Dirty Football Data: When the Machine Labels Guadalajara as Chivas

I was at my screen at one in the morning in Manchester when an item tagged "football" came through my feed. I read it three times. No team. No player. No coach. No competition. No transfer, no broadcast revenue, no league table, no passage of play. I went back through all sixteen information points the system had extracted. Points relevant to football: none. The content was a sensitive family matter in Guadalajara, in the state of Jalisco, Mexico. I will not retell that story. I will not name anyone involved. The person at the centre of it never chose to become a public figure, and a football feed has no right to make them one. What I want to talk about is the label that dragged it in here. In 2026 I wrote that Manchester City paying 50 million pounds for Kyle Walker from Tottenham was an act of madness. I argued that a full-back pushing that high would leave space behind him, that City's defence would break in the Manchester derby. I bet my editor they would lose at least three home games before the turn of the year. They lost one. They won the Premier League with 100 points. Walker provided six assists. People laughed at me over Walker. Three years on, they laughed until they cried over the price of defenders. What I learned was not to stay quiet. What I learned is that every label has a cost. A wrong label does not fade. It flows into databases, into aggregates, into the work of the writer who comes after me, and it stays there. Football content today runs on a conveyor belt the audience never sees. An article enters the system. A classifier assigns it a section. An extractor pulls out names of people, places and organisations. An entity-linker matches those names to existing records: clubs, players, competitions, countries. Then come sentiment scoring, interest ranking, aggregation, packaging. Finally, a headline appears in front of someone like me. Every layer is a chance to fail. And from what I have observed over years of working with aggregated datasets, the weakest layer is entity linking. Proper nouns in football are a semantic trap built almost perfectly. Barcelona is both a city and a club. Napoli is both the largest city in southern Italy and the team bound to the memory of Diego Maradona. Chelsea is both a district of London and one of the wealthiest clubs in Europe. "Athletic" is an English adjective and the name of Athletic Club of the Basque Country. "Real" means "royal" in Spanish and prefixes Real Madrid, Real Sociedad, Real Betis. "Rangers" is a club in Glasgow and a baseball team in Texas. Lyon, Monaco, Valencia, Sevilla, Bordeaux: each name is a city and a club. In the case I read that night, the disruptive token was "Guadalajara". To a Mexican, Guadalajara is first of all a city. To an entity-linker trained on football data, Guadalajara is Club Deportivo Guadalajara, Chivas, founded in 2026, one of the most storied clubs in Liga MX, playing at Estadio Akron. One mapping. That is all. A place name pulled onto a club name. The classifier then looked at the contaminated entity string, saw a football club, and concluded: this belongs in the sports section. Nobody at any layer asked the simple question: does this article talk about football at all? I have spent twenty-seven years in this trade looking at labels, and I sort them into a few familiar families of error. The most common is a place name that doubles as a club name. It is also the most dangerous, because it turns matters wholly unrelated to football into "football news" with a single word. The next kind is a common word doubling as a proper noun: United, City, Sporting, Galaxy, Real. This family pollutes feeds more quietly. Then there is the shared-name person. Football contains thousands of people who share a name with a player, a coach, a referee. There is also the social story that mentions a player in exactly one sentence, usually a throwaway one, and gets pulled into the section by the whole chain. One more: an old article republished by a sports source, so that the old section tag clings to it forever. The cost of these errors is not one misplaced article. The cost is the downstream flow. When a national feed counts articles about a club, that count is inflated by pieces that say nothing about the club. When a fan sentiment index is computed from articles, the index absorbs a criminal case. When a language model learns from that dataset, it learns both the wrong label and the sensitive content the label dragged along. Based on my experience watching matches across many seasons, I hold to one principle: a data error at the first layer travels further than an editorial slip at the last, because the last layer still has a human reading it, and the first does not. I have been wrong in my own way. At the 2026 World Cup in Russia, I wrote that Iran would keep a clean sheet against Spain, that they had nailed their penalty area shut into an impenetrable block. I called it a concrete defence. I overlooked something basic: the speed of ball circulation and the physical gap. Diego Costa tore that defence open with a near-post run and a finish on 54 minutes. Spain won 1-0. Iran did not lose because of their defence. They lost because of the fear I could see before kick-off and chose not to look at. I wrote a 1,500-word correction the moment the match ended. I did it for a professional reason: I voluntarily leave a trace of my own errors in public so readers can check them. The data pipeline I described above cannot do that. It has no corrections page. Football and esports share one law: whoever fakes it gets exposed. Football is the sport where every claim has to face the weekend scoreboard. A data system that grants itself exemption from that scoreboard is no longer journalism. It is a duplicating machine. I disagree not because I want to be different. I disagree because the majority has been wrong with me before. And this time, I have to check myself first. There are three ways I could be wrong. The comfortable one: this is a one-off. Modern classifiers are highly accurate, and I am exaggerating a single error out of a million rows. Statistically that is fair. But error rate is not the only variable. Error position is the decisive variable. An error in a single article is harmless. An error at the index layer, where everything below inherits the result, multiplies. Another possibility: the machine is not the culprit. People are. Sports journalism has cut the checking desk, and that old desk was what stopped pieces like this one. If I had to name a root cause, I lean here. And the possibility I fear most: I am talking about a technical matter when the real issue is ethical. A sensitive family case deserves a serious editorial desk, a clear legal standard and a clear privacy boundary. It does not deserve to sit in the "football" section of any site. This is the point I want to state plainly: a sensitive-content gate belongs upstream of the section classifier. An article showing signs of domestic violence should be held at the first layer, no matter how many sports keywords it carries. I do not need to know how accurate the model is. I need to know whether it knows when to stop. At 43 I still speak hot, but the fire has learned to wait. Here is a verifiable prediction: within two seasons, at least one major British football publisher will publicly release its content classification standard, a document stating what goes in which section, on what grounds, and how errors are fixed. I have a smaller, easier proposal too: add one line at the end of every automated aggregation, giving the original source, the timestamp, and the number of machine layers it passed through before reaching the reader. Fans are angry with me because I break their dreams. Dreams built from truth last longer.