The 'Football' Tag on the Kate del Castillo File

The file arrived on a morning in the middle of the regular season. The classification label said, clearly: football.

I opened it. Inside were Kate del Castillo, Joaquín "El Chapo" Guzmán, Sean Penn, the Mexican government and the Sinaloa Cartel. No formation diagram. No PPDA. No xG. No transfer fee. Not a single player's name.

I read it a second time, then a third. By the third read I understood the problem was bigger than one misapplied label: that label had passed through how many layers of people, and not one of them stopped it.

Eighteen years covering this industry, from my first assignments in Madrid to newsrooms in Guangzhou, had taught me to expect strays in the sports feed. A transfer story filed under basketball. A Winter Olympics item sitting beside tennis. Those were errors at the margin. This one was an error at the core.

A criminal case file, labelled football, and almost nobody in the production chain qualified to notice.

Four layers between a story and your screen

Collection scrapes thousands of sources: foreign outlets, press releases, social media. Classification assigns a topic label. Editing is where a human reads, fixes and rewrites. Distribution pushes everything to apps, feeds and search engines.

The cost sits almost entirely in the middle two layers. Classification is cheap because machines do it. Editing is expensive because people do it. When sports output grows faster than editing budgets — and in Vietnam that has been true for years without interruption — the editing layer gets compressed.

Nobody removes the editing layer. They simply give it less time per file.

An editor with forty minutes per story spots immediately that a file about a Mexican courtroom is sitting inside football. An editor with four minutes does not. That editor reads the headline, reads three lines, checks whether any name looks familiar, and ships it.

Modern sports content classification is not optimised to describe an event. It is optimised to place an event where it earns the most reads.

"Kate del Castillo" is a string with high search volume. A labelling model only needs to see that string appear densely in its training corpus, without understanding what it belongs to, to slot it into a large category.

In Vietnam the risk runs higher, because a large share of sports content is translated and re-edited from foreign sources. Entity recognition on foreign names is weaker, and the last human in the chain often lacks the context to know that this "Guzmán" is a drug lord, not a defender.

What the file actually contains

Kate del Castillo was born on 23 October 2026 in Mexico City, daughter of the actor Eric del Castillo. She became famous across Latin America and with overseas audiences as Teresa Mendoza in the series "La Reina del Sur" from 2026, and later in the Netflix series "Ingobernable".

In the early 2010s she posted a public message expressing more faith in Joaquín Guzmán than in governments. The Sinaloa drug lord saw it. His lawyers made contact. A film project about Guzmán's life was discussed.

In October 2026, del Castillo arranged for the actor Sean Penn to meet Guzmán in Mexico. Rolling Stone published the resulting interview on 10 January 2026. Two days earlier, on 8 January 2026, the Mexican navy recaptured Guzmán in Los Mochis, Sinaloa, after his escape from Mexico's highest-security prison in July 2026.

The meeting and the article became the centre of a years-long argument: over the safety of those involved, over access to a wanted man, and over the line between interviewing and aiding. Mexican prosecutors opened an investigation into financial transactions connected to the film project. Del Castillo denied wrongdoing, said she had been treated unfairly, and has never been convicted in that strand of the case. In 2026 she helped produce a documentary series telling the story from her point of view.

That is what the file contains: power, money, fame and criminal law. An entirely serious story, and an entirely non-football one.

Worth stating plainly too: this file does not close. Investigations, legal disputes and documentaries keep regenerating it on a cycle. Each time it regenerates, it passes through a fresh classification pipeline, and each pass raises the odds of landing in the wrong section.

Three mechanisms that keep a bad label alive

Entity-based labelling. The system recognises proper nouns. In the training corpus, certain names appear densely in sports contexts. "Guzmán" is the surname of many footballers and managers. "Sinaloa" has appeared in reports about Mexican football. "Mexico" is a national team with enormous search volume. Three overlapping signals are enough for the model to assign a sports label without reading the rest of the sentence.

Label inheritance. A misfiled item at the top layer produces a bad summary. That bad summary becomes input for the layer below. After a few cycles the wrong label becomes the original label, and nobody can trace where it started.

Category-based distribution. On most platforms a story sits in one main section. Sports sections generally hold steady traffic and command higher advertising rates than entertainment. Economic pressure pushes content toward whichever section pays better, regardless of what the content is about.

Here I have to say something few people in this industry want to hear. Guangzhou taught me this: money cannot buy a match, but it can buy the man standing next to it. In football, money buys referees, agents, silence. In sports media, money buys position in the feed. Same logic, different pitch.

Based on my experience tracking matches, I always verify the source before trusting a metric. A metric with a bad source produces a bad conclusion, and that bad conclusion gets cited onward as fact. A classification label works exactly like a metric. It is a promise about the content, and that promise can be broken.

The damage does not stop at one misplaced story. It produces what I call label inflation: when the "football" tag appears on stories with no football in them, the tag's informational value falls. Readers learn the label cannot be trusted. By the time they meet a decent tactical breakdown filed in the right place, they can no longer tell it apart from everything else.

For someone who does this work, that is a direct loss. Analysis is only worth something if the reader believes it is where it says it is.

For readers the cost is more concrete. A tag is the last filter you have when you open a feed carrying hundreds of stories a day. When that filter breaks, you have two choices: read everything, or read nothing. Both are worse than reading exactly what you came for.

And as search systems increasingly demand that content deliver new informational value — something a serious analysis has to demonstrate — label inflation works against the interests of the very platforms that generate it. Cheap labels produce cheap traffic, and cheap traffic does not hold readers.

Where I might be wrong

There are three places where I suspect I have overstated the case.

I assume the file is the product of an automated system. It may have been hand-labelled by an overloaded person. If so, my analysis of machine models is only half right. I do not have system logs, so I am inferring from the file's structure.

I assume a misplaced article causes harm. There is a counter-argument worth weighing: sometimes a story outside its section reaches a new audience. If someone opens football for results and stops at a file about power and law, they have read something they were not looking for. That is a form of accidental benefit.

And this is where I hesitate most. I tend to treat every system error as evidence of a larger decay. A single misclassification may be just a single misclassification. I have been right when I picked a team nobody picked. I have also been wrong when I turned one case into a general rule. The line is thin, and I am standing on it.

What I will not retract is the core: a system can distribute a criminal case file under a football label for hours without a single checkpoint.

Three checkpoints a decent pipeline must have

Cross-entity validation. A story earns a sports label only if it contains at least two verified sports entities: team, player, competition, coach. Personal names do not count when they share a surname with a well-known figure from another field.

Verb validation. Criminal files use investigate, charge, prosecute. Match analysis uses score, assist, transfer. A file with a high density of legal verbs sitting in a football section is a file that needs a human.

Label provenance logging. Every time a label is changed, leave a trace. No trace means no accountability, and no accountability means the error repeats.

What I predict

People need data to make predictions. I only need to watch the crowd and walk the other way. Right now the crowd is heading toward "a system bug is a small thing, let it go".

I disagree. Over the next eighteen months, sports content platforms will be forced to disclose more about how they label, not for ethical reasons but for advertising reasons. When a label loses value, the ad rates attached to it fall with it. The people paying will be the first to demand an audit.

If that happens, what gets fixed sits in the editing layer, not the machine layer. Humans come back, with less power and more responsibility.

This is a hot take, and I accept being pushed back on.

A file about a drug lord landing in the football section sounds like a small detail. But "anomalous" is always the signal easiest to ignore and the one that says the most. I learned that in a newsroom where everyone laughed at me for proposing a nineteen-year-old start ahead of a forty-million-euro signing.

That time I was right. This time I do not know yet. But I will not wait until I know to write it.