World-class jurors, voting alone and blind, almost never agree on a photograph. Yet seventeen times in six years, most of them chose the same one. Chance agreement accounts for about one such moment. The crowd did not see these works coming; a machine judging blind did. We call them the Spark.
We asked what would happen if every juror kept exactly the number of picks they really made, but spread them at random. A juror who picked 40 works still picks 40. Under that rule, a work collecting 8 votes should appear about once in 1,700 years of finals. It appeared twice in six.
How to read it. Bars on a logarithmic scale: each step up is ten times more works. At 4 votes reality is already 2.4 times above chance; at 8 votes it is about 570 times above. In 20,000 simulated competitions, none produced 17 works with 6 or more votes; the most was 6. That puts the probability of the observed result arising by chance below 1 in 20,000 (p < 0.00005, permutation test1). This model treats all finalists as interchangeable. They are not: the jurors converge on something visible in the frames themselves, which is exactly what a machine judging blind picks up (Finding 3).
That is what makes the Spark stand out. Agreement among jurors is measured with Fleiss' kappa2: 1 means every juror voted alike, 0 means no more agreement than chance. In every final from 2020 to 2025, kappa stayed between 0.07 and 0.11, barely above chance.
This is not a weakness of the jury. Expert panels are often far less unanimous than their reputations suggest: in one well-known study, wine judges tasting the same wine three times rarely gave it the same score3. Against that background, a near-unanimous verdict is a signal.
If the Spark were just popularity, the audience would have seen it first. It did not. The public ratings of the seventeen works are scattered across the whole range, from the 22nd percentile of their final to the 100th, averaging 56 where pure chance gives 50. An independent witness saw something else. An AI judge scored every finalist blind, never told the votes. It placed all seventeen in the upper 40% of their final, and ten of them in the top 11%.
How to read it. Each dot is one of the seventeen Spark photographs, placed by its rank within its own year's final (100 = best of that final). Audience versus jury votes across all finalists: Spearman ρ = −0.05, no relation. AI score versus jury votes: ρ = 0.33, 95% CI 0.29–0.36. The AI's ability to tell Spark works from all other finalists, AUC4 = 0.84 (0.5 is a coin toss). Six runs under six different instructions each found them: AUC 0.75–0.84, every run p ≤ 0.004. The test was later repeated with the same model on Amazon Bedrock and reproduced within statistical error: AUC 0.79 with the original instruction (0.81 with a second one) on a 100-photograph pilot, against 0.85 for the original run on the same photographs.
The Spark is not one lucky season. It appeared in each of the six finals, under each year's jury, in a competition with no genre categories, where a landscape competes directly with a street photograph or a wildlife frame.
We do not claim to know what it is. We can only show where it is. On a handful of photographs, judges who usually disagree suddenly converge. Popularity does not predict it. And it is visible, it seems, even to a machine that has never met them. Some would call it talent, or mastery. Some would call it a spark of something higher. We simply call it the Spark, and we keep looking for it in every new entry.
None of this says a machine judges better than the jury. Where the work is already outstanding, the jury still leads, and no model replaces a world-class panel at that level. What the Spark shows is narrower and, to us, more interesting: when the best eyes converge, they leave a trace a model can detect.
Data. All 2,687 photographs that reached the FotoAward final round in 2020–2025 (2,686 were scored by the AI; one 2022 finalist with no votes is missing from the image set), with each juror's published ballot (10–11 jurors per final). Jurors vote independently and blind, seeing only the image and its title in random order. No names of jurors or authors were used.
Chance model. For each year, every juror's real number of picks was redistributed at random over that year's finalists, 20,000 times. Expected counts come from the exact Poisson-binomial distribution. Because it keeps each juror's own generosity, the model does not exaggerate the effect.
Agreement. Fleiss' kappa2 per final; Cohen's kappa5 for pairs of jurors.
The AI judge. One of the latest top frontier models (Model 1 — Anthropic's Claude Opus 5.5) scored every finalist blind, under a short instruction fixed in advance. Images were renamed at random, stripped of titles, mixed in batches of 20, and scored by fresh instances with no access to votes, names or the web. Separation was measured by AUC4, with bootstrap confidence intervals6 and permutation tests1. The replication ran the same model through Amazon Bedrock in us-east-1, under the identical blind protocol.
Audience. Public site ratings of the same works, ranked within each year's final.
Limits. Seventeen works is the entire population, so precision is limited. All data come from one competition. The AI recognised 1 of 375 test images when asked; prior exposure cannot be ruled out entirely.