Prediction models work well where the target is measurable. A football result is measurable; so is a horse race. A dance contest result is a judged opinion combined with a public vote, and no amount of data changes that. This article sets out what a model can genuinely learn from such an event, what it cannot, and why "AI prediction" claims in this category deserve more scepticism than most.
What a model can actually learn here
Judging panels are more predictable than they look, and that is the part a model handles well. Given a few seasons of published scoresheets, it can pick up real regularities:
- Which dance styles score above average, and by roughly how much.
- Individual judges' tendencies — who scores tightly around the middle, who uses the full range.
- The effect of running order on scores, which is measurable and consistent across formats.
- Score trajectories: contestants improving week on week versus those who peaked early.
None of this is exotic. It is ordinary regression on a small, clean dataset, and it produces a defensible estimate of how the panel will score a given performance.
Where it stops
The panel is only half the result. The other half is a public vote, and the public vote has no training data: totals are usually never published, only the outcome. A model therefore cannot learn the function it most needs. It can proxy it with popularity signals — social media following, screen time, press coverage — but those are correlates of a correlate, and they break in exactly the situations that decide contests: a sympathetic story after an injury, a backlash, a viral moment in week six.
There is a deeper problem. A model trained on past results learns the judging panel of those seasons. Panels change, formats change, the scoring scale is adjusted. When the thing generating the labels is itself a set of people making subjective calls, the model inherits their habits rather than discovering a hidden truth about dancing.
How to read a prediction claim
If a site advertises model-based predictions, four questions settle whether there is anything behind it:
- What is the sample? A contest produces a few dozen eliminations per season. That is a small dataset by any standard, and small datasets support simple models only.
- Is the record published in advance? Predictions listed after the fact prove nothing. A verifiable claim is timestamped before the broadcast.
- Is it measured against the odds or against chance? Beating a coin flip is trivial; beating the bookmaker's price after margin is the only result that means anything.
- Does the model handle the vote at all, or does it only predict the judges' scores and quietly leave the harder half out?
Most published "AI predictions" in entertainment categories fail the second and third questions. That does not make them fraudulent — it makes them unmeasured, which is a different problem with the same practical consequence.
What machine learning is actually used for in this industry
Away from the marketing, the techniques are real and mostly invisible to players:
- Odds compilation. Operators price thousands of markets automatically and adjust to incoming money in real time. The model here predicts betting behaviour, not results.
- Risk and fraud detection. Unusual patterns of stakes across accounts are flagged for review — a classification problem with plenty of data.
- Responsible gambling monitoring. Several regulators now expect operators to detect markers of harm automatically: rising deposit frequency, chasing losses, play at unusual hours.
- Personalisation. Which game is promoted to which player, which is a recommendation system like any other.
Notice the pattern: these work because the target is well defined and the data is plentiful. Predicting who wins a dance contest is the opposite case on both counts.
A reasonable position
Treat a model's output on a judged contest as one input among several, weighted by how much of the result it can actually see. A tool that estimates the panel's score is doing something real; a tool that claims to predict the final outcome is claiming to predict a public vote it has never observed. The arithmetic in our piece on how these markets are priced matters more for a practical decision than any model output, because the margin is knowable while the vote is not.
Frequently asked questions
Can AI predict the winner of a dance contest
It can estimate how the judging panel will score a performance, which is a genuine and measurable task. It cannot predict the public vote, because vote totals are generally not published and there is no training data. Since the vote decides many eliminations, the overall prediction is far weaker than the judging component suggests.
Why are judged sports harder to model than football
Because the label the model learns from is a human opinion rather than a measured outcome. The model reproduces the habits of those particular judges, and when the panel or the format changes, what it learned no longer applies.
How can I tell whether a prediction service is measured
Check whether predictions are published before the event with timestamps, and whether performance is reported against the bookmaker's odds rather than against chance. Without both, there is no way to distinguish skill from selective reporting.
Where is machine learning genuinely used in betting
In odds compilation, fraud and risk detection, responsible gambling monitoring and personalisation. These work because the targets are well defined and the datasets large — the opposite of a judged contest with a few dozen results per season.