Mislabeled Football Data: How an Artificial Intelligence Story Slipped Through the Scouting Net

At 3:12 in the morning, my news monitoring board pushed up a line flagged in red. The metadata field read clearly: Domain Label — Football. I opened the file and read the thirteen information points inside. Not one club. Not one player. Not one match, not one contract, not one surgery. The entire content was about an artificial intelligence researcher leaving Anthropic and OpenAI, about the probability of extinction driven by superintelligence, about a race between the largest technology laboratories on the planet. Fourteen years in this job, I am used to reading skewed payrolls, unsigned contracts, and injury files with altered dates. This time the skew sat at the very top layer of the information production chain: the label. Modern football runs on data. A club in the V.League or the Thai League maintains at least three information sources: a match-data provider, a scouting department, and a media monitoring system. These three do not speak to each other directly. They speak through labels — through metadata. Someone tags a record as transfer, injury, discipline, finance, and from that moment the whole downstream chain trusts the tag. A wrong label does not produce a small error. It produces an error that propagates. I once sat comparing scouting data from two centres in Southeast Asia. One used an automated tagging system, the other used manual readers. Within one month, the automated side pushed seventy-two records into player files that had nothing to do with football: corporate finance news, federation election news, and one health insurance press release. The manual side missed twenty-three genuine stories. The cost of those two error types is completely different — but only one of them causes a player file to be mispriced. That is the error type I was holding in my hand. THIRTEEN OUT OF THIRTEEN Thirteen information points in the file, thirteen times the same subject: Jacob Coxon, a former OpenAI and Anthropic staffer; Evan Hubinger, an Anthropic researcher; the superintelligence alignment problem; an extinction probability figure above ten percent. I tried applying the nine-dimension football analytical framework to this document, exactly the procedure I use for every club dossier. Tactical and technical dimension: no system to measure. No xG, no PPDA, no possession share, no formation. Nothing to compare, not even against itself. Finance and transfer market dimension: no transfer fee, no wage, no contract length, no sell-on clause, no agent commission. Not one number that could sit beside a club balance sheet. Results and public-opinion cycle dimension: no season, no title race, no relegation place. No pressure on any manager. League landscape dimension: no league. Only two technology companies. Rules and compliance dimension: no financial fair play, no transfer window, no disciplinary sanction. The call for a pause on capability increases is a technology policy proposal, not a competition regulation. Management and dressing-room dimension: no dressing room. Only a resignation. Risk dimension: the risk mentioned is existential risk, sitting outside every football risk framework I have ever built. Media narrative and expectation dimension: no fans, no tickets, no shirts. Industry transmission dimension: no academies, no agents, no broadcasters. Nine out of nine dimensions returned an empty result. Label-to-content match rate: zero out of thirteen. Noise rate within a single record: one hundred percent. When a player dossier returns empty on nine out of nine indicators, I do not conclude the player is poor. I conclude I am reading the wrong dossier. That principle applies to data as much as to people. THE FINGERPRINT OF A FAKE RECORD A contract signed in invisible ink: the fingerprint of a deal that was never published. I wrote that line for transfers with no paperwork. It holds here too. This record is a deal that never existed, stamped with a label typed by a human hand or misassigned by a model. I traced back three layers. The first layer is the Domain Label field. The second is the topic classifier. The third is the final human reviewer. None of the three stopped this error, because all three share one blind spot: they check whether the label is valid in format, not whether the label matches the entities inside the text. In other words, the system checks the spelling of the label, not the content of the label. This is an error I have seen in medical files. A surgery recorded with the right date, the right diagnosis code, the right surgeon, but the wrong leg. Every data field was valid. Only reality was wrong. THE VALID PART OF THE COUNTERARGUMENT There is an argument I must record, because it is not unreasonable. Someone could say: artificial intelligence is entering football. Injury prediction models, player tracking systems, valuation algorithms for young players — all of them built by these very laboratories or by the people who left them. So a story about them could be an early signal for the football industry, years before it becomes a headline. That argument is partly right. But it is right at a different layer. It is right for the sports technology beat, where the question is which tool will change how scouting works. It is not right for the match-data desk, where the question is who plays left wing this weekend. Blending those two layers into one label is wrong. Not because the other story is worthless, but because it belongs to a different queue. Money never dies; it only changes address and waits for whoever is awake enough. Data is the same. A mislabeled record does not disappear. It sits in the database, waiting for some trend report to pick it up, fold it into a chart, and turn it into a conclusion no one can trace back to a source. WHO PAYS In football, the cost of dirty data does not fall on the people who create it. It falls on the people who read it. A scout who trusts a trend report will go and watch the wrong player. A journalist who trusts an aggregated table will write the wrong thing about a transfer. A supporter who reads two different articles built from the same dataset will lose faith in both. One skewed figure in a payroll is the first crack in the whole system. I heard that crack once, in 2026, when three names absent from the official match registration list were still collecting a regular monthly wage. Back then I learned that evidence must be cross-verified through two independent sources. Today I learned one more thing: the first source and the second source can both be wrong, if both of them read from a single label assigned by a single system. Injuries have files, surgeries have invoices, and the truth has exactly one keeper. With data, the truth has no keeper. It only has a labeller. Key insight: the fault does not lie with the artificial intelligence article. The fault lies in the labelling layer — where a text containing no football is still granted passage into a football system. A FORWARD-LOOKING THOUGHT I am not proposing we discard automation. Fourteen years of reading paper files left me humble enough to know that human eyes also miss things, and miss them more quietly. I propose a single check, cheaper than every other step: before any record carrying a football label enters any workflow, the system must reconcile the label against the entities. If the text names no club, no player, no competition and no football governing body, the football label must be revoked automatically. Whoever signs a contract answers for the contract. Whoever assigns a label should answer for the label.