In the 103rd minute of a World Cup knockout match, Croatia scored a goal that appeared to keep its tournament alive. The goal was initially awarded. After a Video Assistant Referee review, however, referee Espen Eskås overturned it. In his stadium announcement, heard in the official broadcast, Eskås said: “Croatia player number 20 touched the ball and headed to the goal scorer. Final decision, offside.”
The announcement did not explain the ruling. Mario Pašalić was not the goalscorer; he flicked the ball to Joško Gvardiol, who scored. More importantly, the replay raised two separate questions. The ball passed close to Igor Matanović’s head, although any contact was imperceptible to many viewers. Portugal’s Renato Veiga then bent at the waist and clearly headed the ball before Pašalić played it. Under soccer’s offside law, the legal effect of Veiga’s header depended on whether it was a deliberate play or a deflection.
That was a judgment for the match officials. Fans may continue to debate whether they made it correctly. This article does not attempt to settle the call. Controversial decisions, and unsatisfying explanations of them, are part of World Cup history.
What happened next created a different problem. FIFA Media tried to validate Eskås’s ruling by invoking Connected Ball Technology and FIFA’s semi-automated offside technology, or SAOT. In a July 2 post, FIFA said the ball data had “proven” Matanović’s contact and had allowed the referee to “correctly determine offside and disallow the goal.” A July 3 follow-up cited additional “heartbeat graphic” images.
Those statements did not validate the ruling. The technology could provide information about contact and player position, but it could not decide whether Veiga deliberately played the ball. By presenting one sensor reading as proof that the entire ruling was correct, FIFA Media made a carefully limited decision-support system appear to be the decision-maker. The response damaged confidence in technology whose published design gives little reason for that loss of confidence.
What the Technology Actually Does
FIFA’s system begins with two sources of data. The connected match ball contains an inertial measurement unit that records the ball’s movement 500 times per second. A change in that data can mark the moment the ball was kicked or touched. The sensor records what happened to the ball; it does not, by itself, identify the player who caused the movement.
Dedicated cameras provide the second source. They track the ball and relevant points on each player’s body. SAOT combines the ball and player-tracking data to identify a possible contact point, locate the players at that instant and generate proposed offside information.
That is the automated part. FIFA’s published procedure then requires video match officials to check the proposed kick point and offside line. If they disagree with either, they may select the kick point or draw the line manually. FIFA describes SAOT as a “support tool” intended to help officials make faster, more reproducible and more accurate decisions.
The design is sensible because offside involves more than geometry. Under Law 11, being in an offside position is not itself an offence. A player must then become involved in active play. Law 11 also distinguishes a defender’s deliberate play from a rebound or deflection. A deliberate play generally resets the analysis; a deflection does not.
FIFA’s own 2026 testing manual states this limit directly. In a test involving contact by a defender, the system is expected to recognize the contact without classifying it as a deliberate play or a deflection. That classification remains for the officials. It is not a defect in SAOT. It is a boundary around the question the system was designed to answer.
The IFAB VAR protocol draws the same division. Video officials may provide information and recommend review, but the final decision is always taken by the referee. Factual questions such as player position are usually handled through a VAR-only review. Questions requiring judgment, including some offside determinations, ordinarily call for an on-field review. Law 5 likewise makes the referee responsible for decisions concerning play.
Because Croatia’s goal was initially awarded, the VAR protocol also required a clear and obvious error before that decision could be changed. Sensor and tracking data unavailable to viewers could contribute to that finding. But the protocol does not say that an SAOT alert or connected-ball reading automatically establishes clear and obvious error.
What FIFA Media Got Wrong
FIFA Media’s July 2 post moved directly from a limited technological claim to an endorsement of the entire ruling. It said the connected-ball data had “proven that contact was made” by Matanović, thereby allowing Eskås to “correctly determine offside and disallow the goal.”
Even if the data established Matanović’s contact, it established only one part of the sequence. The officials still had to determine when the contact occurred, where Pašalić was at that instant, how Veiga’s intervening header should be classified and whether Pašalić then committed an offside offence. Connected Ball Technology could provide information relevant to the first question. Player tracking could address the second. Neither system could answer the third.
FIFA Media’s language erased those distinctions. The phrase “correctly determine offside” suggested that the technology supplied everything necessary to decide the incident. FIFA’s technical materials say otherwise. They describe measurements, an automated proposal, manual validation and a final decision by the referee.
The July 3 follow-up did not repair the problem. It said that additional heartbeat images showed touches by both Matanović and Veiga. But Veiga’s contact was never in doubt. Viewers could see it. The disputed question was whether his header was a deliberate play or a deflection, and the ball sensor could not decide that question.
The graphic also did not demonstrate FIFA’s attribution of the disputed signal. FIFA did not publish a continuous timeline aligning the sensor readings with the video, identify the relevant timestamps or explain how each signal was matched to a player. A still image of Matanović near the ball and a separate heartbeat trace did not, on their face, establish that the pictured event caused the displayed signal.
That omission does not show that the technology was wrong. It shows that the post did not prove what FIFA claimed. If FIFA wanted the heartbeat data to answer the controversy, it could have aligned the signals attributed to Perišić, Matanović and Veiga with the corresponding video frames. Even then, it would still have needed to identify Veiga’s deliberate-play question as a separate judgment made by the officials.
The posts also omitted the basic account of the review. They did not say whether SAOT generated the initial alert, whether the video officials accepted or changed its proposed kick point or line, what information they communicated to Eskås, whether Eskås saw the heartbeat data, or why Veiga’s header was treated as a deflection. FIFA possessed the system data, the video-operation-room communications and the referee’s account. Its public explanation supplied none of that chain.
The result was backwards. Instead of explaining how a human referee used a support tool, FIFA Media used the tool to certify the referee. Fans who disputed the ruling were invited to judge SAOT as though it had decided the entire incident. A well-designed support system began to look like an opaque and unreviewable decision-maker.
What a Useful Explanation Requires
The public controversy was not a formal appeal. FIFA nevertheless chose to invoke an automated system to defend a disputed result. Once it did, a general description of the technology was no longer enough. FIFA needed to explain how the technology was used in this decision.
That distinction appears in broader transparency standards. The OECD AI Principles call for understandable information about a system’s capabilities and limitations, as well as information that allows people affected by an output to challenge it. The European Union Artificial Intelligence Act separately addresses general instructions for interpreting system output and explanations of the system’s role in a particular decision. Those provisions did not govern this match. They are useful here because they identify the same practical distinction: describing a system is not the same as explaining a result.
When an organization invokes automated evidence to defend a disputed decision, its explanation should answer four plain questions:
- What did the system measure or report?
- Who reviewed the output, and did the reviewer accept or change it?
- What other evidence did the reviewer consider?
- What remained for a person to decide, and what did that person decide?
This does not require disclosure of source code or every data point. It requires an account of what happened. The need is greatest when the system detects something people cannot perceive or reproduce for themselves. A viewer who cannot see Matanović touch the ball cannot test a 500-hertz sensor reading. FIFA must connect the reading to the event, the event to the review and the review to the ruling.
What FIFA Should Have Said
FIFA did not need to persuade every viewer that Eskås was right. Nor did it need to concede that he was wrong. It needed to state what the technology supplied and what the officials decided. If the review followed the process described in FIFA’s published materials, an adequate explanation could have said:
The connected ball recorded a signal that FIFA attributed to contact by Igor Matanović. Player-tracking data placed Mario Pašalić in an offside position at that moment. The video match officials reviewed the sensor data, the player-position data and the available video before referring the incident to the referee.
The referee determined that Matanović touched the ball and that Renato Veiga’s contact was a deflection rather than a deliberate play. On that basis, the referee ruled that Pašalić committed an offside offence and disallowed the goal.
If that was not the process, FIFA should have described what actually occurred. If the video officials decided the factual questions and reported them to Eskås, FIFA should have said so. If Eskås did not see the heartbeat data, FIFA should not have implied that the data allowed him to reach his decision. And any explanation should have stated plainly that the connected ball detects ball movement but does not decide whether a defender deliberately played it.
Such an explanation would not end disagreement about the call. It would put the disagreement in the right place. Viewers could question the referee’s judgment without being led to believe that SAOT had made a judgment it was neither designed nor authorized to make.
