The Microscope Video That Needs Its Raw File

By Ellis Ward, Resident Expert on AI Systems and Research Norms, Reporting from the Uncanny Valley

When Nikon announced the winner of its 16th annual Small World in Motion competition in mid-September, the story looked straightforward: a striking research video, a rare disease made visible, a deserving prize. Within days, scientists were arguing that the video's vivid colors concealed a methodological problem that the competition's rules explicitly prohibit. Nikon is now re-reviewing the entry. No decision has been announced.

What the winning video showed, and why it matters

The first-place entry was submitted by Dr. Ning Xu, an optical engineering researcher whom Nikon lists as affiliated with Tsinghua University; other outlets have described Xu as a National University of Singapore researcher, and the exact affiliation has not been settled across sources. The video captures the abnormal beating of airway cilia taken from a child diagnosed with primary ciliary dyskinesia, or PCD, using diffractive super-resolution microscopy at 100-times magnification.

Cilia are the microscopic, hair-like structures that line the airways. In a healthy person, they beat in coordinated waves, sweeping mucus and debris upward and out of the lungs. In PCD, those cilia can beat abnormally, making it difficult for the airways to clear themselves and contributing to recurrent respiratory infections. Diagnosing and understanding PCD depends on observing ciliary motion directly, which is exactly what makes a video more scientifically valuable than a photograph.

Xu explained the reasoning plainly. As PetaPixel reported, Xu wrote: "To understand this disease, you need to watch the cilia through dozens of beat cycles, not just a still image." The dispute is not about whether motion matters. It is about what happened to the image afterward.

False color is normal. This situation is not quite that.

Before describing what critics saw, it helps to understand what routine microscopy processing actually looks like, because the line between normal practice and the objectionable is easy to blur if you approach it from outside the field.

False color is common in microscopy. The resulting colors in a published image almost never correspond to the "real" appearance of a structure. They correspond to which dye was used, which fluorescence channel was active, or which algorithm was assigned to which feature. A 2010 paper that has since become a widely cited reference in image-handling guidelines established that colorized images are acceptable provided the processing is disclosed and described. That standard predates generative AI, but its core logic, namely that color is a representation tool and not a lie so long as the tool is named, remains the working principle across the field.

The contest rules for Small World in Motion do not prohibit color processing. They do state, in plain terms, that AI-generated videos are not permitted.

What critics observed

When scientists began examining the winning video after the September announcement, the objections were specific rather than vague unease about colors.

Edward Phelps, a researcher at the University of Florida, described what he saw to multiple outlets: purple structures that "pop in and out of existence" with no biological explanation. Phelps noted the purple structures resembled mitochondria in shape, but extracellular mitochondria of that apparent size do not occur in biology. He also identified blue nodules that resembled nuclei but did not behave as nuclei behave, and red structures he could not account for at all. As PetaPixel reported, Phelps said the green cilia "also appear from nowhere and do not match the known size." The concern is not that the colors are false, which is expected, but that the structures the colors are attached to may not reflect structures that were actually there.

Ian Donovan, a researcher at UT Southwestern, took a different approach. He ran the video through Google Gemini and reported finding an invisible SynthID watermark, the AI-content identification system pioneered by Google DeepMind, embedded in the footage. Gizmodo noted separately that most detectors that try to spot AI without a watermark fare poorly, and that without SynthID or an equivalent it can be difficult or impossible to determine an image's provenance with complete certainty. Donovan's finding is a reported claim, not an established fact about the video's provenance.

The competition's fifth-place winner, Patrick Hickey, addressed the rules directly. Hickey stated that the rules clearly barred generative AI and that every entry must be captured under a microscope. His position was that the question was not about interpretation but compliance.

"The point of image processing is to help us to see, and quantify, structures that we otherwise would not be able to resolve, not to hallucinate and make up things in your biological sample that look like other structures." Valentin Dunsing-Eichenauer, researcher, calling for release of the raw data

Xu's defense, in his own words

Xu responded to the criticism in LinkedIn comments. PetaPixel reported that those comments were later not findable as of October 1. The substance of Xu's position, as reported before the comments disappeared, was this:

"AI was not used to generate the experimental movie, the cilia, or their motion." Xu acknowledged that an AI method was applied afterward, writing that "it was applied afterwards to the reconstructed grayscale data to distinguish and color structures with similar morphology." He also stated that the colors are false and "do not correspond to fluorescence channels identifying specific cells or organelles," and that regions below the cilia were rendered to improve visual presentation "without making anatomical claims about what those rendered features represent."

Xu's argument, in short, is that the underlying data is real, the cilia motion is real, and the AI's role was limited to post-processing visualization, not content generation. The question the scientific community is pressing is whether that distinction holds up when the post-processing introduces structures that do not correspond to measured data.

Nikon subsequently updated language on its website to reflect what the post-processing involved. The updated description reads: "An unsupervised neural-network method was subsequently used as part of the post-processing and visualization workflow to distinguish and visualize features in the grayscale data, creating a more vivid and visually engaging video."

The timeline

The sequence of events is worth mapping clearly, because the timing of disclosures matters to how the dispute reads:

Mid-September 2026: Nikon announces Xu as the first-place winner of the 16th Small World in Motion competition.

Shortly after: Scientists, including Phelps and Donovan, begin posting objections publicly. Donovan reports his SynthID finding.

September 26: Nikon posts on LinkedIn that it is "carefully re-reviewing the information provided during the initial vetting, along with additional supporting materials from the entrant," and notes that Xu is "respectfully complying and has provided detailed technical documentation."

October 1: As of PetaPixel's reporting, Nikon had not responded to USA TODAY, and Gizmodo reported no update as of that Friday evening. Xu's original LinkedIn comments had become unfindable. No final decision has been announced.

BBC News reported that Xu did not respond to their request for comment. BBC News also provided context on false-color imaging practices in the broader reporting.

What would settle the dispute

The scientists who have raised objections agree on what resolution looks like: release the raw grayscale video that was captured before any post-processing was applied, along with a complete methods description of what was done to that file and in what order.

Melanie White, a researcher at the University of Queensland, told Nature that scientific images are data, and that scientists "need to be able to trust that what we are seeing is grounded in the underlying measurement." The raw file is, in effect, the measurement. Everything applied after that point is representation. The critical question is whether the representation introduces structure that the measurement did not actually contain.

Former judge Andrew Moore offered a less technical framing. He said the situation recalled "old photographs restored with AI," and observed that "if it was your grandma who was face-swapped, it's going to be unsettling." The comparison is not precise, but the intuition it points to is real: restoration and fabrication can look identical from the outside, and the difference is entirely in the method, not the output.

Dunsing-Eichenauer called directly for the raw data to be released, a step that would allow independent verification of which structures in the final video correspond to structures in the underlying grayscale image.

Why this matters beyond one competition

A microscopy competition is a small venue. The underlying question is not. Scientific imaging has become more powerful, more processed, and more opaque to readers in roughly the same period that generative AI has become easy to apply without specialized training. Fluorescence microscopy and super-resolution methods both produce raw data that bears very little resemblance to the final published image. The processing pipeline between raw acquisition and published figure is now long, and large sections of it can be automated.

The contest rules that Nikon wrote represent one institution's attempt to draw a line. That line, in this case, runs between AI as a processing tool and AI as a generative one. Xu's defense rests on being on the permissible side of that line. The critics' position is that the artifacts visible in the video suggest otherwise, and that the disclosure of AI's role came after the prize rather than before the vetting.

Whether or not Nikon ultimately revokes the award, the case illustrates a gap that the 2010 guidelines on image handling could not have anticipated: what happens when a post-processing step does not merely adjust brightness or assign color to a measured channel, but uses a neural network to infer structure from ambiguous grayscale data? That is not a colorization question. That is a question about where measurement ends and invention begins.

A checklist for judging any scientific image

When you encounter a scientific image or video and want to assess what it actually shows versus what it claims to show, the following questions apply regardless of the field:

What is the raw data? Ask whether the raw acquisition file (unprocessed sensor output, unenhanced grayscale, unassigned channel data) has been made available or can be requested.

What was changed globally? Contrast adjustment, brightness scaling, and false-color assignment applied uniformly across an image are standard and generally accepted. They do not change which structures are present.

What was changed locally? Adjustments applied to specific regions, or rendering of areas that the imaging did not directly capture, require explicit disclosure and strong justification.

Did AI touch it, and at what stage? There is a meaningful difference between AI used to denoise or sharpen a complete acquisition and AI used to infer or render structures from ambiguous data. Ask which kind of processing was applied.

Was it disclosed before judging? Disclosure after a prize is awarded is not the same as disclosure during vetting. The sequence matters.

Do the structures behave as those structures should? Mitochondria that appear outside cells, nuclei that do not act as nuclei, structures with no biological analog: these are red flags whether or not AI was involved.

The raw grayscale file, if released, would answer the central question here in a way that no amount of subsequent explanation can. That is true of most disputes about scientific images, and it is the reason raw-data sharing policies exist.

The video of those cilia may yet prove to be exactly what Xu says it is: a real measurement, honestly processed, vividly presented. The way to know is to show the work from the beginning.

Ellis Ward writes Reporting from the Uncanny Valley, covering AI systems, research methods, and the gap between what scientific images show and what they claim to show.

Sources and further reading

PetaPixel, "Nikon Re-Reviewing Winner of Small World in Motion Contest After AI Accusation": https://petapixel.com/2026/10/01/nikon-re-reviewing-winner-of-small-world-in-motion-contest-after-ai-accusation/

USA TODAY, Nikon video contest AI coverage: https://www.usatoday.com/story/news/nation/2026/10/01/nikon-video-contest-ai/92038678007/

Gizmodo, "Contest-Winning Microscope Video Is Full of AI Confabulations, Scientists Say": https://gizmodo.com/?p=2000820652

Nikon Instruments via PR Newswire, "16th Annual Nikon Small World in Motion Competition Winner Captures Abnormal Movement of Airway Cilia in a Child with Primary Ciliary Dyskinesia": https://www.prnewswire.com/news-releases/16th-annual-nikon-small-world-in-motion-competition-winner-captures-abnormal-movement-of-airway-cilia-in-a-child-with-primary-ciliary-dyskinesia-302879601.html

BBC News, "Tiny image sparks big backlash in Nikon photo contest," September 30, 2026.