For more than three decades, science photographer Felice Frankel has assisted MIT professors, researchers, and students in conveying their work visually. Over this period, she has witnessed the emergence of different tools designed to aid in the production of striking images: some beneficial, while others counterproductive to the goal of delivering a reliable and comprehensive depiction of the research. In a recent editorial featured in Nature magazine, Frankel explores the growing application of generative artificial intelligence (GenAI) in imagery and the challenges and ramifications it poses for research communication. On a more personal note, she ponders whether there will continue to be a role for a science photographer within the research community.

Q: You’ve pointed out that once a photograph is captured, it can be deemed “manipulated.” There are methods by which you’ve adjusted your own images to create a visual that more effectively conveys the intended message. Where do you draw the line between acceptable and unacceptable modification?

A: In a broad sense, the choices made regarding how to frame and compose the content of an image, along with which tools are employed to generate the image, are already a modification of reality. It is important to remember that the image is simply a representation of the subject, not the subject itself. Choices need to be made during the image creation process. The key concern is not to manipulate the data, and in most images, that data pertains to the structure. For instance, in an image I created some time ago, I digitally removed the petri dish containing a yeast colony to highlight the remarkable morphology of the colony. The data represented in the image is the morphology itself. I did not alter that data. Nonetheless, I always disclose in the text if I have manipulated an image in any way. I explore the concept of image enhancement in my guide, “The Visual Elements, Photography.”

Q: What steps can researchers take to ensure their research is presented accurately and ethically?

A: With the rise of AI, I identify three primary concerns regarding visual representation: the differentiation between illustration and documentation, the ethical considerations surrounding digital modification, and the ongoing necessity for researchers to receive training in visual communication. For many years, I have been attempting to create a visual literacy program for both current and future cohorts of science and engineering researchers. MIT has a communication requirement primarily focused on writing, but what about the visual aspect, which is increasingly integral to a journal submission? I would wager that most readers of scientific papers go directly to the figures right after reviewing the abstract.

We need to mandate that students acquire the ability to critically assess a published graph or image and determine if there is anything peculiar about it. We must address the ethics of “nudging” an image to appear a certain pre-defined way. In the article, I recount an experience when a student altered one of my images (without seeking my permission) to align with what the student wished to visually convey. Naturally, I did not authorize it, and I was disheartened that the ethical implications of such alterations had not been considered. We should initiate, at the very least, discussions on campus, and ideally, establish a visual literacy requirement in conjunction with the writing requirement.

Q: Generative AI is here to stay. What do you envision for the future of visual science communication?

A: For the Nature article, I determined that an impactful way to challenge the use of AI in image generation was through an example. I employed one of the diffusion models to generate an image using the following prompt:

“Generate a photograph of Moungi Bawendi’s nano crystals in vials set against a black background, fluorescing at varying wavelengths based on their size when stimulated with UV light.”

The outcomes of my AI trials often resulted in cartoonish images that were barely recognizable as reality — let alone as documentation — but a time will come when they might be. In discussions with colleagues within the research and computer science fields, there is a consensus that we need to establish clear standards for what is permissible and what is not. Most importantly, a GenAI visual should never be accepted as documentation.

Nonetheless, AI-generated visuals will indeed prove beneficial for illustrative purposes. If an AI-generated visual is to be submitted to a journal (or shown in a presentation, for that matter), I firmly believe the researcher must

• clearly indicate whether an image was generated by an AI model;
• specify which model was utilized;
• include the prompt used; and
• provide the image, if available, that assisted in formulating the prompt.