When 'Granada' Is Not Granada CF: A Classification Error and the Limits of Machine Reading

In a small apartment in Seoul, I keep a professional habit that has become almost a ritual: I open any piece of news and underline every proper noun before reading a single line of content. Not out of curiosity about the people involved, but because proper nouns mislead readers more than anything else — including the systems that read on our behalf. This week, an internal analysis file landed on my desk labelled 'football.' I read all thirty-six data points. There was not a single club. Not a single player. Not a match, a table, a transfer, a minute of stoppage time. There was a young person who had died, and a network of universities scrambling to handle the fallout.

That analysis, in turn, flagged its own error. It stated plainly: the source article has nothing to do with football, and the 'football' label is a mistake. The notable thing is not that the system failed — systems always fail. The notable thing is the mechanism behind the failure, because it can repeat in any newsroom that lets machines read the news for it.

The real story of the source article is not about football. A 21-year-old Spanish exchange student from the Canary Islands, enrolled at the University of Granada, was killed in the state of Morelos, Mexico. A man identified by authorities as the suspect is under investigation, and criminal responsibility has not been established by any court. After the death, the University of Granada and the University of La Laguna suspended their cooperation agreements with the Autonomous University of the State of Morelos. A series of other institutions — Barcelona, Zaragoza, Extremadura, Jaén, Pablo de Olavide in Seville — said they were 'reviewing' similar agreements. The Conference of Rectors of Spanish Universities proposed folding the Foreign Ministry's country security reports into the criteria for approving student exchanges.

That is a story about risk governance in education, about the duty of care of institutions, and about how a personal tragedy becomes a policy trigger. None of it belongs to football. So why did a classification system label it football?

The answer is almost certainly a named-entity collision — where place names and university names happen to match the names of professional football clubs. The University of Granada collides with Granada CF. The University of Barcelona with FC Barcelona. The University of Zaragoza with Real Zaragoza. Pablo de Olavide University in Seville with Sevilla FC. The University of Extremadura with CF Extremadura. The University of Jaén with Real Jaén. The University of La Laguna in the Canaries with the footballing region of Las Palmas and Tenerife. Even Morelos strikes a Mexican club that once existed. In one document, at least eight names could trigger a keyword-list classifier.

This is the crux: the system does not understand content, it only matches name patterns. For a classifier running on a gazetteer of places and entities, the word 'Granada' is a signal. It cannot tell whether Granada is a city with a university or a club playing in La Liga. Same name, different context, and the machine picks the wrong context. There is nothing mysterious about the error. It is a precision failure, not an ambiguity.

I once misread a contract live on air, and I paid for it with thirty days of silence, reviewing every tape to understand where I went wrong. The lesson was not 'never be wrong' but 'never speak with certainty before three cross-sources.' The same lesson applies to machines. A system that tags a femicide case as football does so not out of malice. It does so because no one forced it to check whether a club, a player, or a match actually exists in the text.

What struck me most was the response of the second-stage analysis. Instead of forcing a football frame onto a criminal case, it marked 'N/A — insufficient information' across all nine analytical dimensions and spent most of its length describing the pipeline error. That is correct behaviour. In this profession, the hardest skill is not analysis but refusing to analyse when there is no data. Outsiders are usually certain. Insiders are usually silent. A good system must know when to stay silent.

Setting the technical error aside, I see something more troubling for the sports media industry. We are building content machines whose speed outpaces our own ability to verify them. When speed beats discipline, error stops being an exception — it becomes a product on an assembly line. A football headline slapped onto a murder case, if pushed to market, causes double harm: it insults the truth of the victim, and it poisons readers' trust in everything else on the site.

I have seen newsrooms burn money racing for speed on transfer deadline day and pay for it with corrections that never erase the first impression. Money comes and goes. The name of the dead — no one can correct that twice.

Second, this mislabel exposes something most people in the trade have not yet admitted. Football and esports share one virus: rumours with no clauses. An automated news-reading system is also the system that will push transfer rumours faster than a denial can respond. When the machine cannot tell Granada the city from Granada the club, it also cannot tell a real negotiation from one staged to inflate a price. I have said this many times at the analysis table: to me, a player's price is not the number on the screen, but the sum of the rejections. But a machine cannot read rejections. It only reads the name that gets mentioned.

One detail in the analysis is worth most to anyone in media: the real case moved through three clear stages. Stage one, institutions directly tied to the victim acted first — Granada and La Laguna suspended cooperation. Stage two, the national coordinating body stepped in — the rectors' conference proposed a risk-assessment framework. Stage three, peer institutions 'reviewed' their agreements to share collective responsibility before the framework became official. That is a very study-worthy model of how responsibility propagates, and it has nothing to do with football. But it operates exactly like a transfer crisis spreads through a market: the directly exposed party reacts first, the regulator follows, the rest stand back and wait for a signal.

I realised that if the system had placed the article in the right domain, it could have become a case study on risk assessment in student exchange programmes — an education-policy topic, not a sports one. The notable point is that Morelos had already been flagged for high crime rates. The warning signal existed. The problem was not a lack of information, but the weight parties assigned to it. I once wrote that a pandemic does not kill the transfer market, it merely exposes who is playing with real money. Here too: tragedy does not create a gap, it exposes a gap that was already there.

Back to the professional question. I am not writing this to tell the story of a system that failed. I am writing it to ask those running content pipelines a question: if you cannot point to a club, a player, or a match in a document, whose interest does your 'football' label serve? Does it serve the algorithm, or the reader? Every contract has three numbers — the announced, the real, and the one they want you to believe. Every content label does too. And readers, sooner or later, find the real number.

At 56, I do not believe the word 'football' written over a murder case. At the negotiation table, I only believe clauses. At the news desk, I only believe sources. Three independent sources, clear timestamps, clear probability levels — and if there is none, the most honest answer is silence.

The lesson is not in fixing a label. The lesson is in building a gate: before any system is allowed to call a document football, it must prove that the document contains a team, a player, or a real match. If not, the only correct output is a cross — and a person sitting down, re-reading from the start, exactly as I did for thirty days after my live mistake.

I leave a thought more worth pondering than any of this: we are building systems that read news faster than humans, but we have not yet taught them what they are not allowed to read. The truth is, the ability to refuse is the final measure of a decent system. And my industry, football or esports, still owes its readers something very old: silence at the right moment.

(Analysis based on public records and re-reading of the text; the case is under investigation and no criminal responsibility is implied.)