Labeling errors in football data pipelines: when an algorithm misreads a news item

On September 24, a news item from the state of Puebla, Mexico, passed through the automated classifier of a sports data feed. It was labeled "football." Inside there was no club. No player. No scoreline. No match report, no minute of stoppage time. The content concerned a kidnapping and a death. The algorithm labeled it anyway, and that label drifted downstream without encountering a single checkpoint.

This is not a trivial detail. Across seven years of cross-checking different broadcast feeds before typing a single line about a yellow card, I learned that the most serious error in this trade is not misreading a passage of play, but letting a news item that has nothing to do with the pitch enter the system unchallenged. Numbers do not lie, but the people who record them can — and this time, the recorder was a machine.

Modern football data pipelines do not run by hand. Thousands of items a day pass through automated taggers before reaching an editorial desk. The system extracts keywords, checks them against a fixed taxonomy, and decides where an article belongs: results, transfer market, discipline, or supporter culture.

That approach works most of the time. The problem lies in the small remainder. When an item contains words matching the name of a university that once fielded a team, or a town that once appeared in a lower division, the tagger can slip. In the Puebla case, two fragments overlapped: a public university, and a place name. Both have historic ties to Mexican football. Neither carries a football meaning in this item. No club is named. No player, coach, competition, contract, or league table appears.

The case also exposed a larger problem: many classification pipelines have no slot for "society" or "crime." When there is nowhere to put something, it falls to the nearest place — and with content containing a few overlapping tokens, the nearest place is usually sport. I call it a token-level false positive, a fault that is not as rare as we assume, and not easy to spot, because it raises no error flag. The system does not know it is wrong. It only knows it matched.

To grasp the scale: if a feed ingests a few thousand items a day, and only a small fraction are mislabeled, then across a single season that figure is enough to skew any aggregate built on top of it. The crowd may be absent, but the referee still has to strain his eyes — and a machine has no eyes to strain.

The consequence does not stop at an article being filed in the wrong place. A mislabeled item enters the very data store that future training models learn from. If items that are not football are mixed in, the model will learn sentence patterns that do not belong to football. A small error, but one with an expiry date — and this time the expiry date is the day the model gets audited.

Across seven years as a discipline reporter, I ran one check before trusting any number: does it belong to the match, is the match real, and do at least two independent sources confirm it. The second step — is the match real — sounds redundant, until you realize a machine never asks itself that question.

Back in June 2026, when La Liga returned after the pandemic, I built a spreadsheet of 214 matches to measure the effect of missing crowds on card counts. A colleague asked why I did not simply pull the ready-made data online. The answer was simple: ready-made data does not come with a guarantee that it belongs to the thing I am measuring. I had to verify every row, every match, every card. The result showed yellow cards down 12 percent on the previous season — but only because I had stripped out items that did not belong to the sample. Had I left them in, that 12 percent could have been a different number, and no one could have checked it back.

An automated pipeline has no concept of "does not belong here." It only has "match" or "no match." To it, an item containing the word "university" and a place name is enough to pass. It has no mechanism to stop and ask: does this article have a club, a player, anyone who works in football? Only a human can pose that question. And when no human poses it, the system defaults to yes.

People watch players run; I watch when they stop. Data is the same: what matters is not how fast it flows, but whether it stops in the right place to be checked. The chaos of a feed is the surface; beneath it lies the law of numbers — and the numbers here say the mislabeling rate in automated systems is far from small.

One further detail is worth noting: the Puebla item had relatively clear sourcing — state authorities, local government, and a direct statement from the family. It was not rumor. It was accurate as an event, and wrong only as a classification. That is the hardest kind of error to catch: a factually correct item filed under the wrong heading. And that wrong heading, over time, shapes how a system understands football.

I do not believe in luck; I believe in slow-motion replay. In football, referees have VAR to review. In a data pipeline, we need an equivalent: a mandatory pause before any item is admitted to the store. That pause must be designed by people, not decided by the algorithm. But that pause does not exist, and does not exist systematically.

There is an easy argument: labeling errors do not matter, because a human at the end of the line will catch them. I do not believe it.

Humans only catch errors when they read. But most data in a feed is never read — it is only counted, aggregated, fed into tables. A mislabeled article can sit in a store for months untouched, until a model draws a false conclusion from it. By then, no one can trace it back, because the label still sits there as though it were fact.

Ironically, it is emotion that spreads the error fastest. Stories that shock, outrage, or intrigue are the most shared — and those are also the ones most likely to slip past a tagger, because they carry the most prominent keywords. Let "popularity" steer labeling, and you end with a feed gathering precisely the items least related to football, but the loudest.

Here lies a thin line between emotion and rules. In football, a referee cannot book a player because the crowd howls; he books because of the law. That law exists, is written down, is checkable. In a data pipeline, no such law exists for labeling. No clear definition of what makes an item football. No recorded precedent. No appeal mechanism for mislabeled items. A system without law cannot be held to account — and a system that cannot be held to account will be wrong forever.

I once recorded a player's card count incorrectly and spent two weeks reviewing 47 foul incidents in a single match. I recount that not for sympathy, but to say that every data error has a price, and that price is only paid when someone is accountable. In football, there are victories no one notices — and there are errors no one sees, until it is far too late.

The question I carried away from the Puebla item is not how to teach the machine to read better. That is the easy question. The harder one is: who is accountable when a machine mislabels, and how do we catch the error before it hardens into a conclusion? A match lasts 90 minutes, but discipline lasts a whole season — and with data, that discipline must last longer still. If this industry truly values the number, the first task is to ensure the number belongs to the right match.