How do we use AI images ethically? While I obviously am comfortable using AI images, I also have limits. I have guidelines and boundaries for my own use of AI image generation. For example, I don’t use images of real people without their consent or emulate the style of living artists. I aim to use AI images to improve the quality of the training I build and to show more diversity. AI also raises concerns about the training data, environmental impacts, and copyright questions. I use AI images, but I also know AI images can be used in ways that cause harm. Here’s my current thoughts and considerations for using AI images ethically and appropriately.
Obviously, I’m not a lawyer, and nothing I say here is legal advice. This is my understanding of fair use and copyright. Check with your own attorney if you have legal questions.
My goal here isn’t to convince you to use AI images if you’ve already decided it goes against your ethical values. If that’s you, I respect that, and I’m not here to change your mind. I’m writing for those who are using AI images and want to do so carefully.
Training data
I know lots of folks who oppose AI, and AI images particularly, because of the training data. I think it’s totally reasonable to use Adobe Firefly for image generation because it was trained only on licensed stock and public domain images. Because they have permission for everything they used to train their model, that’s the safest option. That’s an easily defensible ethical stance.
Most AI image generation models were trained on a much broader set of data. That includes photos and art that are available online but not in the public domain. Just like LLMs were trained on copyrighted text, image generation models were trained on copyrighted images. That means AI companies are relying on a fair use argument.
I’ve seen a range of legal opinions on AI training. This is clearly an area where the law is still evolving. So far, the decisions support training models on copyrighted materials as long as they were legally obtained. Anthropic is paying a settlement for using pirated books to train on, but the books they legally bought were allowed under fair use. That’s where I generally fall too; as long as the training data wasn’t pirated, training on a broad range of publicly available images is fair use.
I don’t follow the legal landscape in the EU as much as I do in the US. However, it looks like the EU AI Act will require more transparency on data sources. (Any readers with more insight on the EU, feel free to comment and correct me if I’m wrong about that.) I hope we do get more transparency on training data so people can make informed choices about which tools they use or avoid.
Transformative use
Fair use is the US doctrine that allows some use of copyrighted content under certain circumstances. Fair use is very “squishy” though; it’s not a clear cut rule. It’s not as simple as saying there’s a clear line based on any single factor. You have to weigh four factors, and judges sometimes decide individual cases differently even with similar circumstances.
If someone tells you that something is definitely not fair use because it uses the whole source while ignoring the other three factors, be very skeptical. That’s not really how fair use works. Yes, the amount of the source material used matters, but that’s only part of the question.
One of the other four factors is the purpose and character of your use of source content. That’s why parody and satire are often allowed even of copyrighted or recognizable sources; that’s a protected purpose that constitutes fair use. The judge in the Anthropic case ruled that AI training is a transformative use; it transforms the original into something else. Because the character of the use is transformative, it’s fair use.
For my own work, that ruling means I’m less concerned with the input for AI and more concerned with the output. Because of course you can make things with AI that are copies (or very close to copies) and aren’t really transformative. Just because an AI image model lets you generate something obviously based on Star Wars or the Simpsons doesn’t make it transformative enough to be considered fair use. And regardless of the legal landscape, transformative use is more ethically defensible.
Don’t emulate living artists
One of my own guidelines for using AI image generation is that I don’t use it to emulate the style of any single living artist. You want to make a version of Starry Night with a cat in it? Cool, go for it! Van Gogh isn’t alive to be affected by it. But any artist who is alive now, who you (at least hypothetically) could hire to create something? Nope. While it was trendy to do Studio Ghibli style remixes for a while, I opted not to participate in that.
Copying the style of a living artist potentially affects the market for that person’s art—and that’s another of the four factors weighing against fair use. Plus, that use doesn’t feel right to me. Copying a single style doesn’t feel transformational. Plus, why use these tools to simply copy when they’re capable of more?
Instead of emulating a specific artist, use AI to develop a new style. That’s one of the reasons I still use Midjourney even though other image generators are now better at consistency and prompt adherence. Midjourney is the best for creating new styles and mashing up elements of multiple images together.
It’s like the old academic joke: if you copy from one source, it’s plagiarism, but if you copy from a bunch of sources, it’s research. AI image styles feel similar. If you copy from one artist, it’s plagiarism. If you copy from a bunch of artists, it’s transformative and fair use.
AI images can’t be copyrighted
Like anything else produced by AI, AI-generated images can’t be protected by copyright unless a human makes substantial changes to it. Work that is primarily guided by the decisions of a human but is “AI-assisted” may be copyrighted. Read AI and Copyright for more details and examples of what kinds of edits and decisions make something “AI-assisted” rather than “AI-generated” and an explanation of the March 2026 US Supreme Court decision.
In my opinion, that doesn’t matter too much for images within training. If you were using stock images previously, you didn’t own the copyright to those images either.
While there are multiple AI tools for creating business logos, that feels really risky to me. Even if you’re a small business and don’t plan on trademarking your logo, wouldn’t you like the option of doing so in the future? Pay for a human to create your logo so you can protect that IP.
The safest tool related to IP claims is Adobe Firefly, which also gives you indemnity against IP claims with paid plans.
As a side note, this is one of the issues with using AI to generate course content too, especially courses to sell. Someone can duplicate your content and sell it, and you might not have much recourse. This is an area of the law to keep watching, since future court decisions or legislation may change it.
Commercial use of AI images
Check the ToS for your AI image tool. Most free tools for AI images don’t license your results for commercial use. You generally have to use at least a minimal paid plan to get commercial use rights.
Don’t use images of real people without permission
Don’t use images of real people as reference images for AI unless you have explicit permission to do so. That’s not just for cases of deepfakes or sexually explicit images; that goes for everything.
I see way too many examples on LinkedIn where people have done “fun” transformations of people’s photos. Most of that isn’t malicious. It’s stuff like people making composite group images or turning workshop participant images into superheroes or whatever. But people should have the right to not have their images used that way, where your identity is clear.
Plus, even with good intentions, sometimes the images are problematic. I remember seeing someone’s AI-generated image of people they were looking forward to having a beer with…except I noted multiple people who don’t drink shown in that image. I shouldn’t have to explain that you shouldn’t generate images of a Jewish person eating bacon or a Mormon drinking alcohol, but apparently that’s a thing that need to be explicitly mentioned.
Even if you have stock images of real people, you may not have permission to use those images as references in AI. If you need character images, just use AI to generate your reference image and work from there.
Environmental impacts of AI images
The water use for AI data centers gets a lot of attention, but every time I actually dig into the sources and data, I see that the water use is relatively minimal. I don’t want any AI data centers drawing water from the Colorado River where water use is already such a problem. However, I think data centers can be built in places where water is more available, and they can be built and managed in ways that reduce their impact.
Electricity use is potentially a problem, especially at scale. AI models are much more efficient than they used to be, so they don’t use as much energy as early models (at least for regular prompting). I’d love to see more investment in renewable energy in the US. Where I live in North Carolina, data centers don’t pay taxes on electricity. I’d like that to change; I think data centers shouldn’t be incentivized so much that it’s hard to track the real costs.
You can also reduce your energy use by using cheaper models when possible. I often use Midjourney’s draft mode if I know I need a lot of iteration to fine tune something, partly because that reduces my energy use.
It’s important to keep the water and electricity use of AI in perspective with other tasks though. Check out the What Uses More? calculator to compare the energy and water use of different tasks.
For example, would you think twice about the water and energy use for a one-hour Zoom call with 10 people? Probably not.
Guess how many images could you generate with AI for the same resource use as that one-hour Zoom call?
According to the calculator, that one-hour Zoom call is equivalent to generating 1000 images. I generate a lot of AI images, but even I have only generated about 1500 images in Midjourney over the course of two years. Are you going to generate 1000 images in a year? How many Zoom calls will you have in that same year?
Transparency in AI use
One question that comes up often when I present about AI images is about transparency and labeling AI images. This is an interesting question because I don’t think there’s a clear answer that fits every situation. I had a client project where I clearly labeled everything I generated with AI but was asked to remove the disclosures. I can’t find the source right now, but I remember reading somewhere that Americans were generally less worried about having AI labeled, but Europeans (especially Germans) value transparency more. Cultural differences play a role here, as does organizational culture.
I used to label every AI-generated image here on my blog, but these days I admit that I figure people mostly know even without the label. But maybe I should be more explicit? From an ethical standpoint, transparency feels like the more defensible position.
Counter bias in AI images
AI image generators tend to exaggerate stereotypes. If you don’t actively work to counter bias in AI images, it can be worse than stock images. The good news is that AI images can give us even better, more inclusive representation than stock images. We’ve had the ability to create images of nonbinary characters and people of varied racial, ethnic, and national backgrounds for a while. We’re now finally getting to the point where disability representation is better. Inclusive images is one area where I think AI can give us better outcomes than stock photos. Using AI for social good is an ethical consideration too.
Review AI images
Just like you need to review AI-generated text for accuracy, review your AI-generated images. Beyond obvious AI errors, check the background objects, watch out for bias, and verify the results. If you always accept the first image result without refining your prompt or editing the result, that’s a problem. Your human judgement is more important than ever.
Further reading on AI ethics and legal considerations
What do you think?
What are your ethical guidelines for AI image generation? Do you prefer to see AI images transparently disclosed?
I expect a fair amount of disagreement on this. I’ve had some great discussions about this though, even with people who don’t agree. I’m hoping for some nuanced conversations about balancing our ethical considerations. Let me know your thoughts in the comments.
