AI image generators are everywhere. Midjourney, DALL-E, Stable Diffusion, Adobe Firefly: if you run a website or create content, there's a good chance AI-generated images are already in your workflow.

Two questions matter if you use them for business. Can you own the images you generate? In the US, mostly not: copyright requires a human author, and prompts alone aren't enough. And can AI images infringe someone else's rights? Yes, sometimes, and the publisher is usually the one who gets the letter.

Here's what actually matters if you're using AI images for business. (This is general information, mostly about US law, not legal advice.)

Can You Copyright AI-Generated Images?

Short answer: mostly no.

The US Copyright Office has been consistently clear on this. Copyright requires human authorship. If an AI creates the image, no human author exists, and no copyright attaches.

The key rulings so far:

Thaler v. Perlmutter. Stephen Thaler tried to register a copyright for an image he said was created autonomously by his AI system. The Copyright Office refused, a federal district court agreed in 2023, and on 18 March 2025 the D.C. Circuit affirmed, holding that the Copyright Act "requires all eligible work to be authored in the first instance by a human being." The D.C. Circuit also noted that the human authorship requirement does not rule out copyright for work made by or with the assistance of AI. The Supreme Court declined to hear the case on 2 March 2026.

Zarya of the Dawn (2023). This one added nuance. Kristina Kashtanova created a comic book using Midjourney-generated images with her own text and arrangement. The Copyright Office ruled that while the overall selection and arrangement of the comic could be copyrighted, the individual AI-generated images could not. The parts a human created were protectable. The parts the AI created were not.

Where the line sits now. In January 2025 the Copyright Office published Part 2 of its report on AI, covering copyrightability. It says AI output is protected only where a human author has determined sufficient expressive elements, and that providing prompts alone is not enough. Human selection, arrangement or modification of AI output, and human work that incorporates it, can be protected. So if you sketch a composition, use AI to generate elements, then significantly modify, combine, composite, and paint over those elements in Photoshop, there's a stronger argument for human authorship. The more human creative control you exercise, the stronger your claim.

But if you're grabbing raw Midjourney outputs and putting them on your website? In the US, those images most likely have no copyright protection. Anyone can use them, and you may have no basis for a takedown notice if someone copies them.

For businesses, this means AI-generated marketing images, product visuals, and blog illustrations you create may not be protectable intellectual property. That's worth thinking about before you build a brand identity around them.

The Other Side: AI Images Can Infringe Existing Copyrights

This is the risk most people underestimate. Just because an AI generated an image doesn't mean that image is free of copyright problems. In fact, the way these models work makes infringement a real concern.

How AI Image Models Learn

Models like Stable Diffusion, Midjourney, and DALL-E were trained on billions of images scraped from the internet. Many of those images are copyrighted. The models learn patterns, styles, compositions and, critically, specific visual elements from that training data.

Researchers have demonstrated that diffusion models can and do memorize training images. A 2023 study from Google and UC Berkeley showed that Stable Diffusion could reproduce near-exact copies of training images under certain conditions. This isn't theoretical. It happens.

When you prompt an AI to generate "a professional photo of a sunset over the Golden Gate Bridge," the output is a statistical remix of every sunset and Golden Gate Bridge photo the model was trained on. Most of the time the result is sufficiently novel. Sometimes it's not.

The Lawsuits

Getty Images v. Stability AI. This is the big one. Getty sued Stability AI (maker of Stable Diffusion) claiming the model was trained on millions of Getty's copyrighted images without permission. Getty presented generated images that contained remnants of the Getty Images watermark. The case did not settle. In the UK case, Getty dropped its claims about how the model was trained before judgment, because there was no evidence the training took place in the UK, and it dropped its claims about the model's outputs too. In November 2025 the UK High Court rejected Getty's remaining copyright claim, finding that the trained model is not itself an "infringing copy", and upheld only narrow, historic trademark claims over reproduced watermarks. So the UK court never ruled on whether the training itself infringed. Getty has permission to appeal, and a separate US case is still ongoing. Other models, other lawsuits, same open question.

Class action by artists. A group of artists including Sarah Andersen, Kelly McKernan, and Karla Ortiz filed a class action against Stability AI, Midjourney, and DeviantArt. Their argument: these models can generate images "in the style of" specific artists because they were trained on those artists' copyrighted work without consent. Parts of the case survived initial motions to dismiss, and litigation continues.

Training cases outside images. In the US, Thomson Reuters v. Ross Intelligence and the 2025 rulings in Bartz v. Anthropic and Kadrey v. Meta show courts reaching different answers on AI training depending on the facts. None of them involved images; there's more on them under the fair use question below.

Real Risks for Businesses Using AI Images

This isn't abstract. Here are the scenarios that actually get companies in trouble.

An AI image looks too much like a copyrighted photo

You prompt Midjourney to create "a businessman shaking hands in a modern office." The output happens to closely resemble a popular stock photo from Shutterstock or Adobe Stock. You publish it on your website. The stock agency's automated crawler finds the visual similarity. You receive a demand letter.

This is a realistic risk. Stock agencies use visual similarity technology to find matches, and they don't care whether the match was generated by AI or copied manually. If the output looks like their image, their legal team sends a letter.

AI generates recognizable brand elements

Ask an AI to create "a sports drink bottle on a gym bench" and it might generate something that looks suspiciously like a Gatorade bottle, complete with the lightning bolt logo. AI models have ingested millions of branded images and can reproduce trademarked logos, product designs, and trade dress without you explicitly asking for them.

Publishing an image with recognizable brand elements exposes you to trademark infringement claims, even if you never mentioned the brand in your prompt.

AI generates celebrity likenesses

Prompt an AI for "a famous actor in a leather jacket" and you might get an output that clearly resembles a real person. Even if the result isn't photorealistic, a recognizable likeness can violate right of publicity laws in many US states. These laws give individuals control over commercial use of their name, image, and likeness.

Celebrities and their estates actively enforce these rights. Using an AI-generated image that resembles a real person in commercial content (ads, product pages, social media marketing) is a liability.

How to Protect Yourself

AI-generated images aren't inherently dangerous, but treating them as automatically safe is a mistake. Here's a practical workflow.

1. Scan AI outputs for visual similarity before publishing

This is the most important step. Before any AI-generated image goes on your website, blog, or social media, check whether it visually resembles existing copyrighted images.

PixGuard can help with part of this. Upload your AI-generated image with a free account and it checks for visible stock watermarks and copyright metadata, compares the image against a database of reference image fingerprints, and gives it a 0 to 100 risk score with the signals behind it. A fingerprint match catches copies of known images, not every look-alike, so a low score isn't permission to use an image. For a closer look, run a reverse image search on the output too.

2. Check for watermark residue

AI models trained on watermarked images sometimes reproduce faint or distorted watermark artifacts in their outputs. These are easy to miss at a glance.

Use our free watermark detector to check AI-generated images for residual watermark patterns. It looks for visible watermarks and flags unusual pixel patterns; a flag isn't proof of a watermark. If you find traces of a stock watermark, don't publish the image until you've looked into it.

3. Check metadata for AI generation markers

Many AI generators now embed metadata indicating the image was AI-created. This is increasingly important as platforms and regulators push for AI content labeling. Some stock agencies specifically flag and reject AI-generated content.

Our free metadata checker shows the EXIF, IPTC and XMP fields in an image file, including XMP fields some generators use to label AI-created images. Know what's in your image files before they go live.

4. Use models with known, licensed training data

Not all AI image generators are equal in terms of legal risk.

Adobe Firefly: Adobe says it's trained on Adobe Stock images, openly licensed content, and public domain material. That should reduce the risk of generating outputs that infringe specific copyrighted works.

Shutterstock's AI generator: Shutterstock says it's trained on its own licensed library, and it offers indemnification on some plans. Check the current terms for yours.

Models like Stable Diffusion (open source, trained on broad web scrapes) carry higher risk because their training data includes copyrighted material without explicit permission.

If copyright safety matters to your business, choosing a model with transparent, licensed training data is one of the most effective steps you can take.

5. Keep your prompts and generation records

Save your prompts, settings, model versions, seeds, and timestamps for every AI image you use commercially. If a copyright dispute arises, this documentation demonstrates that you generated the image through AI rather than copying it directly. While this doesn't guarantee you're in the clear, it shows good faith and helps establish that any similarity was unintentional.

Some organizations are building formal AI image generation policies that require logging all prompts and outputs. If you're using AI images at scale, consider doing the same.

The Legal Landscape in 2026

The law around AI-generated images is still being written. Here's where things stand and what's still open.

What's settled

Purely AI-generated images without significant human creative input are not copyrightable in the US.
Using copyrighted images to train AI models without permission is being actively challenged in court.
Existing copyright, trademark, and right of publicity laws apply to AI outputs the same way they apply to any other content. Generating an infringing image with AI is not a defense.

What's still in play

The fair use question. In the US, no appeals court has yet decided whether training an AI model on copyrighted works is fair use. District courts have split on different facts: Thomson Reuters v. Ross (February 2025) found that training a non-generative legal research tool on Westlaw headnotes was not fair use, and that ruling is on appeal to the Third Circuit, while Bartz v. Anthropic and Kadrey v. Meta (June 2025) found that training language models on books was fair use on their records. None of these cases involved images.
The EU AI Act includes transparency obligations for some AI-generated content, such as deepfakes. If you publish in the EU, check which of them apply to you and from when. Other jurisdictions are considering similar rules.
Proposed US legislation. Multiple bills have been introduced in Congress addressing AI and copyright, including proposals to require AI companies to disclose training data and to create licensing frameworks. Check the current status of any bill before relying on it.
Model-level solutions. Some AI companies are building opt-out mechanisms for rights holders and adding content credentials to outputs. These technical solutions may reduce friction but don't resolve the underlying legal questions about existing models already trained on copyrighted data.

What happens next

More clarity will come as the major lawsuits reach rulings, appeals or trial, but it isn't yet clear how these cases will come out. Whatever happens on training, you are already responsible for making sure the AI outputs you publish don't infringe existing rights.

In the meantime, the risk falls on the publisher. If an AI-generated image on your website infringes someone's copyright, you're the one who receives the demand letter. Not Midjourney. Not Stable Diffusion. You.

Don't Assume AI Images Are Safe

The convenience of AI image generation is real. But "an AI made it" is not a copyright defense, and it's not a guarantee of originality.

Before you publish AI-generated images on your website, social media, or marketing materials, scan them. Check for visual similarity to known copyrighted works. Check for watermark residue. Check the metadata.

A free PixGuard account includes 30 image scans, valid for 30 days, to check AI-generated images (or any images) for copyright risk signals. Upload an image or scan your website and you'll get a 0 to 100 risk score for each image checked, with the signals behind it. No credit card required.

Or check a page on your site free, no signup: enter a page URL and we check the 3 largest images on that page, up to 3 times a day.