The Mislabeled 'Football' Tag and a School in Iztapalapa: The Flaw Sits in the Data-Classification Layer

On my dashboard in Barcelona, one data entry carried the label "football." I opened it. No club. No player. No scoreline. No xG. Only a school called Colegio de Bachilleres Plantel 6, in Iztapalapa, in the east of Mexico City. Masked individuals threw firecrackers onto the grounds. A teacher was injured at the top of the head. A woman suffered an anxiety crisis. An evacuation. And an investigation still open. In the summer of 2026, I saw the Opta ghost – and since then, my eyes no longer trust what they see. I tell this story for a reason larger than mocking an algorithm. It touches the very core of my trade: classification. In more than five decades as a data journalist, I learned that the most dangerous error is not a wrong number, but a correct number placed in the wrong drawer. A correct fact, misfiled, will reach the right person at exactly the moment they should not have received it. And when that drawer is called "football," readers have every right to believe they are reading about football. On Friday, October 9 – the original gives no year, a detail I will return to – a group of masked individuals appeared around the area of Colonia Unidad Habitacional Vicente Guerrero, on Avenida Antonio Díaz Soto y Gama. According to the first versions, they threw firecrackers into the schoolyard. The school was evacuated. Classes were suspended. Parents and neighbors gathered outside, waiting for information. Forces from the Secretaría de Seguridad Ciudadana, the SSC, were deployed and kept a presence. Mexico City's C5 camera system was reviewed. A teacher suffered a head injury of about one centimeter, reportedly without needing transfer to hospital. Another woman experienced an anxiety crisis. By the time the original piece was written, no one had been identified, no motive had emerged, and it was still unclear how the individuals had entered the grounds. Then I read the sourcing section. Thirteen of twenty information points carried no source at all. The rest cited only vague phrases: "first reports," "first versions," "circulated information," "emergency personnel." That is the sourcing profile of the first hours after an incident – weak, and fragile. As a data person, I care less about the incident itself. I care about the flow that made it drift. A modern news process runs through three layers: collection, labeling, distribution. The labeling layer today is often automated. A model reads text, detects a topic, applies a tag. For a piece containing words like "school," "field," "crowd," the model may jump to the sports drawer. But "field" in "schoolyard" does not share a semantic field with "field" in "stadium." That is the classic problem of natural-language processing: homonyms with different meanings, and a model with no eyes to see context. There is a subtler trap I want to flag. In Mexico, the term for organized violent groups in school settings is sometimes loosely conflated with football supporter groups. That overlap in naming is enough for a model – and for a hurried editor – to jump to the sports drawer. But the original piece makes no such connection whatsoever. Any such link, if it exists, is speculation. And speculation, in my trade, is an unpaid debt. I counted a third time. Twenty information points. Not one mentioned a club, a player, a coach, a league, a match, a transfer, or a tactic. Not a single line. If this were a football article, it would be the emptiest football article I have ever read – and I have read thousands in my life. The second notable point lies in how the original writer self-corrected. At first, there was a framing that hinted at "explosive devices" and a brawl. Later, the author narrowed it: the available information supports only firecrackers, with no other explosives confirmed. A writer willing to withdraw a framing is more credible, not less. But that self-correction also confirms something: at the moment of publication, at least two competing versions were circulating among local sources. When two versions coexist, the one published first is usually only the fastest, not the most accurate. And here is where I want to pause a little longer. When I teach young reporters, I always repeat one line: correlation is not causation. Here, the variant of that line is: a label is not a fact. An entry tagged "football" does not thereby become football. But in the publishing machine, a label holds a quiet power. It decides who sees the content, in which drawer, alongside what. Mislabel it, and you do not merely file the news in the wrong place; you create a kind of false truth, confirmed by the very system that applied the label. Readers do not re-check the label. They trust it. The irony is that my trade, the data trade, grew on a simple belief: numbers do not lie. But the thing that lies, as I have written before, is the person reading the number. And now, the "person reading the number" is sometimes a model that does not read, only matches patterns. It does not understand where "school" differs from "stadium"; it only knows the two words once appeared near each other across millions of texts. That is a way of learning that is very good at frequency, and very poor at meaning. I once believed in feeling. After Opta, I believed in probability. After COVID, I believed in structure. What is the structure here? A process with no gatekeeper at the labeling layer. A system that believes classification is the machine's job, while verification is the human's – but the human has been pushed to the final layer, where the product is already on the shelf. In the summer of 2026, when stadiums stood empty, I had a rare privilege: real-time data access for a second-tier team in Catalonia. I learned that when the crowd disappears, some metrics rise, some fall, and no metric explains itself. Data never speaks on its own. It speaks only when a person places it in the right context. A wrong label is exactly that context, swapped out. Drawing on my experience following matches, I extracted one principle before trusting any statistical table: you must know how it was sampled, by whom, and for what purpose. A number cut off from how it came to be is an ownerless number. A data entry cut off from its proper content domain is no different. I am 68 years old, but data is younger than I have ever seen it – each season it grows another layer of teeth. I do not leave behind the question of who applied the wrong label. I leave behind another question: in an industry swallowing data faster than it can digest, where do we place the checker? If the answer is "at the end of the line," then incidents like the one in Iztapalapa will recur – quietly, in drawers few people inspect, until a careful reader discovers that what they are reading does not belong where it was filed. And perhaps, in a season when transfer noise drowns out real signal, keeping the right label on the right content is the most modest yet most necessary form of data discipline.