each measure actually does, who has to cooperate for it to work, and what researchers and regulators have said, the picture is more modest and more useful: three different kinds of defence, each with a clear weak point.
General information, not legal advice, checked on 27 October 2026. This area changes quickly, so check the sources at the end for newer developments. TraceMyAssets does not stop AI training; where we fit is described near the end.
Three different goals
“Stopping training” can mean three things, and a measure that helps with one usually does nothing for the others:
- Keep your work out of the data a company collects (opt-outs and crawler blocks).
- Make your work less useful if it is collected anyway (tools that alter the image).
- Make using it unlawful or costly (law, transparency and licensing).
1. Keeping your work out: it depends on who cooperates
The main technical tool is robots.txt, plus newer machine-readable ways of saying “do not use this for AI training”. They work by asking. A crawler that respects the file stays out; a crawler that does not, gets in. That is why the policy side matters.
In the EU, the General-Purpose AI Code of Practice includes a copyright chapter in which signatories commit to respect robots.txt and to identify and comply with appropriate machine-readable protocols for reserving rights. The European Commission ran a consultation (December 2025 to January 2026) to help it publish a list of generally agreed machine-readable opt-out solutions, to be reviewed at least every two years. Commission consultation page.
The weak point: the Code of Practice is something companies choose to sign, so it only binds the ones who do, and it governs future collection, not what a trained model already contains. And a crawler block on your own website does nothing about copies of your art posted on other people's sites, which is where much of it ends up.
2. Making your work less useful: a moving target
Tools such as Glaze and Nightshade change an image in ways that are hard to see, aiming to disturb a model that trains on it. They are free, they were built for artists, and they attracted a lot of attention.
Then researchers tested them. A paper at USENIX Security 2025, LightShed, presents a method that detects images protected by this kind of perturbation and removes the perturbation. For Nightshade the authors report a very high detection rate and say the method generalises across perturbation techniques; the abstract gives no Glaze-specific results. These are the authors' own results. LightShed on USENIX.
The honest reading: this is an arms race. A protection that holds today may not hold against next year's methods, and the tools are updated in response. That does not make them worthless, it makes them a layer that you cannot rely on alone, and one that cannot be applied retroactively to images already collected.
3. Law, transparency and licensing: slow and uneven
- UK: the government had floated an exception for AI training with an opt-out for rights holders. In its March 2026 report it did not adopt that option, citing practical and economic problems with opt-out systems, and kept the status quo while it watches court cases and international developments. It said it would not legislate on training-data transparency yet, and that a review would look at the tools creators can use to control their works online. Fieldfisher summary.
- EU: rights holders can reserve their rights against text and data mining, but only in an appropriate machine-readable way. What counts as appropriate is still being settled, as the Commission consultation shows.
- US: the artists' case Andersen v. Stability AI was the first filed against AI image companies and has moved slowly. A case-tracking report from February 2026 says the summary judgment hearing was moved from 4 November 2026 to 17 February 2027, so a ruling was not due before then when this was written. Case tracking note. Check the docket for later changes.
The weak point is time. None of this changes anything for the work you post this week. It decides what is allowed in the future, and what happens to the data already collected is part of what the courts are still weighing.
Where watermarks and monitoring fit
An invisible watermark does not stop training, and no honest vendor should say it does. In our own tests the mark was removed by resizing, cropping and stronger compression (see what an invisible watermark survives). What a mark, a dated record and monitoring can do is different: they help you find copies of your work on public pages and show that you published first. That helps you act on a copy, not on a training run. See three ways to find copies of your artwork.
What a sensible artist does with this
- Use the crawler blocks and site settings you control, and treat them as a request, not a lock.
- Treat image-altering tools as one layer that may weaken over time. Do not rely on them alone.
- Keep working files and first-publication dates. They help in any dispute, whatever the law turns out to be.
- Follow the places where rules change: the EU list of machine-readable opt-outs, the UK review of creators' tools, and the US court dockets.
- Look for copies of your work, because that is the part you can act on today.
The short checklist is in Can you stop AI from training on your artwork?
What we did not check
We did not test whether any particular company honours robots.txt, and we did not test Glaze or Nightshade ourselves. We report what the sources above say, and we flag that LightShed's numbers are the authors' own. Court dates move, and reports about them can be wrong, so use the docket, not this post, for the current state of a case.
Sources
- LightShed: Defeating Perturbation-based Image Copyright Protections (USENIX Security 2025)
- Fieldfisher: UK government maintains status quo on AI and copyright (March 2026 report)
- European Commission: consultation on protocols for reserving rights from text and data mining
- ChatGPT Is Eating the World: Andersen schedule pushed back
Related reading
Protect your first artwork
Register it, add an invisible watermark and review possible copies. You decide what to do. No credit card needed.