On September 22, Nikon published a warm profile of the man who had just won its most prestigious video competition. It described Dr Ning Xu as a research fellow at the National University of Singapore who builds microscopes combining advanced optical techniques “with AI-assisted post-processing and visualization.”
Nine days later the company announced it was re-reviewing his winning entry, after scientists accused him of using AI in it.
That sequence is the most useful thing about this story, and almost nobody covering it has mentioned it. The use of AI was not a secret dragged out of anyone. It was in Nikon’s own publicity, written by Nikon, before the complaints started. What is actually in dispute is narrower and much harder: where processing ends and fabrication begins, and whether anyone checked.
The short version
- Dr Ning Xu won the 2026 Nikon Small World in Motion competition and a prize of £2,250, roughly 3,000 dollars
- The video shows cilia, the hair-like structures lining the airways, beating in a sample from a child with primary ciliary dyskinesia, a rare genetic disorder diagnosed by how cilia move rather than how they look
- Scientists flagged artifacts: structures appearing and vanishing between frames, and cilia at sizes biology does not produce
- One researcher reported a SynthID hit, Google’s invisible watermark, suggesting at least part of the video passed through a Google generative model
- Xu’s defense is specific. He says AI was applied after reconstruction, to distinguish and color structures in greyscale data, and that it did not generate the cilia or their motion
- The rules are blunt: AI-generated videos are not permitted
- Nikon is re-reviewing the original vetting plus new documentation from Xu. A spokesperson said that at this time the company does not see that any rules were violated
- His LinkedIn comments and then his profile disappeared, which did more damage to his case than any of the technical criticism
What the video was supposed to show
The subject matters, because it is the reason the entry was impressive in the first place.
Cilia are microscopic hairs that line the airways in their millions, beating in coordinated waves to push mucus and debris up and out of the respiratory system. In primary ciliary dyskinesia, that beating goes wrong, the airways cannot clear themselves, and the result is chronic lung, sinus and ear infections.
The diagnostic catch is that PCD cannot reliably be identified from a still image. The cilia often look normal. It is the motion that is abnormal, which means diagnosis depends on capturing extremely small structures moving extremely fast. Building an instrument that can do that is a genuine engineering achievement, and it is what Xu and his collaborators spent years on.
His own framing of the competition, given to Nikon before any of this blew up, was that “life is not a still picture. Life moves.” It is a good line, and the irony of where it ended up is not lost on anyone.
What the critics actually found
The objections came from inside the field, which is why they stuck.
Edward Phelps, an associate professor at Florida University, posted that there were “many serious problems with this video.” His specifics were not vibes. Purple structures “pop in and out of existence” between frames. Green cilia “appear from nowhere and do not match the known size.” Nuclei behave in ways nuclei do not. Scaling appears that “does not occur in biology.”
Andrew Moore, a former Small World in Motion judge, said the footage “pointed toward incompatible interpretations” and reached for an analogy that captures the discomfort better than any technical argument. It reminded him, he said, of old photographs restored with AI: “The high-res, colorised photo may look nice, but if it was your grandma who was face-swapped, it’s going to be unsettling.”
Then came the piece that turned a dispute into a story. Ian Donovan, an MD and PhD student at UT Southwestern Medical Center, ran the video through Google Gemini and said it came back flagged with SynthID, Google’s invisible watermark for content its models produce. If that result holds, at least some pixels in that video went through a Google generative system.
| The objection | Why it matters | Xu’s answer |
|---|---|---|
| Structures appear and vanish | Real objects do not blink out between frames. Generative infill does | Rendering “to improve the visual presentation” |
| Cilia at impossible sizes | Cilia dimensions are well characterized. Wrong scale suggests invented geometry | Not addressed directly |
| Nuclei behaving oddly | A model guessing at structure will get cell biology subtly wrong | Not addressed directly |
| SynthID watermark detected | Indicates a Google generative model touched the file at some stage | No public response |
| Scaling “does not occur in biology” | Goes to whether the footage represents a real measurement at all | Says the experimental movie and motion are real |
The line nobody had agreed on
Here is the part that makes this more than a cheating story.
Scientific imaging has always been processed. Nobody looks at raw sensor output. Microscopy data arrives as greyscale intensity values, and turning that into something a human eye can interpret requires choices: brightness, contrast, which channel becomes which color, deconvolution to undo the blur the optics introduced, denoising to pull signal out of a very dark frame. All of that is normal, all of it is published, and none of it is controversial.
What changed is that the best tools for several of those steps are now neural networks. ML denoising and ML upscaling are mainstream in microscopy because they work better than the classical methods. The trouble is that a model trained to remove noise is also, structurally, a model trained to guess what should be there. On good data that guess is close enough to be useful. On marginal data it starts inventing plausible biology.
Xu’s claim is that he sits in the second-to-last row. In his words, AI “was not used to generate the experimental movie, the cilia, or their motion.” It “was applied afterwards to the reconstructed grayscale data to distinguish and color structures with similar morphology,” and the regions beneath the cilia “were rendered mainly to improve the visual presentation.”
Read carefully, that last clause is an admission of something more than coloring. Rendering regions to improve visual presentation is not the same as assigning colors to data you already measured. Whether it crosses into generation depends entirely on what the renderer was given and what it produced, which is exactly the thing nobody outside the review can see.
The contest rules, meanwhile, leave no wiggle room on the extreme case. AI-generated videos are not permitted. Dr Patrick Hickey, who placed fifth with a video of mitochondria and chloroplasts in Turtle Vine leaf cells, said the prohibition was “quite clearly” among the rules.
Why the scientists are angry about the prize money least of all
£2,250 is not what this is about.
Hickey put the stakes plainly: “If you spend a lot of your time in a dark room looking down a microscope, it’s really nice to be able to show people and to show the public.” He described the appeal as “an interesting fusion between science and art, and a lot of the images that scientists are making have some very abstract and artistic qualities to them.”
That is the thing people in the field feel is at risk. Small World works because the images are astonishing and real at the same time. The astonishment is load-bearing, and it collapses if a viewer has to wonder whether they are looking at a measurement or a rendering. As one researcher put it on Bluesky, there is “something profoundly cynical and sad about using synthetic images to cheat in a contest that is supposed to be about how beautiful and cool biology is.”
The same erosion is visible everywhere else. The heuristics ordinary readers were taught for spotting fakes stopped working some time ago, which is why spotting AI generated images now depends far more on provenance than on looking harder. The consequences are no longer confined to hobbyist embarrassment either. When a synthetic clip of a strike on Iran’s Kharg Island circulated this year, the Pentagon had to publicly confirm the bombing never happened.
Competitions have turned out to be an unusually good stress test for all of this, because they put a prize on top of an unverifiable claim. Earlier this year an AI-judged contest with a six-figure purse handed its prize to an entry that gave the Cyclops three eyes. Different failure, same root cause: nobody had decided in advance what would count as correct, or who would check.
The detail that keeps this from being simple. Nikon’s own Masters of Microscopy profile, updated September 22, described Xu as building systems that combine optics “with AI-assisted post-processing and visualization.” In a separate interview he encouraged other researchers to “embrace AI to reveal the hidden beauty in your data.” He was not hiding a workflow. Either the vetting process read that and considered it acceptable, or it did not read it. Both answers are a problem for the competition rather than for him alone.
The self-inflicted wound
None of the above is why Xu is in trouble today.
He posted a defense on LinkedIn. The comments then became unfindable, and his profile appeared to have been deleted entirely. Whatever the reason, the effect is the one you would predict: a researcher who had a technical argument to make stopped making it in public at the exact moment people were waiting to hear it.
And the argument was winnable, or at least testable. Scientists have repeatedly said the dispute could be resolved in an afternoon if he released the raw greyscale footage. That file either shows cilia moving in the pattern the final video depicts, in which case the processing was cosmetic and he is vindicated, or it does not. There is no third outcome, which is why the raw data is the only piece of evidence that matters and the only one still missing.
Nikon, for its part, is doing the cautious thing. It is re-reviewing the original vetting along with the technical documentation Xu has supplied on his equipment, imaging methods and processing, and it says he is complying. A spokesperson’s position so far is that the company does not see that any rules were violated. That is a holding statement, not a verdict.
What to watch next
- Whether the raw greyscale video is released. This is the whole case. Everything else is inference
- What the SynthID hit actually came from. A watermark on a published file could come from a generative step in the pipeline, or from a later edit, or be a false positive. The provenance of the detection needs the same scrutiny as the video
- Nikon’s conclusion, and its reasoning. A quiet decision with no published standard would leave every future entrant guessing
- Whether the rules get rewritten. “No AI-generated videos” is unusable as written when ML denoising is standard equipment. The rule needs to name steps, not technologies
- Disclosure requirements. The obvious fix is mandatory processing disclosure plus raw data on request, which is what journals have been moving toward
- Whether Xu speaks again. He has a credible technical defense available and currently no public account
The bottom line
The most likely reading of the evidence is not fraud and not innocence. It is a researcher doing real work on a real instrument, using AI tools that his field has genuinely adopted, pushing the visualization further than the data could support, and entering it into a competition whose rules were written before anyone needed to distinguish between denoising and inventing.
The artifacts the critics found are hard to explain away. Structures that blink in and out between frames and cilia at the wrong scale are the signature of something filling in gaps, and “rendered mainly to improve the visual presentation” is a phrase that concedes more ground than it looks like.
But the failure is not only his. Nikon described his AI-assisted workflow in its own publicity and awarded him first place anyway. That means the vetting either approved it or never looked, and a competition that cannot tell the difference between measured and generated footage has a structural problem that disqualifying one entrant will not fix.
The fix is boring and available: say which processing steps are allowed, require entrants to declare them, and ask for the raw data. Microscopy already knows how to do this. It just never had to write it down before.
Sources and further reading
- PetaPixel: Nikon re-reviewing winner of Small World in Motion contest after AI accusation
- CNN: Winner of Nikon microscopy contest faces criticism over AI usage
- Nikon Small World: Ning Xu, Studying Disease through Motion
- The Scientist: Nikon announces Small World in Motion 2026 winners
- UNILAD Tech: Nikon launches investigation after contest winner’s admission
- Google DeepMind: How SynthID watermarking works
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
