Monday 3rd August 2026. Each week, between 2.30pm and 3pm, I talk to Sonja McLaughlan on Track Radio about technology and sport. Following a dramatic weekend of boxing at the Commonwealth Games, this week we talk about whether AI can judge a boxing match.
Dimeji Shittu vs Ankush Panghal
In Glasgow this last weekend, Dimeji Shittu was beaten by India’s Ankush Panghal in the gold medal bout. Shittu started off well, winning the first round unanimously. He was then deducted points by the referee for ducking his head too low and Ankush was judged to have won the last round, winning the bout overall.
Commentators weren’t happy, feeling that Shittu was by far the better boxer and should’ve won: the judges clearly disagreed. Given that scoring is so subjective, some have asked whether AI could replace judges, make it fair and far more transparent. Is AI up to the job?
This is my personal take.
iBoxer
If we’re ever going to use AI, then we’ll need data, and lots of it. Between 2008 and 2012, Dr Simon Goodwill from the Sports Engineering Research Group at Sheffield Hallam University developed a system for GB Boxing called iBoxer. It is now in its fifth Olympic cycle. Its task was to gather as much information as possible on all aspects of a fight so that tactical decisions could be made on how to approach a fight against known opposition. Video was always included.
One thing they created was a punch counter. Performance analysts watching a fight would track punches manually by pressing buttons: this allowed them to look at their total number, timing, the momentum of the fight and so on.
At their training base in Sheffield, GB Boxing also have a unique bird’s eye view of the ring which allows coaches to look at positioning, distance between the boxers and movement across the ring. All this information, combined with intelligence about every boxer they might encounter has helped GB Boxing to punch above their weight (apologies for the pun).
What would AI have to do to judge a bout?
Pass lots of exams for a start. But in terms of specifics, we need to first ask what does a judge does that AI would have to copy (and do better).
In an amateur bout, there are five boxing judges watching around the ring, all with a slightly different view. The simplest part of their task is to assess the number and quality of punches. But that’s the least of it. They also look at positioning, who is advancing, who is retreating, who is holding the centre of the ring. They try to figure out who is in control, which boxer has a clear strategy that appears to be working. If this sounds all a bit opaque, it’s because it is. A lack of transparency is one of the key criticisms of judging in boxing.
An AI system, then, would have to be video based, and would track all parts of the boxers’ bodies around the ring. It would assess boxers objectively with a set of pre-defined metrics about number of punches, punch quality, hit points, movement across the ring, momentum and control.
Keeping it simple: what is a good punch?
Quantifying these from video isn’t easy. Even counting punches is hard. Was there contact? Was it a glancing blow? Did it affect the opponent? I once asked a boxing coach what the definition of a good punch was.
“Ooomph,” he said, doing some sort of air punch with his fist.
“One with real power,” he added.
That was all I got. He knew instinctively what a good punch was from looking at hundreds of fights, but couldn’t really come up with a useable definition. In physics terminology ‘power’ isn’t even the right word (throughout the sporting world power is used as a proxy for force or impulse). An AI system processing video might be able to estimate impact forces using Newton’s laws much like Hawkeye and other systems use lift and drag forces to create trajectories of the ball around a tennis court or a football stadium. But calculating multi-body forces of boxers from video is a whole new ball game (apologies for another pun).
A different approach would be to ask lots of judges to assess the same videos of boxing matches, score them and use the results to train an AI system. But then we’re back to the original problem: judges aren’t consistent and the AI would inherit the inconsistencies.
An AI judge: Oleksandr Usyk vs Tyson Fury
And yet it appears to have been done. In 2024, an AI judge ran alongside three judges in a professional fight between Oleksandr Usyk and Tyson Fury. The only thing I could find out about the AI judge was that it gave the same result as the human judges (an Usyk win). There was no information on how it worked and no data came out apart from the final score.
My guess is that it was probably an automatic punch counter with statistics attached, although I’m happy to be proved wrong. The lack of transparency is really not helpful. We need to be able to trust systems like this and we can only do that with the evidence before our eyes.
How AI might help
GB Boxing have arguably the best boxing performance analysis system in the world (of course, I would say that). Often the boxer has to beat both the opponent and the judges, particularly in fights with a partisan crowd. Shittu had a game plan and knew exactly what he was supposed to do to win. He was magnanimous in defeat.
“Whatever the referee says goes”, he said.
“If I’m sitting here blaming judges, blaming refs,” he said, “I’m not going to get better.”
He and the coaches will look at the bout afterwards to see what went wrong. Given the immense amount of information in the iBoxer database, they might even be able to understand how much was boxer error and how much was judging bias. Training an AI model with the data might enable them to predict what to do to guarantee a win if a similar opponent, judges and referee came up again.
I should imagine, not losing two points for ducking his head will be a start.
Professor Steve Haake OBE 