In early 2024, a striking illustration in a published article sparked a lively debate on social media. The image showed a rat with a penis and testicles larger than the rest of the animal’s body. The authors noted that the illustration was generated by an artificial intelligence model called Midjourney, but was obviously inaccurate, as it depicted four testicles and included strange, misspelled text, such as “sserotgomar cells.” Somehow, it passed peer review.

celebrating NatureCovers: past, present and future

“This was the first widespread example of how an AI-generated image became a scientific paper, and it should not have been published,” says Elisabeth Bik, a scientific integrity consultant in San Francisco, California, who wrote about the incident on her blog. AI tools at the time weren’t good enough to create credible user-requested illustrations, as the rat figure demonstrated, but that was then. Two years later, “AI is much better and continually improving, and we are at a point where we can no longer tell fake from real,” Bik says.

Even so, errors continue to appear. In April, a study by researchers in China was retracted by the New England Journal of Medicine due to image manipulation. The numbers on a tape measure, shown at the top of the figure, were incorrect, exposing the use of an artificial intelligence tool. In a comment on the post-publication discussion forum PubPeer, one of the authors notes that they had used an artificial intelligence tool to adjust the location of the measuring tape, which had been placed incorrectly during an emergency medical procedure. “The irregular numbering is an unintentional artifact of this adjustment,” they wrote.

How to use AI to make a graphical summary in minutes

Graphics, which include schematics, data figures, and diagrams, are a crucial part of scientific publishing. And they can substantially affect the influence of an article: an analysis of eight million graphs published in scientific articles found “a significant correlation between scientific impact and the use of visual information, with higher impact articles tending to include more diagrams” (P.-S. Lee et al. IEEE Translation. Big data 4117–129; 2018). However, many researchers have neither the resources nor the skills to create informative and aesthetically pleasing images themselves.

The potential for modern AI systems to help researchers generate graphs is “huge,” says Sebastian Porsdam Mann, an ethicist at the Center for Advanced Studies in Biosciences Innovation Law at the University of Copenhagen. AI tools can make scientific illustration accessible to everyone, allowing researchers to better communicate their science in a fraction of the time and at a lower cost than ever before, he says. And researchers are interested in having good articles that explain complex topics, with good data visualizations and graphical summaries, he adds.

‘Good design requires mastery’: scientific illustrators outline the future of AI

But, as with text, AI image generators like Midjourney and OpenAI’s DALL-E still make mistakes, the effects of which can range from personal embarrassment to professional censorship. “It’s still early days and it’s still largely untested waters,” says Mann. But if there’s even the slightest hint that you’ve done something wrong using AI tools, “journals will probably take it very seriously right now.”

Below are some guidelines to help researchers navigate this rapidly changing landscape.

See the publisher’s rules.

The first step for anyone who wants to use artificial intelligence tools in their scientific articles is to consult the policies of their preferred journals. There is little agreement on the use of AI tools among academic journal editors: some allow AI-generated images as long as their use is disclosed, and others ban them entirely.

The PLOS editor, for example, allows the use of artificial intelligence tools, but authors must report how they used them (including the names of the tools used, how they were used, how their result was evaluated, and in which sections of the article they were used).

AI-generated images and videos are here: how could they shape research?

Other publishers have a more restrictive stance. Cellular reports (published by Cell Press) prohibits any use of graphical summaries created by AI and imposes restrictions on the use of AI in data visualizations. Springer Nature (which publishes Nature) prohibits the use of generative AI for images, but makes exceptions for AI-generated images and videos in articles that discuss “specifically AI,” adding that “such cases will be reviewed on a case-by-case basis.” The policy also allows researchers to use AI imaging tools “developed with specific sets of underlying scientific data that can be attributed, verified, and verified for accuracy, provided that ethical, copyright, and terms of use restrictions are respected.” (NatureThe editorial team is independent of its editor.)

But journals’ guidance on AI use is rarely specific, says Mann, who researches AI policies in academic publishing. When talking about AI-generated images, people often think of art and illustrations. But scientific graphics can also include diagrams, visual summaries, figures representing data from a study, and images used as evidence. “The ethical issues are very different, whether you’re using images as evidence or to explain things,” he says.

Do not manipulate the original data.

In April, biologist Mikael Elias of the University of Minnesota in Saint Paul posted a series of compelling Western blots on the social networking site But these images were created by entering a single message into ChatGPT: “generate Western Blots representing an experiment in a nature magazine article.”

Scientific Figures Who Stand Out: Resources for the Artistically Challenged

“While data creation has always existed, this makes it unprecedented. [sic] easy and accessible,” Elias wrote on the blogging platform Substack (see go.nature.com/4wacqm9). “I fear an avalanche of manufactured parts, with very few ways to distinguish them from legitimate work. Not tomorrow, but soon.”

Bik agrees. She says that although there are clues in Elias’ images that the spots are fake, they would be easy to miss. And he strove to suggest any cases in which it would be acceptable to generate images to be provided as primary evidence using artificial intelligence tools.