Do AI-Generated Images Have Copyright? The 5 Questions You Must Know About Commercial Midjourney and Stable Diffusion Use in 2026

Do AI-Generated Images Have Copyright? The 5 Questions You Must Know About Commercial Midjourney and Stable Diffusion Use in 2026

Designer Xiao Wang used a key visual generated by Midjourney in the materials for his company's new product launch, but the client was later sent a legal notice by another company on the grounds that the AI image was suspected of infringement. For the first time, he realized that AI image generation seems to have zero barriers to entry, but copyright issues are more complicated than with traditional stock libraries. This is not an isolated case. Over the past two years, several major jurisdictions—the United States, the European Union, China, and Japan—have all seen controversial rulings on the copyright of AI works, and the core logic of each ruling is not entirely consistent.

This article summarizes several core issues that need to be clarified for commercial AI image generation as of the time of writing. They include copyright ownership, whether commercial use is allowed, the legality of training data, the risks of users uploading images to the model, and differences in domestic and international judicial practice. All conclusions are given with their reasoning as far as possible, and parts involving uncertainty are explicitly noted; no absolute conclusions are drawn on the reader's behalf.

Who actually owns an AI-generated image?

This is the most frequently asked question and the one with the most complicated answer. From a legal-tradition standpoint, the object of copyright protection is "original human intellectual achievement." If an image is generated entirely automatically by AI and the human only entered a short prompt, the mainstream view in most jurisdictions is that the image lacks "original human participation" and does not constitute a work.

The U.S. Copyright Office's position in its public guidance is relatively clear: purely AI-generated images are not eligible for copyright registration. However, if there is significant human participation in the creative process—such as multiple rounds of manual retouching, combining multiple AI outputs, and adding original elements—the human-created parts can be protected. In China, in the 2023 Beijing Internet Court "First AI Text-to-Image Case," the court held that a work that had been repeatedly adjusted through prompts and infused with the user's aesthetic judgment could constitute a work and received copyright protection. The logic on both sides is similar, but the threshold details differ.

Whether commercial use is allowed comes down to the agreement

From a legal standpoint, whether an image belongs to you and whether you can use it commercially are two different things. Midjourney stipulates in its terms of service that paying subscribers own the generated images and may use them commercially, but the terms also state that Midjourney itself reserves the right to use these images for training and promotion. Stable Diffusion's open-source license is looser by default; the generated images themselves do not belong to the platform, and whether they can be used commercially depends on whether you used a particular model that carries commercial restrictions.

Several mainstream domestic platforms—ByteDance's Jimeng, Alibaba's Tongyi Wanxiang, Baidu's Wenxin Yige, and MiniMax's Hailuo—describe their commercial terms inconsistently in their respective user agreements. It is recommended that, before commercial use, you read the latest version of the user agreement for the platform you are using, with the terms on the official public page taking precedence. This step may seem tedious, but it is far cheaper than being sued afterward.

The legality of training data

The reason an AI model can paint in a certain artistic style is that its training data contains a large number of samples of that style. The question then arises: the copyright of those samples belongs to their original authors, so did the model producer obtain authorization? The answer differs across jurisdictions.

After the EU AI Act took effect in 2024, it set out clear requirements for the transparency of training data. Model providers need to disclose the sources of the data they use and respect the copyright-reservation statements of original authors. The United States is still debating this matter across multiple lawsuits—a typical case being Getty Images suing Stability AI, which still has no final verdict. China's regulation of training-data legality is set out in principle in the Interim Measures for the Management of Generative Artificial Intelligence Services issued by the Cyberspace Administration of China, which requires training data to be lawful and compliant. At the user level, as a downstream user you usually cannot trace the training data. However, if you use a model that has been found infringing by a court to produce images, your usage risk rises accordingly.

The compliance path for domestic AI image generation

If your clients are in China, give priority to filed (registered) domestic AI image-generation services, as the compliance risk is relatively controllable. As of the time of writing, service providers that have been publicly filed include ByteDance, Alibaba, Baidu, SenseTime, Zhipu, and MiniMax. These models perform stably in Chinese-language scenarios, their terms of service are written in Chinese, and the appeal path is clear when problems arise.

There are also many domestic applications built on filed models, such as Lingtu, an AI image-generation app available in the iOS China store. It aggregates a Midjourney-style atmosphere engine, a Flux-style photorealistic engine, and a Nano Banana-style fast engine, with Chinese interaction, localized prompts, and no need for a VPN—suitable for domestic users who don't want to bother with international services. You can download it by searching for Lingtu in the App Store, save the generated images to your local photo album, and, as always, read the latest version of the agreement terms in the app before commercial use. The advantage of this kind of tool is that the compliance boundaries are clearer for domestic users, and the generation speed is fast too.

The risk of user-uploaded reference images

Many AI image tools support uploading reference images for style transfer or img2img redrawing. There are two levels of risk here. The first level is that if the uploaded reference image itself carries someone else's copyright, your redrawn result will be treated as a derivative work, doubling the risk of secondary commercial use. A common misconception is to take an image someone posted on Xiaohongshu or Pinterest as a reference, thinking "I'm not using it directly." But legally, the copyright traceability of a derivative work is not waived just because an AI step was added.

The second level is portrait rights. If the reference image contains a real person's face, then even after AI redrawing it may still involve infringement of portrait rights, especially in commercial scenarios where the rules are stricter. A safe approach is to use reference images that are either taken by yourself, public-domain works, or properly licensed materials, and to avoid grabbing images directly off the internet. If a face in the generated image "looks a lot like a certain celebrity," it is also advisable to adjust it manually—facial recognition tools can now reverse-search it.

A compliance self-check list before commercial use

Before officially putting AI-generated images to commercial use, it is recommended to run through these five self-checks. First, confirm whether the platform's current terms of service clearly allow commercial use; the terms for free users and paying users differ greatly, and Midjourney's free tier strictly cannot be used commercially. Second, did you upload any reference image with copyright or a portrait during generation? If so, abandon that image immediately and regenerate. Third, do the generated results show any trace of copying a specific artist's style, such as using the artist's name directly as a prompt? If so, rewrite the prompt and regenerate.

Fourth, did you do significant human creative processing, such as reworking, compositing, and text layout—this affects subsequent copyright registration. Fifth, if it is for an important client project or large-scale deployment, keep complete generation logs and prompt records, so you can produce proof of process if questioned. Once these five items are done, the legal risk in conventional commercial scenarios is basically under control.

Industry cases and trend observations

Observing the direction of rulings over the past two or three years, you can see two trends. First, countries are drawing an increasingly fine distinction between "pure AI generation" and "AI-assisted human creation," and the threshold usually falls on "whether there is enough human decision-making to reflect originality." Repeated prompt adjustments, multi-version screening, and manual post-processing are all plus points. Second, platforms are shifting risk away from themselves; almost all commercial-grade AI image service terms include a clause stating "the user warrants that their input does not infringe, and the user is responsible for any disputes arising from the input."

This means that even if you use a lawful platform, if you input unlawful material, the responsibility still lies with you. AI image generation has lowered the production threshold, but the compliance threshold has not lowered—it has simply moved from the production end to the usage end. Understand this, and your commercial decisions will be more secure.

FAQ

I made the image with Midjourney. Can the client use it directly as a logo?

Not recommended. The core requirements of a logo are exclusivity and trademark registrability, but AI-generated images cannot obtain copyright in most jurisdictions, meaning others can generate nearly identical images and the risk of trademark squatting is high. If you must use AI to produce a logo prototype, you must do extensive manual redesign—vectorization, font customization, and color specifications all completed by hand by a designer.

Are images from a domestically filed model completely risk-free?

No. Filing resolves the platform's operational compliance. For a specific image of a user's, the copyright ownership, scope of commercial use, portrait rights, and reference-image risk must still be confirmed item by item according to the points above. Filing is a necessary condition, not a sufficient one, and cannot be equated with no risk.

Will including an artist's name in the prompt cause infringement?

It depends on the jurisdiction and scenario. For private use it does not constitute infringement, but in a commercial project, using a living artist's name directly as a prompt is suspected of improper imitation and unfair competition. Deceased artists and public-domain works carry lower risk. A safe approach is to switch to a style description—for example, changing "Van Gogh style" to "Post-Impressionist brushstrokes with rich color."

I used a photo I took myself as the reference image. Do I own the copyright of the redrawn image?

Part of it is yours. The copyright of the original reference photo belongs to you. If the redrawn image involves significant AI intervention, in most jurisdictions the AI-involved part does not constitute a work; it depends on whether you subsequently did significant human re-creation. On balance, the copyright ownership of a redrawn selfie is clearer than that of pure AI generation.

How do I handle copyright differences in cross-border projects?

The safest approach is to follow the strictest jurisdiction. If the same project is to launch in China, the U.S., and Europe, prepare documentation according to the requirements of the EU AI Act plus the U.S. Copyright Office guidance, including training-data transparency, prompt records, and evidence of human creation. Doing this raises the project delivery cost, but it avoids a blowup in any single market.

AI image generation has pushed the marginal cost of producing images to nearly zero, but the complexity of copyright has not decreased—it has increased. Spend ten minutes running through these common questions before commercial use, and you'll avoid a good many pitfalls down the line.