Copyright, Generative AI, and Scientific Visuals
Understanding copyright in scientific image generation: what's protected, what's fair use, and how copyright-safe models can help researchers use AI confidently.
There's a lot of discussion about generative AI and copyright. Real concerns have been raised about what comes out of these systems and what went into them in the first place to make them so good. From a scientist's perspective, though, a lot of the debate misses how I use AI and how I think about copyright. (Standard caveat: I'm not a lawyer, and rules vary by country.)
When I use AI to prepare scientific visuals, I assume most image models were trained on copyrighted material, including scientific figures from books and journals. Probably some of mine. But scientific figures are an unusual case in copyright, and it's worth being precise about what's protected.
You can't copyright a molecule. You can't copyright a scientific fact. What can be copyrighted is a specific expression of a fact: a particular drawing style, layout, composition, graphical treatment, and the exact pixels in a published figure. That's effectively what many journals control: the specific representation, not the scientific meaning.
So if you want to reuse a figure image from a journal, you usually need permission. But if you redraw the same molecule or recreate the same idea with a different rendering, you're often creating a new expression rather than reusing the journal's protected one. That's one reason scientific image generation can feel fundamentally different from generating a Disney-branded character or a trademarked logo.
Fair use (in the U.S.) can protect some educational and research uses—especially classroom use, commentary, and transformative reuse—but it's always context-dependent. Here is a good explainer on copyright and generative AI: https://creativecommons.org/2023/08/18/understanding-cc-licenses-and-generative-ai/
Even if the output of a model doesn't infringe on copyright, people still care how the model was trained. That's a real ethical concern, and it's one reason some scientists avoid image models entirely. If I trained a model on scientific figures I would avoid stealing other people's work to do this. If I went to the library and stole a journal, scanned it, and trained on it, I'd expect to get in trouble. But only because I stole the journal, not for reusing the content – that is covered under fair use laws.
Frontier AI labs have been caught doing essentially this, stealing copyrighted materials. And they are being held to account for it. On the technical side I'm very confident that we can get LLMs that are just as performative without stolen work. I think we are not far off from being able to use frontier models without feeling like we are aiding theft.
For most scientists, the best approach is simple: use AI for scientific visuals, and don't intentionally steer models toward copyrighted characters, brands, or distinctive proprietary styles especially if you'll publish commercially or try to claim copyright in the final work.
But if you want to be meticulous about risk, because of where you plan to use the output, or because you prefer not to rely on models trained on disputed data, there is another option:
You can use models trained specifically on licensed, copyright-safe materials. Adobe Firefly was an early prominent example of this option. More recently, providers like Bria have made this category easier to access in a developer-friendly way.
That's why ChemIllusion now offers copyright-safe model options through Bria. If your goal is to create figures you can confidently use in papers and commercial materials, these models can reduce uncertainty. ChemIllusion works well with these models. Our process of augmenting a model's handling of chemistry using prompt and context engineering helps squeeze the best results out of them.
Generative AI can genuinely improve how science is communicated: clearer figures, faster iteration, better teaching visuals, and less time lost to editing details. Copyright uncertainty has kept some researchers on the sidelines. My hope is that offering copyright-safe model options makes it easier for more scientists and educators to use these tools confidently and responsibly.