AI image generators and your photo rights in 2026

In 2022, the discussion about AI image generators was hypothetical. In 2026 it's a daily topic for working photographers. Models like Stable Diffusion, Midjourney, and Flux have been trained on datasets with billions of images, a significant portion of them without photographers' permission. What are your rights? What can you actually do? This article lays it out, honest about what's achievable and what isn't yet.

How does your photo end up in an AI model?

Many major image generators are believed to rely on large web-scraped datasets or similar collections of online images. With open models, it's sometimes clearer which datasets played a role; with closed commercial models, that often stays murky. LAION-5B and Common Crawl are the best-known public datasets, both built by mass-scraping the web, including photographers' portfolios, press archives, social media, and commercial stock banks. The dataset designers defend this as "research-only", but the models that train on them get deployed commercially.

How do you know if your work is in there? For LAION there are search tools (haveibeentrained.com, spawning.ai) where you can enter your name or a photo URL. The result is useful insofar as you learn whether a specific photo is in a specific training set. What you don't learn: whether a style the model has "absorbed" can be specifically traced to your work.

The legal status as of 2026

Three ongoing lawsuits set the direction:

Getty Images v Stability AI. The UK High Court issued a first significant ruling in November 2025. Getty largely lost on the broad copyright claims around training, but won in part on points related to trademark use and Getty watermarks reproduced in generated output. The case mostly shows that the legal position is still unsettled.

Andersen et al. v Stability AI (US District Court, Northern California). A group of visual artists alleges Stability "baked in" their style into the model. An early Stability motion to dismiss was largely denied in 2024, the case continues.

Pictoright v Meta. The Dutch collecting society Pictoright is litigating in the Netherlands over compensation for image use on Meta platforms. That case doesn't map one-to-one onto AI training, but it's relevant to the broader discussion of online image use and collective compensation.

The EU AI Act obligates providers of general-purpose AI models to publish a sufficiently detailed summary of the training content used. That doesn't automatically mean photographers get paid, but it makes the discussion about provenance and use less optional.

Opt-out: what works, what doesn't

What works reasonably well

Adjust your robots.txt. Known AI crawlers (GPTBot, CCBot, ClaudeBot, anthropic-ai, Google-Extended) honour a Disallow directive in robots.txt. This prevents future scrapes by these specific bots. Example:

User-agent: GPTBot
Disallow: /

User-agent: CCBot
Disallow: /

Helps for the future, not for what's already been downloaded.

Spawning.ai's "Have I Been Trained" opt-out. For LAION specifically, you can register your photo URLs. Stability AI committed that opted-out photos won't be used in future Stable Diffusion versions. How well this works in practice is hard to verify.

What doesn't work

• Watermarks on photos: AI models learn to ignore watermarks. Visible watermarks deter casual misuse, they don't stop training.
• C2PA / Content Credentials metadata: nice standard, but metadata is routinely stripped when a photo gets republished.
• Glaze / Nightshade: tools that "poison" training data with subtle changes to the photo. Works in research settings, in practice newer models bypass it within weeks.

The annoying part: none of these measures fixes the problem after the fact. They mainly help you stand on firmer ground from now on.

What you can actually do now

Document

Keep a record of where your photos appear, with or without credit. For data on how widespread the problem is, see image theft statistics 2025. Not to sue every individual AI training photo, but to build a file for:

• Possible future collective settlements. Pictoright is working on this kind of file in the Netherlands.
• Negotiations with clients who use AI tools in their workflow.
• Evidence in case a specific AI output is recognisably traceable to your photo (for example through watermark reproduction).

ImageTrace's automatic recurring scans (25 to 500 cycles per month depending on your tier) can do this documentation work for you, so you don't have to check manually every time.

Add AI clauses to your license terms

Standard contracts from a few years ago don't mention this. Add explicitly: "The licensee is not entitled to use the licensed work, or derivatives thereof, for training algorithmic or AI systems." Most national photographers' associations now have a template clause.

Be critical of tools that "enhance" your work

Adobe Firefly, Topaz Labs, Magnific, Photoshop Generative Fill: all useful, but read the terms. Some tools send your photo to their servers and use it for further training. Others don't. Adobe says Firefly was trained on licensed Adobe Stock content and royalty-free content.

Know what your portfolio is worth

When a collective settlement eventually comes, or if an AI company becomes open to individual settlements, the question is: what's your portfolio worth? For a concrete method to calculate that, see how to calculate damages for unauthorized photo use. That depends on how often and where it's been used, whether there's recognition in a specific niche, and how well you can demonstrate it.

And recognising derivative work?

The most practical question for photographers today: how do you know if your photos are being used, in what context (originals or AI-modified versions), and where? Reverse image search finds exact copies and strongly similar images. For recognisable AI output of your style there's no off-the-shelf detection method yet, but the foundation remains: knowing where your work actually lives.

ImageTrace compares your images per scan against millions of web pages and gives you the overview: copies, strongly similar images, and the URLs where they live. That's not "AI detection", but it is the portfolio map you need to determine where your position is strong and where it isn't. A scan of one photo costs five euros in Standard, less at higher tiers.