If you have ever found your distinctive brushwork showing up in an AI image generator's output, you already know the uneasy feeling. Your style took years to develop, and now a text prompt can summon something close to it in seconds. The good news is that you are not powerless here, even though no single tool will fully lock your work away from every scraper and dataset builder out there.

This guide walks through what "training" on your images actually involves, which metadata and watermarking tools exist today, the opt-out registries worth signing up for, platform settings that genuinely help, and the honest limits of all of it. Think of this less as a fortress and more as a set of locks on different doors. Some will slow people down. A few might stop casual scraping entirely. None will stop a determined bad actor with command-line skills, and anyone telling you otherwise is overselling.

What "Training" Actually Means for Your Images

When people talk about AI "training" on artwork, they usually mean one of two different processes, and the distinction matters for how you protect yourself. The first is large-scale dataset scraping, where a company or research group crawls publicly accessible web pages and downloads images along with any text near them, like alt tags or captions. LAION-5B, one of the most widely used training datasets, was built this way from billions of image-text pairs pulled off the open web.

The second process is fine-tuning, where someone takes an existing model, often an open-source one like Stable Diffusion, and trains it further on a smaller, curated set of images, sometimes just 20 to 50 pieces from a single artist. This is how "style LoRAs" get made, and it requires far less effort than building a dataset from scratch. Someone can right-click save a few dozen images from your portfolio site in an afternoon and have a usable style model by evening.

Knowing which scenario you are worried about changes your approach. Opt-out registries and metadata tags target the first scenario, the broad crawlers. They do almost nothing against the second scenario, because a person manually downloading your images never checks a registry or reads your metadata, they just save the file. That is why platform settings and access friction, covered later in this guide, matter just as much as the technical tagging methods.

Metadata and Watermarking Options

Metadata tags are small pieces of machine-readable text embedded in an image file that signal how it should or shouldn't be used. The most relevant one right now is the "do not train" field defined by groups like the Content Authenticity Initiative and supported in tools such as Adobe's Content Credentials. When you export an image with this tag through Photoshop or Lightroom, it travels inside the file's metadata and, in theory, any scraper that respects the standard will skip it.

There is also glaze and Nightshade, two research tools from the University of Chicago's SAND Lab. Glaze subtly alters pixel patterns so that if a model does train on the image, it learns a distorted version of your style rather than an accurate one. Nightshade goes further and tries to actively corrupt the model's understanding of a concept, for example making a model associate "dog" with cat-like features if enough poisoned images get swept into a dataset. Both are free downloads, both require you to run every image through the software before posting, and both add processing time, typically a few minutes per image depending on your computer.

Visible watermarks still have a place too, even though they feel old-fashioned. A watermark placed across a focal point of the image, not just in a corner, is harder for someone to crop out cleanly. Some artists layer a semi-transparent diagonal text watermark at around 15 to 20 percent opacity across the center third of the piece, which keeps the work viewable while making clean extraction more work than it is worth for casual scrapers.

Opt-Out Registries Worth Knowing About

A handful of organizations now maintain lists of artists who have asked not to be included in AI training sets, and some AI companies have agreed to check these lists before scraping or to remove flagged content after the fact. Spawning.ai runs "Have I Been Trained," a searchable database where you can look up your own images by uploading a sample or searching by filename, and then flag matches for removal from certain datasets, including LAION derivatives.

The European Union's AI Act, which started phasing in obligations through 2025 and into 2026, includes provisions requiring companies training general-purpose AI models to respect machine-readable opt-out signals under existing copyright exceptions. This is a developing legal area, and enforcement mechanisms are still being worked out, so if you are relying on this for a specific dispute, a conversation with an intellectual property attorney familiar with EU law is worth the fee before you assume it covers your situation.

In the United States, there is no equivalent federal registry yet, though the Copyright Office has been collecting public comments on AI training and copyright since 2023. Some individual companies, including Adobe and Shutterstock, run their own internal opt-out systems for content uploaded to their platforms specifically, which is a narrower but more directly enforceable option if you already sell through them.

Signing up for these registries takes maybe 20 minutes total and costs nothing, so there is little reason not to do it even if the practical effect is uncertain. Just keep your expectations calibrated: a registry listing is a signal, not a legal shield, and it relies on the scraping party choosing to honor it.

Platform Settings That Help

Where you post your work matters almost as much as how you tag it. Some platforms have built opt-out toggles directly into account settings, and turning these on takes two or three minutes.

• DeviantArt: has an account-wide setting under Privacy & Safety called "AI Opt-Out" that blocks the platform's own licensed training partners from using your uploads, though it does not stop third-party scrapers who ignore the site's robots.txt file.
• ArtStation: added a "NoAI" tag option that can be applied per-project or site-wide, which also embeds a machine-readable flag in the page.
• Instagram and Facebook (Meta): Meta's data use settings include an option to object to your content being used for AI training, found under Privacy Center, though you have to actively submit the objection form rather than toggle a switch, and Meta has stated it still reserves rights in some regions depending on local law.
• Personal portfolio sites: adding a line to your robots.txt file specifically disallowing known AI crawler user-agents, like GPTBot or CCBot, blocks compliant bots from even accessing the images in the first place.

That last option, the robots.txt edit, is probably the single most effective technical step available to you if you host your own site, because it works at the server level before an image is ever downloaded. It takes about 10 minutes if you are comfortable editing a text file, or you can ask your web host's support chat to do it for you.

What These Measures Can't Do

It is worth being blunt here. None of these tools stop someone from taking a screenshot of your work on their phone and feeding that image into a training pipeline manually. Screenshots strip metadata automatically, so a "do not train" tag embedded in the original file is gone the moment someone captures the screen.

Glaze and Nightshade are also not invisible or permanent solutions. Independent researchers have published papers demonstrating that certain image processing techniques, like specific denoising or upscaling passes, can reduce or remove the glazing effect, and the SAND Lab itself updates the tools periodically to stay ahead of these countermeasures, which means older glazed images may become more vulnerable over time as detection methods improve.

Opt-out registries only work if the company doing the scraping agrees to check them, and plenty of smaller AI startups, hobbyist model trainers, and operations based outside jurisdictions with enforceable AI regulation simply do not check. There is also no way to retroactively pull your images out of a dataset that has already been used to train a released model, since the model has already learned whatever patterns it learned. Removal from a dataset list does not undo training that already happened.

Common mistakes

A few patterns show up repeatedly among artists trying to protect their work, and they are worth naming directly.

• Assuming a watermark in the corner is enough. It is trivial to crop out and barely slows anyone down.
• Applying Glaze or Nightshade once and assuming the protection is permanent, when in fact every new image needs separate processing before upload.
• Posting full-resolution files when a smaller export, say 1500 pixels on the long edge instead of 4000, would show the work just as well online while giving a model less detail to learn from.
• Treating a single opt-out registry signup as complete coverage, rather than one layer among several.

A Simple Monthly Habit

Rather than trying to do everything at once, build a short routine you actually repeat. Once a month, spend 20 minutes on three things: check whether any new work you posted needs glazing and metadata tagging before it stays public, search your name or a signature piece on Have I Been Trained to see if new matches surfaced, and glance at your platform privacy settings in case a site has added or changed an opt-out option since your last visit.

Keep a simple note, even just a text file, listing which images you have already processed with Glaze or Nightshade and which export resolution you used, so you are not guessing six months from now. If a specific infringement case comes up, meaning you find a close derivative of your work being sold commercially, that is a different and more serious situation than general training concerns, and it is the point where talking to an attorney who handles copyright or AI-related disputes makes sense rather than relying on these preventive tools alone.