Nikon will redraw its AI rules after disqualifying its microscopy video winner

Robert Hirst did not need software to see that something was wrong with the winning video of cilia in a child's airway ^[2]. "I have been diagnosing PCD for 20 years by examining their cilia waveform, length, cell size and shape," the University of Leicester scientist said ^[13]. He said he "knew immediately that the video was fabricated" ^[12]. Hirst is lead scientist for the NHS centre for PCD diagnosis at the university. He said the cells and cilia "look nothing like those from PCD patients" ^[14].

The same processing step now has three names. Xu described it as AI-assisted post-processing "to distinguish and visualize features in the reconstructed grayscale images from the super-resolution optical imaging" ^[5]. Nikon ruled that the entry "did not comply with the competition rules regarding generative AI" ^[1]. Hirst's word was fabricated ^[12]. According to the BBC, Xu believed he had acted within the rules and cooperated with Nikon's investigation ^[6]. Neither report quotes the wording of the rule.

The expert check that decided the outcome happened after the result was public. Nikon said it began reviewing the video following skepticism online about its authenticity ^[3]. It reached its decision "after thoroughly re-evaluating the video and supporting materials and consulting with members of our judging panel" ^[4].

Nikon's promised fix covers "rules and evaluation procedures for future competition entries" ^[8]. I think the second item matters more. A sharper rule would have told Xu where the line sat before he entered. A different procedure would have put someone with Hirst's training in front of the winner before it was announced. Nikon also said the decision "should not be interpreted as a judgment of the entrant's professional reputation, scientific contributions, or intent" ^[9]. So Nikon has ruled on eligibility only. Hirst's question, whether these cilia look like PCD cilia at all, is outside its verdict.

Two groups use a rule like this: entrants deciding what to declare, and judges deciding what to check. Hirst named the people downstream of both. "This controversy has impacted a lot of patients, scientists, doctors and PCD support networks around the world," he said ^[15].

Anyone running a program that accepts processed images can place it on two axes. The first is how the rule is written. It can name a tool category, such as generative AI, or it can name an effect: any step that adds or infers structure the instrument did not record has to be declared, with the raw capture attached. The second axis is timing. A specialist in the subject either checks the entry before the award or sees it after it is public. Nikon's microscopy contest, now 15 years old ^[11], ended up in the box with a tool-label rule and an after-publication check.

The box to aim for is an effect-based rule with a pre-award check. It has a cost. Judges need a specialist for each subject, and entrants have to hand over raw files and a step-by-step processing log, which some will not want to share. One test settles it. If a program cannot name the specialist who checked last year's winner against its raw capture before the announcement, it is in the same box Nikon was in.