On the screen, tiny filaments beat in rhythm, coloured green, red and violet. They are cilia from the bronchial wall of a child with a rare genetic disease, which prevents these cilia from properly clearing his lungs.
The video is magnificent. In September, it won first prize in the Small World in Motion competition, organised by Nikon, one of the most prestigious scientific imaging competitions in the world.
On 9 October, Nikon stripped it of its prize.
What happened
As soon as it was published, the video raised doubts among specialists in this disease. The structures shown did not look like what they usually observe in their patients. Some suspected the involvement of generative AI.
After re-examining the video and consulting its jury, Nikon concluded that it did not comply with the competition's rules on generative AI. The company stated that its decision did not constitute a judgment on the author's reputation, scientific work or intentions, and that it would review its rules for future editions.
First prize now goes to a Vietnamese researcher, Nguyen Nam Nhat, for a video showing a small worm and a single-celled organism.
The author's version
The researcher does not deny having used AI. But he maintains that it was used neither to make the video, nor to create the cilia, nor to invent their movement, which genuinely come from his instrument. According to him, AI only made it possible, after the fact, to distinguish and colour the different structures in originally black-and-white images. He thought he was staying within the rules.
The rules, however, prohibit AI-generated videos. That is the whole question: colouring and separating structures with AI — is that enhancing a real image, or starting to fabricate another one?
Scientific research has strict rules on images, well predating AI. Since the mid-2000s, major journals have for example required that any brightness or contrast adjustment apply to the entire image, and never to a single area, and that any enhancement be disclosed. The reason is simple: in science, an image is not decoration, it is evidence. Modifying it, even with good intentions, risks showing something that does not exist.
Why it matters
The Nikon competition rewards beauty as much as science. But the affair touches on a much broader question.
AI tools are increasingly used to clean up, enhance and colourise scientific images. This is often useful. But an AI that "improves" an image can also add plausible details that were not there. In a holiday photo, that is inconsequential. In an image of diseased cells, it can lead to false conclusions.
And science is already under pressure. The large online library arXiv has had to ration submissions, overwhelmed by texts written with AI. Images are the next front.
What we take away
It was specialists in the disease, not detection software, who spotted the problem. They saw that these cilia did not look like those of their patients.
This may be the most reassuring lesson of the story: faced with increasingly perfect images, the eye of those who truly know their subject remains the best detector. For everyday images, our five reflexes when faced with a dubious photo still hold.