For years, football has treated technology as a way to make refereeing fairer. Goal-line technology answered one clean question: did the ball cross the line? Video Assistant Referee technology, or VAR, promised to catch “clear and obvious” errors.
Semi-automated offside technology brought cameras, tracking data, connected , and digital player models into decisions that used to depend on the assistant referee’s eye.
Now the conversation has shifted. The question isn’t really whether technology belongs in sport anymore. It clearly does. The harder question is how much authority it should have once it gets there.
That makes sport a useful case study for one of the biggest questions in AI governance. When should humans remain in the loop? When is automation helpful decision support? And when does it start quietly reshaping who owns the decision?
Football feels like a strange place to test that question. But it’s also perfect. Every decision is public. Every delay is visible. Every mistake is replayed from 12 angles before anyone has had time to blink. There’s nowhere for bad governance to hide.
Why The Debate About AI Referees Is Bigger Than Football
The debate around AI referees isn’t just about whether football is becoming too technical. It’s about what happens when automated decision making enters a high-pressure environment where trust matters as much as accuracy.
That’s a familiar enterprise problem.
Businesses are already using AI systems to support decisions in security operations, fraud detection, compliance monitoring, customer service, hiring, finance, logistics, and business intelligence. In many cases, the system doesn’t make the final decision. A human does.
At least on paper.
But anyone who has worked with automated recommendations knows the line can blur quickly. A system flags a risk. A dashboard assigns a confidence score. A model recommends an action. The human still clicks approve, but the path has already been shaped.
Sport makes that tension easier to see because the emotional stakes are so immediate. A disallowed goal doesn’t sit quietly inside a quarterly report. It happens in front of a stadium, a broadcast audience, and millions of people on social media who all somehow become experts in geometry within 20 seconds.
The same basic governance questions still apply:
- Who made the decision?
- Can the decision be explained?
- Was the human meaningfully involved?
- What happens when the system gets it wrong?
- Do people trust the process, or only tolerate it?
That’s why AI in sport matters beyond the pitch. It gives enterprise leaders a live, public example of what happens when technology improves decision-making, but doesn’t automatically improve confidence.
What AI Is Actually Doing In Modern Officiating
Modern officiating systems are not all the same. Some technologies answer narrow, factual questions. Others support judgement calls. That difference matters.
Goal-line technology is the simplest version. It answers one binary question: did the whole ball cross the whole line? There’s very little interpretation involved. It’s fast, clear, and easy for people to understand.
VAR is different. It’s a review system. It allows officials to check potential errors around goals, penalties, red cards, and mistaken identity. The technology provides footage and angles, but the referee still has to interpret what happened.
Semi-automated offside technology sits somewhere between the two. FIFA’s 2022 system used 12 dedicated tracking cameras to follow the ball and up to 29 data points on each player, 50 times per second, to support offside decisions. The data points included limbs and extremities relevant to offside calls.
For the 2026 FIFA World Cup, FIFA has taken this further. Advanced Semi-Automated Offside Technology will be used at the tournament for the first time, with clear offsides sent directly to match officials on the pitch rather than only to the VAR team. FIFA says this is designed to speed up decisions and reduce injury risk during the moments between an offside offence and the flag being raised.
That sounds small, but it changes the decision flow. The system isn’t just helping someone review a decision after the fact. It’s starting to shape real-time officiating.
The Premier League’s version of semi-automated offside technology shows how data-heavy this is becoming. Its system uses up to 30 cameras installed around stadiums, with some capturing footage at 100 frames per second. These cameras help automate offside line generation and identify relevant body positions.
This is still not a fully automated referee. But it is a more automated decision environment.
And that’s the important distinction. The future of AI referees probably won’t arrive as a robot with a whistle. It’ll arrive as a growing stack of systems that track, flag, recommend, visualise, and accelerate decisions until the human role starts to look different.
Why Accuracy Alone Doesn't Create Trust
The strongest argument for officiating technology is accuracy. And to be clear, there’s evidence that it helps.
A study published in the Journal of Sports Sciences found that VAR improved decision accuracy in association football from 92.1 per cent to 98.3 per cent after intervention. The same study reported median review durations of 62 seconds for on-field reviews and 15 seconds for VAR-only reviews.
That’s a meaningful improvement. It would be lazy to pretend otherwise. But accuracy isn’t the whole trust equation.
The Football Supporters’ Association’s 2026 VAR survey found that 75.7 per cent of Premier League fans who responded did not support VAR in football. The same survey found that 91.7 per cent said VAR had removed the spontaneous joy of goal celebrations.
At first glance, that looks contradictory. If VAR improves accuracy, why do so many fans dislike it?
The answer is that people don’t experience technology as a spreadsheet of improved outcomes. They experience it as delay, confusion, uncertainty, interruption, and sometimes relief. If the decision is right but the process feels opaque or miserable, trust still erodes.
The comparison with goal-line technology is useful here. In the same FSA data, goal-line technology had 93 per cent support. That tells us something important. Fans are not automatically against technology. They’re against technology that feels unclear, slow, inconsistent, or too involved in subjective judgement.
That’s an enterprise AI lesson hiding in plain sight.
A system can be technically better and still fail the trust test. Accuracy helps. But trust also depends on speed, transparency, explainability, consistency, and whether people understand where human judgement still sits.
The Explainability Problem Hidden Inside Automated Decisions
People can accept decisions they don’t like. What they struggle with is decisions they can’t understand. That’s where explainable AI becomes more than a technical ideal. It becomes part of whether people believe the system is legitimate.
In football, offside is easier to explain than a subjective foul. A player is either in an offside position or they aren’t, although even that can become painfully close. A shoulder. A toe. Half a knee. The kind of detail that makes everyone wonder whether we’ve lost the plot slightly.
But fouls, handballs, interference, and obstruction are different. These depend on context. Intent, movement, advantage, contact, position, and interpretation all matter. The technology can show what happened. It can’t always settle what it means.
That’s why automation works better when the decision is narrow and rules-based. Once judgement enters the process, explainability becomes harder.
The same issue appears in enterprise AI systems. A fraud model may flag a transaction. A security tool may rank an alert as critical. A hiring system may score a candidate. A customer service platform may recommend a response. If the user can’t understand why the system reached that output, they’re left with two bad choices.
They can trust it blindly. Or they can ignore it. Neither is governance.
Good explainability doesn’t mean every user needs to understand the model’s internal mathematics. Most people don’t need a lecture on machine learning weights while they’re trying to do their job. Mercy, really.
But they do need to understand the reason behind the output at the level required to act responsibly. What data was used? What rule or pattern triggered the recommendation? How confident is the system? What uncertainty remains? What should the human check before accepting it?
That’s the difference between a useful decision-support system and a very confident black box.
When Humans Stop Making Decisions And Start Approving Them
Keeping a human in the loop sounds simple. It isn’t.
A human can be present without being meaningfully involved. They can review the output, but not understand it. They can approve a recommendation because the system looks more certain than they feel. They can become the person who signs off the machine’s decision after the real decision has already been made.
That’s the risk of automation bias.
The European Data Protection Supervisor has warned that automation bias is linked to the mistaken assumption that a decision is not automated if a human supervises the process. That assumption can create unrealistic expectations and weak human oversight.
This matters because “human-in-the-loop” can become a comforting phrase that doesn’t describe reality.
In sport, the risk is that officials start to defer to the system because the system appears more objective. In enterprise environments, the same risk shows up when analysts, managers, or operators accept AI outputs because challenging them takes more effort, more confidence, or more time.
That doesn’t mean automation should be avoided. It means oversight has to be designed properly.
A human who rubber-stamps an AI output is not the same as a human who can question it. Meaningful human oversight needs authority, context, training, time, and a clear escalation path. Otherwise, the organisation hasn’t kept a human in the loop. It has kept a human near the loop and hoped that counts.
It doesn’t.
What Other Sports Can Teach Us About Human Oversight
Football isn’t the only sport working through this. Other sports have built different models for how technology and human judgement should interact. The useful comparison isn’t which sport has the best technology. It’s which sport has the clearest governance model.
Automated ball-strike challenges in baseball
Major League Baseball introduced the Automated Ball-Strike Challenge System in 2026. The system tracks the exact location of each pitch against the batter’s strike zone, but it doesn’t simply replace the human umpire. Players can challenge a ball or strike call they believe was wrong, and the result is shown almost immediately to fans in the stadium and viewers at home.
Each team is given a limited number of unsuccessful challenges. If the challenge is correct, the team keeps it.
That structure matters.
It keeps the human umpire central. It gives players a controlled way to question a decision. It limits disruption. It makes the outcome visible. It also creates a clear trigger for when automation enters the process.
This is not full automation. It’s governed escalation.
For enterprise leaders, that’s the interesting part. Not every AI system needs to act automatically. Some systems are better designed as challenge mechanisms, escalation layers, or second opinions.
Cricket's decision review system
Cricket’s Decision Review System, or DRS, is another useful model. The International Cricket Council describes DRS as a technology-based process that assists match officials with decision-making. On-field umpires may consult the third umpire, and players may request that the third umpire review an on-field decision.
Again, the structure is the lesson.
There are clear roles. The on-field umpire has authority. The third umpire has review responsibility. Players have a defined challenge route. The system supports the decision, but it doesn’t float around as an invisible authority with no owner.
That kind of design is exactly what many enterprise AI deployments lack.
A business may introduce AI into a workflow without clearly defining who can challenge it, who can override it, who explains it, and who is accountable when it causes harm. Then everyone acts surprised when accountability gets foggy. It’s not ideal. It is, however, very common.
What enterprise leaders should notice
The best sports technology systems don’t just add more data. They define how that data is allowed to influence the decision. That’s the difference between automation as a tool and automation as a power shift.
In baseball, automation enters through a challenge. In cricket, it enters through review. In football, semi-automated offside technology is moving closer to direct real-time signalling. Each model creates a different relationship between human judgement and machine assistance.
Enterprise AI needs the same level of design. Before adopting automation, leaders need to decide what role the system plays:
- Does it detect?
- Does it recommend?
- Does it rank?
- Does it escalate?
- Does it approve?
- Does it act?
Those are not small differences. They define where accountability lives.
The Real Enterprise Question: Who Owns The Decision?
The deeper question behind AI referees is not whether the technology works. It’s who owns the decision once the technology is involved.
That question is now central to responsible AI.
The National Institute of Standards and Technology’s AI Risk Management Framework defines trustworthy AI through characteristics such as validity, reliability, safety, resilience, accountability, transparency, explainability, interpretability, privacy enhancement, and fairness. It also says these characteristics need to be balanced based on the context in which the AI system is used.
That last part matters. Context changes everything.
An AI system used to recommend a playlist is not the same as an AI system used to flag a fraud case, deny a loan, prioritise a security incident, or influence a hiring decision. The more serious the decision, the more carefully the organisation has to define oversight and accountability.
The OECD AI Principles also frame trustworthy AI around human rights, democratic values, and responsible stewardship. They were adopted in 2019 and updated in 2024 to reflect new technological and policy developments.
The EU AI Act takes this further for high-risk systems. Article 14 says high-risk AI systems should be designed so they can be effectively overseen by natural persons while they’re in use. Human oversight should help prevent or minimise risks to health, safety, or fundamental rights.
Sport is not the same as regulated enterprise AI. A disputed offside call and a denied mortgage are not equivalent. But the governance pattern is similar enough to be useful.
Once automation enters a decision process, responsibility can scatter. The official blames the system. The system provider blames the configuration. The organisation blames the process. The process points back to the human who clicked approve. That’s not accountability. That’s a meeting invitation.
Enterprise leaders need to define decision ownership before automation becomes embedded. That means being clear about the difference between assistance, recommendation, approval, and autonomous action. Those categories should not be treated as technical details. They’re governance decisions.
How Much Automation Is Too Much?
Automation goes too far when it removes human judgement from decisions that still require context, interpretation, ethics, or accountability.
That doesn’t mean humans need to make every decision manually. They shouldn’t. Some decisions are better handled by systems because they’re fast, factual, repetitive, or too data-heavy for people to process well.
Automation makes sense when:
- The decision is narrow and rules-based.
- The system’s output can be verified.
- The cost of delay is high.
- The risk of subjective interpretation is low.
- The human role is clearly defined.
- The process can be explained after the fact.
That’s why goal-line technology works so well. It answers a narrow factual question quickly. There’s very little theatre around the decision. No one needs a five-minute philosophical debate about whether the ball felt like it crossed the line emotionally.
But automation becomes risky when:
- The decision depends on context.
- The output is hard to explain.
- Humans don’t have time to challenge it.
- People are likely to defer to the system.
- Accountability is unclear.
- The system affects trust, fairness, safety, or rights.
That’s where enterprises need to slow down. Not because AI is bad, but because AI changes the shape of responsibility.
The goal shouldn’t be to automate as much as possible. That’s a lazy strategy dressed up as innovation. The better goal is to place automation where it improves speed, consistency, and visibility, while keeping humans where judgement actually matters.
That distinction is going to become more important as agentic AI systems become more common. A dashboard that recommends an action is one thing. A system that takes action on behalf of a team is another. The governance model has to mature with the level of autonomy.
Sport is already showing us what happens when that maturity lags behind adoption. People may accept the outcome. They may even admit it was technically correct. But if the process feels distant, unclear, or unaccountable, trust still weakens.
Final Thoughts: The Goal Isn't To Remove Humans From Decisions
The rise of AI referees shows that accuracy is only one part of responsible automation.
Football has better technology than ever. More cameras. Better tracking. Connected . Faster offside decisions. More detailed review systems. Yet the trust question hasn’t gone away. In some cases, it has become sharper.
That’s the lesson enterprise leaders should take seriously.
The organisations that use AI well won’t be the ones that automate the most decisions. They’ll be the ones that understand which decisions can be automated, which should be supported, and which must remain meaningfully human.
Because once a system becomes good enough to influence an outcome, the real question changes. It’s no longer “Is there a human in the loop?” It’s “What is the human still there to do?”
For leaders working through the same questions around AI governance, operational trust, and responsible automation, EM360Tech keeps tracking the technologies, frameworks, and real-world lessons shaping enterprise decision-making.
Comments ( 0 )