Who taught the World Cup to see?

A striker breaks. A defender holds the line. Two centimetres separate a goal from an offside flag.

At the 2026 World Cup, that call comes back in seconds. Drawn to the centimetre. Live, on the stadium screens.

Watch it land and you reach for the obvious answer: a machine made the call.

It didn't. Not really.

The line you see is the end of a long chain. And the chain doesn't start with a sensor. It starts with people.

Before any camera could find a player, a person had to teach it what a player looks like. Before any system could map a pitch, a person had to mark where the lines are. Frame by frame. Point by point. By hand.

The most advanced sports AI on Earth runs at this World Cup. Under it sits a foundation almost nobody talks about — millions of human-labeled examples.¹

Here is the hidden story: who taught the World Cup's AI to see?

How the machine learns the pitch and players

In the Annota8 Annotation Engine, a person marks the fixed landmarks first. Touchlines. Goal. Penalty box. Centre circle. The halfway line.

Those marks tie a flat broadcast frame to the real, measured pitch. State-of-the-art soccer calibration defines 57 geometric keypoints across 26 labeled marking classes — humans place them by hand.²³

The machine does not see a field. People define the geometry first.

Then come the players. A box around each player teaches detection — this is a body, here, in this frame. Then keypoints for posture: head, shoulders, hips, feet. Those are the limbs that decide an offside.

One angle is not enough. So the same player gets boxed from every view — broadcast high, behind the goal, tactical wide. Box the same body in each frame and a flat image becomes a 3D position. At a 2026 stadium, 16 optical cameras watch at once.⁴

This is illustration, not FIFA's pipeline. But the lesson holds everywhere. The open SoccerNet dataset alone carries 9.37 million hand-placed pitch-line points and 2.36 million labeled player positions.¹

The ball that broke the cameras

Cameras are built for bodies, not bullets. A struck ball can move faster than the eye and faster than the frame. It blurs. It shrinks to a few pixels. A leg swings across and hides it.

That is the optics problem. Production computer vision learns from human-labeled frames, and its accuracy degrades when the data is thin or the object is hard to see.⁵ A ball mid-flight is the hardest frame there is.

So FIFA stopped asking the cameras to do it alone.

The 2026 World Cup is played with the adidas Trionda. Inside one of its four panels sits a 500Hz motion sensor — moved out of the centre, where Qatar 2022 suspended it, and built into the panel itself.⁶ It reads the ball 500 times a second.

Built with Kinexon, that sensor pinpoints the exact moment of a touch to within about two milliseconds⁷ — a precision no camera frame rate can reach.

Eyes outside. A heartbeat within. The cameras watch the pitch. The chip feels the kick. Together they catch what neither could alone — the millisecond the ball was struck.

Is offside fully automatic at the 2026 World Cup? No — the machine measures, people decide

Here is the honest part. Almost nothing at this World Cup is fully automatic.

One thing is: goal-line technology. About seven cameras watch each goal. Millimetre accuracy. The referee's watch buzzes in about a second. It has been the World Cup standard since 2014 — not new, not the ball chip, not offside.⁸⁹

It answers one question. Did the whole ball cross the line? Yes or no. Nothing else.

Offside gets a second automated output, and it's new for 2026. An advanced version of Semi-Automated Offside Technology sends clear offsides straight to the assistant referees on the pitch — not just the VAR booth — so the flag goes up faster.⁴

But read the word again. Semi-automated. The system measures position. The on-pitch official still makes the call.

Everything else stays human. A foul. A penalty. A handball. A ball that sits just over the line. The machine measures. People decide.⁴

And that is the feature, not the bug. The line is drawn to the centimetre because humans defined what correct looks like — frame by frame, example after example. The narrow band the machine can call, it can call because people taught it. The wide band it can't, it leaves to people.

The human layer underneath

Here is the part nobody puts on the broadcast. Before any model can watch a match, a person has to teach it to see. Not the AI. A human, frame by frame.

Look at the open research and the scale of that teaching shows up fast. The SoccerNet game-state dataset — academic, not FIFA's, not ours — carries 9.37 million pitch-line points placed by hand to teach a model where the field is.¹ The same dataset hand-labels 2.36 million player positions, each tagged with role, team, and number.¹

That is one open dataset. Underneath one sport. A model does not learn a touchline because it is clever — it learns because someone drew that line a million times, across angles, lighting and jerseys, until the pattern stuck.

It doesn't hallucinate because someone wrote a bad function. It hallucinates because nobody gave it enough good examples.

The biggest stage in sport learns the way every intelligence learns. From human hands.

I'm Ahmed Rafiq Fahmy. My co-founder and I ran data-annotation ops for a decade — both ex-Affectiva and ex-Smart Eye — before we built the product. What's live today: the Annota8 Annotation Engine (180 annotation UIs across 7 modalities) plus an AI assistant that answers questions about your annotation data. That human-teaching layer is what we build.

Frequently asked