The Fake That Won a Science Prize — and the Twenty Years That Caught It

In September, Nikon’s Small World In Motion competition named its first-place winner: a short microscopy video of tiny, hair-like structures called cilia beating inside the airway of a child with primary ciliary dyskinesia (PCD), a rare genetic disorder. The entry came from Dr Ning Xu, an optical engineer at Tsinghua University. To most eyes it looked like a window into a disease most people have never heard of. That is exactly what made it valuable, and exactly what made it dangerous.

The competition has run for fifteen years and is reserved for images and video taken through powerful microscopes; first place carries a $3,000 prize. Cilia are the hair-like projections that line the airways and sweep mucus out of the lungs, and in PCD they beat abnormally. That abnormal beat is the diagnostic signal. A video of it is not decoration. It is data about a real child’s real condition.

When Nikon posted the winning video, scientists started commenting on its LinkedIn page. Edward Phelps, an associate professor at the University of Florida, said there were “many serious problems with this video” and that “the microscopy community deserves an explanation from Nikon.” A PhD student at UT Southwestern said the source image appeared to carry a watermark indicating it had been AI-generated. Two former competition judges said they had concerns. And Robert Hirst, the lead scientist at the University of Leicester’s NHS centre for PCD diagnosis, said he had known immediately that the video was fabricated.

I have been diagnosing PCD patients for 20 years by examining their cilia speed, structure, waveform, length, cell size and shape. I knew immediately that the video was AI generated and put my reputation on the line by being the first to publicly say so on LinkedIn.

— Robert Hirst, lead scientist, University of Leicester / NHS PCD centre, to The Scientist

Hirst told the BBC the affair had “impacted a lot of patients, scientists, doctors and PCD support networks around the world.” That is the part that separates this from an ordinary photo contest. A fake in a beauty competition is a lie about aesthetics. A fake in a medical context is a corrupted record of an illness, and once it is out there it travels.

Nikon defended it first

Nikon’s first position was that nothing was wrong. “At this time, we don’t see that any rules were violated,” it said in late September, after Xu acknowledged on LinkedIn that AI had been used in the image processing while insisting he had stayed within the rules. Then, on 9 October, after “thoroughly re-evaluating the video and supporting materials and consulting with members of our judging panel,” Nikon reversed itself.

After thoroughly re-evaluating the video and supporting materials and consulting with members of our judging panel, Nikon has determined that the winning video did not comply with the competition rules regarding generative AI. Nikon has adjusted the competition rankings to reflect this decision.

— Nikon Instruments Inc., official statement, 9 October 2026

Nikon was careful about what the decision was not: it was “based solely on the submitted video’s eligibility under the rules outlined for this competition,” and “should not be interpreted as a judgment of the entrant’s professional reputation, scientific contributions, or intent.” First place went instead to Nguyen Nam Nhat of Vietnam, for a video of a tiny roundworm and a single-celled organism exploring their microscopic world.

The line Xu drew — and why it cannot be drawn

Xu did not deny using AI. He drew a line. “AI was not used to generate the experimental movie, the cilia, or their motion,” he wrote on LinkedIn. It had been applied afterwards, he said, “to distinguish and color structures with similar morphology” in the reconstructed grayscale data.

Read that sentence carefully, because it is where the whole problem lives. “Distinguish and color structures” in data a microscope already recorded is, on its face, the same family of operation as the enhancement every microscopy lab already does — denoising, reconstruction, brightening, coloring. The contest even permitted artificially bright colors, because staining a sample is standard practice. Xu’s argument was that he had enhanced real data rather than generating fake data, and enhancement was within the rules.

But a model asked to distinguish and visualize features in reconstructed grayscale images is a model that fills in structure where the data was ambiguous. The scientists who objected said the cilia were too big, that structures “pop in and out of existence,” that cilia “appear from nowhere.” Those are not the fingerprints of enhancement. They are the fingerprints of a model inventing plausible detail. The point is not that Xu set out to cheat. The point is that no rule can cleanly separate the two cases, because they run on the same software.

Why no detector caught it

The revealing detail is that nothing automated caught this. The people who flagged it were specialists who happened to own the exact domain — a PCD diagnostician, a cell biologist, an imaging researcher — each able to tell because they had spent years staring at the specific thing the video claimed to show. Hirst could see it because he has spent twenty years diagnosing PCD by looking at cilia. His expertise is narrow in the way all real expertise is narrow, and it was exactly narrow enough to be the one thing that worked.

That is the uncomfortable lesson. The judging panel included scientists, and still crowned the fake, because nobody’s expertise covers everything, and a fabricated image looks flawless to anyone who has not spent years inside the specific detail it depicts. Detection by expertise is not a system. It is a lottery in which the right person has to happen to look. It worked here because scientists have a culture of looking closely and speaking up — it worked slowly, publicly, and by accident of who follows Nikon on LinkedIn.

The honest counter-argument

The strongest case against reading too much into this is that AI is genuinely transforming microscopy for the better. Denoising, super-resolution and reconstruction tools let scientists see things their sensors recorded but could not resolve, and a contest that banned all of that would be the poorer. The researchers who pushed for the disqualification were careful to say so. Christophe Leterrier, a neuroscientist and former Nikon Small World jury member, called the affair the start of a conversation rather than the end of one.

This should not be seen as the end of a controversy but the start of discussion on how AI is transforming microscopy.

— Christophe Leterrier, neuroscientist, CNRS–Aix Marseille University, and former Nikon Small World jury member

Melanie White, a cell biologist at the University of Queensland, put the goal precisely: “The aim is not to restrict innovation, but to harness AI while protecting data integrity, scientific reproducibility and public trust.” That is the right position, and it is also the hard one, because it asks institutions to do something much more difficult than ban a tool. It asks them to verify how images were made — on every image, forever.

What this means for AI image tools

This generalizes well beyond microscopy. Any AI image tool that “reconstructs,” “enhances” or “cleans up” a picture is doing work on the same spectrum Xu described, and any output that carries a factual claim — a medical image, a product shot, a documentary frame, a scientific figure — now carries a question its pixels cannot answer: was this recorded, or was this filled in? The only defense is process: a record of what was recorded and what was added, disclosed up front rather than discovered later.

We test and review AI image tools on this site, and the Nikon affair is the clearest demonstration yet that the detector is not the answer. The detection system that worked here was a human being with twenty years of a very specific kind of looking, and a community that refused to look away. That culture is slower, patchier and more human than a detector would be. It is also the only thing that caught this one.

The contest that could not check itself

The competition assumed it could tell real from fake by looking at the entry. It could not, because the boundary it needed no longer exists in the tools people use. What finally moved the prize was not a better way of seeing the image. It was a better way of asking how the image was made: one specialist’s voice, a community’s insistence, and a company’s decision to listen.

Nikon has said it will revisit the rules and evaluation procedures for future competitions. That is the right move, but the deeper shift is cultural. Rules written for a world where “real” and “fake” were a binary are not going to survive contact with tools that erased the line. The next contest that crowns a fake will be the one with no twenty-year specialist in the room, and no culture loud enough to correct it. The image can no longer vouch for itself. The making of it has to.

We use AI tools in our research and drafting on this site. Every judgement and every recommendation is a person's.