No single tool stops an AI model from learning from your images. What exists instead is a stack of overlapping defenses: a technical filter that makes style copying harder, a legal reservation that in the European Union forces anyone mining your work to seek permission if you express it correctly, and each platform’s own opt-out settings. Together they reduce the risk. None of them eliminates it.
This matters for illustrators, photographers, designers and anyone who publishes original work online. Below is what Glaze and Nightshade actually do, what Article 4 of the EU copyright directive says, how to record your reservation in a format machines can read, and where the gaps remain.
- Technical layer: Glaze and Nightshade inject near-invisible distortions that confuse AI models during training, but published attacks such as LightShed can strip them.
- Legal layer: Article 4.3 of Directive 2019/790 lets rightsholders opt out of text and data mining, provided the reservation is machine-readable for online content.
- Enforcement layer: The EU AI Act obliges general-purpose model providers to respect that reservation, with obligations applying since 2 August 2025.
What Glaze and Nightshade actually do
Both tools add almost imperceptible distortions to images, designed to confuse AI models during training. Glaze defends your style; Nightshade attempts to punish anyone who trains without permission.
They come from the Glaze Project, a research group of professors and doctoral students at the University of Chicago. The project describes Glaze as a tool to make artistic style harder to imitate, and Nightshade as a tool to disincentivize training on images scraped from the web without consent. Both are free for artists, and the team also offers WebGlaze, a free web service for those without a sufficiently powerful computer.
The University of Cambridge summarized the distinction in a June 2025 note on the weaknesses of these tools. Glaze takes a passive approach: it hampers the model’s ability to extract the features of your style. Nightshade goes further and corrupts the learning process, making the model associate your style with unrelated concepts. According to that same note, the two tools had accumulated nearly nine million downloads.
Do they really work?
They work against training pipelines that make no effort to look for them, but they are not a padlock. Published methods exist to detect and remove the protection.
The best known is LightShed, developed by researchers at the University of Cambridge, the Technical University of Darmstadt and the University of Texas at San Antonio, with a presentation planned at the USENIX Security conference in August 2025. According to Cambridge’s note, LightShed works in three steps: it detects whether an image carries one of these known “poison layers,” learns what they look like from public examples, and removes it to return the image to its original state. The authors published it to warn creators that these tools do not reliably prevent training.
The practical reading is that Glaze and Nightshade are an added cost for anyone who wants your work, not a barrier. And they carry a limit no update can fix: they only protect copies you process from now on. Work already published unprotected and already downloaded remains outside your control.
The Article 4 reservation and its machine-readable requirement
Article 4 of Directive (EU) 2019/790, dated 17 April 2019, requires member states to allow reproductions of lawfully accessible works for text and data mining. But paragraph 4.3 attaches a condition: the exception applies only if the rightsholder has not expressly reserved that use “in an appropriate manner, such as machine-readable means in the case of content made publicly available online.” Recital 18 specifies that such means include metadata and the terms of use of a website or service. One exception cannot be vetoed: mining for scientific research by research organizations and cultural heritage institutions, governed by Article 3.
In Spain, the rule sits in Article 67 of Real Decreto-ley 24/2021. Its third paragraph states that the exception does not apply when rightsholders have expressly reserved the use of their works “through machine-readable means or other suitable means.”
The piece that gives it teeth is the EU AI Act. Article 53.1.c obliges providers of general-purpose AI models to have guidelines for complying with copyright, and in particular for detecting and respecting the Article 4.3 reservation. Those obligations have applied since 2 August 2025, and the relevant recital extends them to any provider marketing a model in the EU, regardless of where it was trained.
How to record your reservation so machines can read it
On your own website you control robots.txt and the terms of use. On a social network, the platform’s settings prevail.
On your site or portfolio, three things matter:
- Terms of use with an express reservation clause against text and data mining, citing Article 4.3 of Directive 2019/790 and Article 67.3 of Real Decreto-ley 24/2021. Recital 18 names such terms as a valid means.
- A robots.txt file listing the AI agents you want to block. Google, for example, documents the Google-Extended token in its list of common crawlers. Blocking it controls whether what Google crawls on your site is used to train future generations of Gemini models and for grounding its answers. Google says this does not affect your presence or ranking in Search.
- Watching AIPREF, a common, machine-readable preference signal. Google indicates on that same page, updated 3 November 2026, that it expects to support it with the user-agent: __usage_ai-training directive in early 2027.
On Instagram, Facebook, LinkedIn or X you cannot touch robots.txt. What remains are each company’s objection forms and settings. A line in your bio saying “no AI training” serves as a declaration of intent, but for online content the directive calls for machine-readable means. On its own, it is the weakest layer.
The practical approach is to combine layers: Glaze or Nightshade on everything you publish from now on, a written and technical reservation on your own site, an objection filed on each network, and high-resolution files only for paying clients. But there is no point pretending the burden is evenly shared. The European system runs on an opt-out logic: your work can be mined unless you say no, in the correct format. That leaves the work to those with the fewest resources. The AI Act moves things forward by obliging large providers to respect the reservation, but its effectiveness will depend on someone checking that they do.