A File Tagged Football, 48 Data Points, and Not a Single Player

On my desk in Liverpool sits a file tagged football. I read it top to bottom, twice. No clubs. No players. No match, no goals, not a single corner kick. Forty-eight information points, and all forty-eight concern Pakistan's internal politics: negotiations between the government and the opposition, a long march planned for 27 September, the legal situation of Imran Khan and Bushra Bibi, an interior minister seeking to block a protest, and commentary on inflation, electricity prices and public spending. The label says football. The content does not. I stopped there longer than I stopped at any transfer rumour all week. Not because the Pakistan story troubled me. Because the label did. In forty-nine years in this trade I have learned something simple: a bad report is caught by readers within ten minutes. A bad tagging system can run silently for months and nobody checks. People call me reckless, but the numbers have never lied to me. And the numbers here are painfully clear. At sixty-five I have watched sports news travel from the typewriter in a local radio studio in 2026 to content pipelines that run entirely on their own. I once sat in the studios of Voice of Israel and Channel 1, commentating on football, the Olympics, athletics championships. Back then a story needed at least three humans before it aired: the writer, the editor, the final reader. In 2026 that process is compressed into a few milliseconds of automatic classification. Thousands of reports a day are stripped into data points, tagged by subject area, and pushed into different analytical pipelines: tactics, club finance, the transfer market, law and governance. Fast, cheap, and on paper, efficient. The problem is that the classifier does not read. It counts keywords. And keywords betray people far more often than we admit. Based on my experience watching and commentating on matches, I recognise a familiar tell here. A match report written by someone who was not in the ground always carries a smell: the right numbers, the right names, but missing the one detail you only get by being there. This file has that smell, only at the data layer. The deep analysis I cross-checked states the domain label plainly: football. But three lines in, it falls apart. No FIFA. No UEFA. No AFC. Not one domestic league. No contracts, no transfers, no player wages. Only political agreements, a protest with a fixed date and time, and arguments about electricity bills rising two and threefold. What is remarkable is that the analysis still patiently ran all nine dimensions of a professional football framework. And all nine returned the same verdict. Here comes the data. I will lay it out like a lawyer, because that is the only way this story is not read as a complaint about technology. Nine dimensions. Tactics and technique: no formation, no pressing scheme, no expected-goals data. Club finance and the transfer market: not one deal, not one contract structure. Results and the opinion cycle: no table, no form. League landscape: no league. Rules and governance: no football rule system invoked. Management and dressing room: no coach, no players. Risk profile: the only risk identified sits inside the data pipeline itself. Media narrative and expectation: no team to hold expectations about. Football industry transmission: not one link. Nine out of nine. The share of football content in this file is zero per cent. And here is the detail that made me sit down and write. The only sporting entity anywhere in that file is Imran Khan — the former captain of Pakistan's cricket team, who lifted the 2026 Cricket World Cup and later became prime minister. He appears as a politician, not an athlete. Across forty-eight information points, that is the sole point of sporting contact, and it is cricket, not football. In other words: a single check — does this text contain any football entity at all — would have caught the error at the door. No large language model needed. No specialist needed. Just one question. So what fooled the classifier? The analysis points fairly bluntly at generic keywords. March. Protest. Leader. Constitution. Plus the presence of currency units and price figures. Together, those four signal groups were enough for a keyword-counting system to tag the text as sport, and even as football. I understand why. In English, march is both a procession and a month. Leader is both a political chief and the man at the top of the table. Constitution is both a founding legal text and a player's physical condition. These vocabulary collisions are a classic trap, and we — the people who write about sport — have lived with it for years without ever naming it. What worries me is not one file. It is what happens next. The analysis rates overall risk as high and states clearly that this is a data-quality risk, not a sporting or financial risk to any club. I agree, and I want to put it more strongly. Because I have been on the receiving end. I was once wrong about the 2026 World Cup. That remains the most expensive lesson I own. In the Croatia-Denmark round-of-16 tie I mispronounced the name of Ivan Rakitić three times on live air. Three times. Not because I did not understand Croatia — I had been analysing their midfield for weeks. Because I trusted my memory instead of checking the source. I spent the following month rewatching the tapes, compiling the passing numbers of Modrić, Rakitić and Brozović in the knockout rounds, and writing a piece that argued against myself. The difference between my 2026 mistake and this label is the speed of transmission. I was wrong once, in front of a few million people, and I had a month to fix it. A wrong label can flow through hundreds of articles and thousands of data tables, and nobody signs their name to it. There is nobody to apologise to. Nobody to fix it. When bad data flows into betting markets, into prediction models, into automatically rewritten news copy, the damage stops being one poor article. It is money, it is trust, and it is an entire information ecosystem eroding from the inside. Data does not kill emotion. It gives emotion a frame. But that frame is only trustworthy when people know where it was built. Now the part where I argue with myself, because a piece built only on numbers and with no room for doubt is just another form of propaganda. The first possibility, which I rate low but cannot entirely dismiss: the football tag may belong to an internal classification system where every sports item is lumped into one group, and this file carrying that tag is ordinary operations. If so, the error is not the classifier's but ours, for reading an operational label as though it were a claim about content. I accept the criticism. It does not save the output, though: whatever the label, a football analysis pipeline should never run on a text with no football in it. Another possibility is that I am overstating the severity. One bad file in a batch of thousands may be a defect rate of one in a thousand. Statistically, that is small. But here is the point I want the data operators to hear clearly: the analysis sets the acceptable threshold for this class of error at zero. Any occurrence greater than zero inside a batch is enough to corrupt downstream output. For data used to make judgements, one in a thousand is not a tolerable margin. It is a hole. And the possibility that unsettles me most: perhaps I have become the conservative in my own industry. I grew up in newsrooms with three layers of human checking, and at sixty-five I still believe a story must pass through a human hand. Perhaps that is nostalgia, not expertise. Perhaps the people building these systems are right: accept a small error rate in exchange for speed, and let the platform correct itself. I do not think so. But I also do not want to pretend I am a neutral party in this story. My proposal is very simple, naive enough to be dismissed: add one mandatory gate that asks exactly one question — does this text contain any football entity, be it a club, a player, a competition, a governing body, a match? If the answer is no, the file is routed back to its proper domain. That gate is cheaper than any model and catches this precise class of error. My judgement, specific enough to be tested within twelve months: at least one major sports outlet will be found to have published or broadcast material built from a mislabelled data pipeline. And whichever organisation installs an entity gate first will not be on that list. The rest is up to readers. Football waits for no one. It only waits for those willing to ask the question.