Inside the data tunnel: why a Saturday Night Live notice carried a football label

In Lyon at night, only one file remained lit on the screen. I opened it, and the label line glowed green: football.

Inside was a short notice about Saturday Night Live. Season 52. Premiere on September 26. Two new faces. A host who is a basketball player. A musical guest. A list of returning cast members. A few names departing. Not a single club. Not a single footballer. No coach, no competition, no goal, no contract.

Only one word fooled an entire data pipeline. That word was "cast."

I sat with that file longer than necessary. Not because it was interesting, but because it was familiar. It was exactly like the nights I spent in a windowless rented room in Moscow, listening to the heater, wondering what I was actually seeing through the narrow gap of my pen. The Moscow room had no windows, yet every night I saw the World Cup shining through that gap. Tonight, what shone through was not a match, but a misclassification — and that error says more about my trade than any match report ever could.

People call me a writer on the edge of the pitch. I only try to capture the breath of the ball before it rolls. But my trade now is not only sitting at the pitch's edge. It is standing between two worlds — one is human emotion on the field, the other is machines that read words. And tonight, the machine read wrong.

This is not a story about a broken file. It is a story about a system learning how to see, and seeing wrong.

Context: a windowless tunnel

Screenwriting for documentaries taught me something journalism never did: everything is editing. A shot means nothing until it is cut into a sequence. A sentence works the same way. A sports outlet works the same way. And a content-classification machine works the same way — it only sees the sequence it is fed.

In recent years, the sports information industry has run on a pipeline: collect — label — analyse — publish. At the top, thousands of articles pour in daily. In the middle, a layer called "domain tagging" decides which field each article belongs to: football, basketball, tennis, or entertainment. At the bottom, analytical models take the labelled pieces and start generating conclusions: tactics, finances, discipline, dressing room.

When the pipeline runs well, it saves a newsroom hundreds of hours. When it runs wrong, it does not fail once. It fails along the whole chain.

I followed one mislabelled file for four months. It was not a football article. It was a brief notice about casting for an American television show. Season 52 of Saturday Night Live, airing on NBC, premiering on September 26. Two new faces joining the cast. A host who is a professional basketball player — Jalen Brunson, of the NBA, not football. A musical guest. Second-year performers returning. Several veteran names leaving, including Chloe Fineman and Bowen Yang.

That was the entire content. Not a single football entity.

Yet the label stayed green.

One multi-meaning English word can contaminate an entire national football dataset. That is what I want to write about here, and I want to write it not as a technical warning, but as a record of how humans and machines together misread a fact.

Back when I was a statistics editor for a local football site in Lyon, I learned that data does not speak for itself. People speak for it. In 2026, at 24, after a 2-2 draw between Lyon and Marseille, I wrote a piece about Nabil Fekir's return from injury. I was supposed to write about expected goals. Instead I wrote about his eyes when the captain's armband was handed back to him. My editor brushed it aside, told me I wrote like a dreamer, and pinned a label on me: "too literary."

I have kept that label to this day. It taught me that labels are never neutral. A label can open a door, and it can close one. For an article, a wrong label only costs an opportunity. For a data pipeline, a wrong label manufactures an entire false world.

Anatomy of a semantic false positive

I call this phenomenon by its name in language processing: a semantic false positive. It means the system encounters a word, notices that the word has appeared in football contexts before, and concludes the article belongs to football. It does not read meaning. It reads traces.

Look at the vocabulary that fooled the system in that file. I list these words not to accuse, but to dissect.

First, "cast." In television English, it is the verb for selecting and assigning performers. In sports English, it appears in the "cast" of a squad, or in the verb "to cast" a player to another club. The same string of characters, two different worlds. The machine sees no difference. The machine only sees a match.

Second, "season." In football, it is the campaign: August to May, with fixtures and a table. In television, it is the broadcast season: September to May, with a writing staff and an advertising budget. The temporal structures overlap almost perfectly — and that is the most beautiful trap of all. A television season can impersonate a football season without even trying.

Third, "return." A player returns from injury. A performer returns after a season away. The same verb, two kinds of identity. To a model that only counts words, both are signals of "squad continuity."

Fourth, "additions." A club adds signings. A show adds new members. In both cases, it is a signal of "personnel reinforcement" — a concept that transfer analysts are highly attuned to.

Fifth, "departures." A player leaves a club. A performer leaves a show. In football, this news tends to come attached to expiring contracts, wages, agents. In television, it comes attached to shooting schedules, scripts, performers' labour agreements.

Once these five words appear densely enough in a short piece, a keyword-based tagging system has no way to tell a casting room from a recruitment room.

And that is the entire story. No conspiracy. No one deliberately doing wrong. Just language — inherently polysemous — colliding with a machine that inherently prefers single meanings.

I once spent four months inside Lyon's Groupama Stadium during the pandemic, recording wind, birds and echoes during closed training sessions with only seven cameras and two cleaners. I proposed a short film called "Football in Silence." The studio rejected it, saying there was no market. But I made it anyway, with two friends. The lesson I carried out of those four months was this: absence has a voice too, and that voice is usually ignored because it does not fit a familiar template.

That Saturday Night Live file was the same. A silence mislabelled as noise.

Ghosts in the dataset

I have a private phrase for files like this: ghost football. Documents that carry a football label but contain no football inside. They wander through datasets, get read by models as real, and quietly inject into conclusions about tactics, finances and dressing rooms things that never existed.

I picture such a file passing through three layers of analysis. The first layer asks about tactics. There is no formation to dissect, no shape to compare, no metric to compute. The model should stop and say: insufficient data. But a model does not know how to say that, unless someone taught it. And if it was never taught, it will invent. It will see "cast" and describe a squad. It will see "season" and construct a campaign. It will see "departures" and write about a dressing-room farewell.

The second layer asks about finances. No transfer fee, no wages, no release clause, no contract length. It should be a blank. But once again, the blank can be filled with imagined numbers, as long as they sound plausible.

The third layer asks about rules and governance. No FIFA, no UEFA, no national federation, no financial fair play breach. But if the system has made it this far, it will find a rule to break.

What is frightening is not that the system reads wrong. What is frightening is that the system never chooses silence when it has nothing to say.

I remember one moment in Sochi, in 2026, when I was 25 and sent to Russia to cover the World Cup. The quarter-final between Croatia and Russia. Luka Modric — a small player, unremarkable in physique — ran 12.8 kilometres, dribbled past three players, and scored from outside the box. After the match, I skipped the official press conference and followed him toward the tunnel. I saw him standing alone, crying, even though his team had won.

That night I wrote about Modric's tears as a statement about pain and release. The piece was the most shared on the site. But what I remember most is not the article. What I remember most is that I skipped a press conference to follow a human being. And precisely because of that, I saw something the data tables could never see.

A windowless tunnel, a player crying alone — and a file carrying the wrong label. All three are things outside the official frame. One touch of the ball is an unfinished poem. The ball rolls on, but the writer stays behind. The writer stays behind to ask: what was missed?

A counter-intuitive angle: a wrong file is not rubbish

Here I want to turn against instinct.

Our instinct when we find a mislabelled file is to throw it away. Quarantine. Remove it from the dataset. Cut it from the model. And that is right — operationally, it must be done. A document that does not belong to the football domain should not sit in a football dataset, even for a second.

But diagnostically, that file is a gift. It is a free test of the entire pipeline.

I once worked in quality control for documentary scripts. People tend to think QC is the most boring stage. I think it is the most honest one. Because it does not judge the correct product. It judges whether the system catches the wrong one.

A Saturday Night Live file tagged as football tells us three things a correct file never will.

First: the system lacks a minimum football-entity threshold. For an article to be tagged football, it should contain at least one club, player, competition or governing body detected in the text. The file had none of those four. It slipped through not because the system is weak, but because the system had never been asked to check that.

Second: the system relies too much on surface keywords and too little on structure. A football article has a typical structure: subject, action, competition context, table consequence. A casting notice has a completely different structure: subject, role, broadcast timing, brand consequence. The two structures overlap only at the lexical surface.

Third: the system has no mechanism for self-doubt. It has no release valve that says: there is a contradiction here. A basketball player appearing in an article tagged football is already a red signal that should trigger immediate re-checking. But no. The label was applied, and once applied, it carried the weight of a destiny.

A mislabelled file is worth more than a hundred correctly labelled ones, because it shows exactly where the system is blind.

And there is a deeper layer I want to name: these semantic errors are not randomly distributed. They cluster in precisely the lexical zone where football overlaps with the performing industries. "Cast," "season," "return," "additions," "departures" — these are words of the stage and the broadcast wave. Modern football, after all, is a competitive performance industry. It has seasons, casts, entrances and exits, audiences, points of view. So the border between the two worlds is thinner than we think — and the machine stands right on that border.

I do not find that sad. I find it worth writing about.

The shadow of the source

There is another aspect of that file I cannot ignore, because it touches a professional principle of mine.

Of 25 information points extracted from the file, 23 carry no stated source. Only two are attributed, and both point to the same outlet: Variety, an entertainment trade publication. The original outlet the file came from is The Express Tribune, a regional English-language site.

So: an article synthesised from a single source, with most details untraceable. Technically, that is an evidence loan at high interest. Professionally, it is a reminder.

I am used to readers asking me: is this true? And the most honest answer is usually not yes or no, but: where did it come from, and how many layers of mediation sit between the fact and the page.

A transfer story passing through three layers — agent, journalist, aggregator — loses half its credibility before reaching the reader. An entertainment story passing through the same three layers does too. But with one difference: a sports story has a powerful self-correction mechanism — the match will be played, and the truth will show on the scoreboard. An entertainment story does not. Once the information spreads, it stays, even if wrong.

What makes an information system trustworthy is not the number of articles it produces, but the number of articles it dares to say it could not verify.

I think of the nights in Lyon, noting non-sporting details on the pitch: a smile, a hug, a glance. I did that because I believe people are made of things that cannot be measured. But tonight I realised the reverse: precisely because the unmeasurable matters, the measurable must be checked twice as hard.

The Moscow room had no windows. But if that window opened, I would want it to open onto a world where facts carry clear sources, and where blanks are left alone as blanks.

The transmission of an error

When I told this story to a colleague in data analysis, she asked me: how bad can a small labelling error really be?

I took a while to answer, because I think the reverse question is the right one: what could stop it?

A labelling error travels the same path as money travels through a league. It starts at a very small point — a word, a keyword match, a label line. Then it passes through analysis layers. At each layer it is amplified: the tactical layer turns it into a casual conclusion; the financial layer turns that conclusion into a number; the governance layer turns that number into a risk; the media layer turns that risk into a story; and finally the reader layer turns that story into a belief.

What is remarkable is this: at no layer does anyone intend to lie. Each layer simply trusts the layer above. And at the very first layer, a green label was applied.

This is why I believe the story of that Saturday Night Live file is not a technical anecdote. It is a lesson in structure. In football we are used to chain failures: a wrong VAR decision leads to a disallowed goal, a disallowed goal leads to a result, a result leads to relegation, and relegation leads to the fate of an entire city. We know small errors do not distribute evenly. They accumulate at junctions.

In sports information, the junction is language. And we are running high-speed systems through those very junctions without installing signals.

I once hosted "Football Night" for about four years, both presenting and producing. That job taught me that audiences do not need to know everything. They need to know that the person speaking to them knows what they do not know. A good host is not one who answers every question, but one who asks the right one.

For a data system, the right question is: does this document contain a football entity? And if not, who is responsible for the label?

A frozen moment

I will close with a small detail, because I believe in small details.

In that file there is a name: Jalen Brunson. A professional basketball player, of the NBA, appearing as a television host. In a document tagged football, he is the only person with any link to the sports world — but not football, and not as a player.

I stopped at that name for a long time.

It is an error. But it is also an image: a person from the sports world, standing between the worlds of sport and performance, accidentally marking the junction between them. If the machine does not ask itself why a basketball player is inside a football document, then the machine has learned nothing about the world.

I remember the night Modric wept in the tunnel. No metric measures that moment. But if a data system mislabelled that night, it would not simply lose a moment. It could generate a player who never existed, in a match that never happened, to explain an emotion that was never felt.

That is what frightens me most when I write about data: not the coldness of the machine, but its capacity to tell a beautiful story about something untrue.

And for a writer like me, someone constantly scolded as "too literary," that fear has a particular shape. I have spent a career believing emotion can carry truth. If a machine learns to use emotion without needing truth, it will do what I do — but in reverse.

That is why I wrote this. Not to complain about an error. But to remind us that every time we paste a label onto a person, a document or a night of performance, we are deciding what will be seen behind it.

The stadium is empty, but memory is never empty. In a dataset, the same should hold: a document with no football in it should be left alone as a blank — not filled with an imagined match.

I learned to listen to the pitch with my heart, because reason had said too many tired things. But tonight I learned one more: sometimes listening with the heart means daring to say, there is no pitch here.

The question I leave is not a question about an algorithm. It is a question about the person who applied the label: if you cannot verify what you are looking at, how can you retell what you have seen?