A prestigious microscopy competition has become an unexpected flashpoint in a debate that reaches far beyond any single video. The question at its center is deceptively simple: when AI tools are applied to scientific images, where does enhancement end and fabrication begin?
A Winning Video That Raised More Questions Than Answers
Nikon Instruments runs an annual competition called Small World in Motion, which celebrates videos captured through light microscopes. This year, the prize went to Ning Xu, an optical engineer at the National University of Singapore. His entry documented the movement of cilia, the hair-like structures that line parts of the lungs, in tissue taken from a child with primary ciliary dyskinesia, a rare genetic disorder that causes chronic lung, sinus and ear infections.
The video is visually striking. Below the cilia, red, purple and blue structures appear. That is precisely what drew scrutiny.
Days after the announcement on September 15, microscopy experts began raising concerns on social media. Edward Phelps, a bioengineering researcher at the University of Florida in Gainesville, wrote on LinkedIn that the purple structures resemble mitochondria but that extracellular mitochondria of that size do not occur in biology. He noted that the blue structures look like nuclei but do not behave like nuclei, and that the identity of the red stain was unclear. The concern, stated plainly: these structures may not exist in the biological sample at all.
On September 22, Nikon Instruments updated its blog post about the entry to acknowledge that Xu had used an “unsupervised” AI model to assist with “post-processing.”
What the Creator Said, and What Remains Unresolved
Xu did not respond to Nature’s request for comment on the biological accuracy questions. He did write on LinkedIn that AI was not used to generate the experimental footage, the cilia themselves, or their motion. According to his account, the AI was applied after the fact to reconstructed grayscale data, to distinguish and colorize structures with similar morphology. The goal, he wrote, was to make the final presentation “visually engaging as well as scientifically interesting,” given that the video was prepared for a competition celebrating the beauty of microscopy.
That explanation draws a line between the underlying data and the visual rendering. But it does not fully close the controversy. Melanie White, a developmental biologist at the University of Queensland in Brisbane, Australia, pointed out that it remains unclear how faithfully the video represents what the microscope actually captured. The AI tool, she noted, may have introduced biological structures that do not exist in the original data.
This is the crux of the problem. Post-processing is not new in scientific imaging. Adjusting contrast, removing noise, and colorizing grayscale data are standard practices. What changes when an AI model handles that step is the degree of interpretive freedom the system exercises. An unsupervised AI model, by definition, is not guided by labeled examples of what structures should look like. It makes its own decisions about how to segment and render what it sees in the data. Those decisions may be visually plausible without being biologically accurate.
Why Scientific Images Are Not Illustrations
This is what most coverage of the story misses. The controversy is not really about a competition, or even about one video. It is about a foundational principle of scientific practice.
Markus Sauer, who studies super-resolution microscopy at the University of Würzburg in Germany, put the distinction clearly: using AI tools to visualize experimental data is not automatically a problem. It becomes one when AI-generated images misrepresent, exaggerate, or alter experimental findings. The line between those two outcomes is not always obvious, and that ambiguity is the real issue.
White’s framing is worth holding onto: scientific images are data. They are not artistic interpretations of data, and they are not illustrations designed to communicate a concept. They are the evidence itself. When a reader looks at a microscopy image in a paper or a competition entry, the assumption is that what appears on screen corresponds to something that exists in the physical sample. If an AI model has rendered structures that the microscope never detected, that assumption breaks down.
The implications extend well beyond competitions. Microscopy images appear in peer-reviewed research, in clinical contexts, and in the training datasets used to build future AI models. If AI post-processing introduces biologically implausible features that go undetected, those features can propagate through the scientific record.
The field does not yet have a clear, shared standard for what AI-assisted post-processing is acceptable in scientific imaging, or how it should be disclosed. This case has made that gap visible.
In Short
An AI tool applied to a prize-winning microscopy video may have rendered biological structures that do not exist in the underlying data. The creator says the AI was used only for post-processing, not to generate the core footage. Experts disagree on whether the result faithfully represents the original measurement. The deeper issue is that scientific images function as data, not decoration, and the field lacks agreed standards for how AI post-processing should be used and disclosed in that context.
Based on reporting from Nature: Machine Learning.