After weeks of re-review, Nikon stripped the 2026 Small World in Motion title from Dr. Ning Xu's cilia video for not complying with the contest's generative-AI rules, and promoted Nguyen Nam Nhat of Vietnam to first place. The dispute turned on how, and whether, AI visual processing was disclosed.
2026-10-09 · by Moe Ameen
Nikon disqualified the first-place winner of its 2026 Small World in Motion competition — the 16th edition of its microscopy-video contest — after weeks of re-review tied to generative-AI accusations. Reporting on October 9, 2026 confirmed the decision: the original winning entry, Dr. Ning Xu's video of abnormally beating airway cilia in a child with primary ciliary dyskinesia, was ruled to not comply fully with the competition's rules on AI-generated content. With Xu removed, Nguyen Nam Nhat of Vietnam was promoted to first place for a video of a tiny roundworm and a single-celled Dileptus, and other placements moved up.
The challenge came from the scientific community, not from Nikon's own vetting. After the win was announced in mid-September, microscopists flagged problems with the footage captured at roughly 100x: structures that appeared and disappeared between frames, cell features that didn't behave the way real cilia do, and — most concretely — reports that running the video through an AI tool surfaced an invisible provenance watermark consistent with AI-generated or AI-edited content. Nikon first said it was re-reviewing the entry against the technical documentation Xu supplied, then, after that review, moved to disqualify.
Xu did not concede that the footage was generated. He argued his use stayed within the rules, saying AI was used to distinguish and visualize features in his reconstructed grayscale images — in his framing, to improve the visual presentation rather than to fabricate an anatomical reconstruction. Nikon's position was narrower and procedural: the entry did not meet the competition's eligibility rules, and the ruling should not be read as a judgment on the entrant's professional reputation, scientific contributions, or intent. The exact extent of the AI involvement was never laid out publicly, which is part of why the case became a disclosure story rather than a clean "it was fake" story.
Nikon also signaled that the episode exposed a gap: the company said the event showed a need to revisit the rules and evaluation procedures for future entries, in a contest that had run for more than a decade without an incident like this. It lands alongside a pattern of similar reversals in image competitions, including a 2023 Sony World Photography Awards winner later revealed as an AI construct and a Hasselblad Masters 2026 finalist's entry pulled for AI use. Treat specific wording and the full timeline as evolving; confirm current details against Nikon's own competition resources.
What actually sank this entry was not that AI touched it — it was that nobody could cleanly say which part was AI and whether that use was declared. The defense ("AI only improved the presentation") and the accusation ("the structures are generated") were arguing about the same ambiguous pixels. That ambiguity is the real risk for anyone publishing AI visuals, and it is a workflow problem: if you can't point to exactly what was captured versus synthesized, you can't disclose it credibly when someone asks. [Kompozy](/) is built so that answer is never a guess. Every output is a typed format with a known provenance — a [Clipped Short](/glossary/clipped-short) is your real footage recut; a Persona Photo is a Gemini face-locked avatar image; a Photo Post is a gpt-image scene; a Persona VFX hook is fal.ai-generated and AI-ranked. You always know which pixels came from a camera and which came from a model, because the format decides it.
That makes disclosure a one-line habit instead of a forensic defense. Kompozy's per-post review gate in [Autopilot](/glossary/autopilot) is where you attach a consistent AI-disclosure line before anything publishes — and because the generation is governed by a [Persona Brief](/glossary/persona-brief) rather than an untraceable one-off prompt, you keep a record of how each asset was produced. From one source you fan that clearly-labeled output across the eight social platforms plus blog and email, with the same disclosure wording everywhere, so you are never caught the way this contest winner was: unable to show, after the fact, what the model did and when. For the provenance-signal side of this story, see our notes on [invisible AI watermarks](/news/anthropic-claude-invisible-watermarks) and the [rise of AI-detection tooling](/news/pangram-9m-ai-detection-funding).
After scientists questioned the winning cilia video and a re-review, Nikon ruled that Dr. Ning Xu's 2026 entry did not comply fully with the competition's rules on AI-generated content. Nikon framed it as an eligibility decision rather than a judgment on the entrant's reputation or intent. Xu argued his AI use was limited to visualizing features in reconstructed grayscale images, not fabricating them; the exact extent of the AI involvement was not laid out publicly, which is why the case centers on disclosure.
With Xu disqualified, Nguyen Nam Nhat of Vietnam was promoted to first place for his video of a tiny roundworm and a single-celled organism called Dileptus. Other placements moved up as a result.
That AI visuals are increasingly checkable by third parties — here an invisible provenance watermark was surfaced by an outside tool — and that unclear disclosure, not just outright fakery, is enough to get an entry pulled. The practical takeaway is to know and be able to state which parts of a visual were captured versus AI-generated or AI-processed, and to attach a consistent disclosure line before you publish rather than defending it after a challenge.
Keep a documented workflow where each asset's provenance is known, then add a plain-language label noting what was AI-generated or AI-edited. A content engine like Kompozy helps because every output is a typed format with known provenance — real footage, face-locked avatar image, generative VFX hook — so you can say exactly what the model did, add a consistent disclosure line in a per-post review gate, and carry the same wording across the eight social platforms plus blog and email.