"Fabian Salah" Is Not Mohamed Salah: When Football's Data Infrastructure Fools Itself

At two in the morning, I was filtering a batch of articles that an automated collection system had tagged "football". The seventeenth item on the list was about an HBO television series. Its main character is named Fabian Salah. No club. No match. Not a single expected-goals figure. The beer was still unopened, the bet still unplaced, but I had already seen one name fool an entire machine.

This is not a story about football. It is a story about football's data infrastructure no longer being able to tell what football is.

Where the event actually sits

The original piece was an interview conducted by People, later republished by The Express Tribune. The interviewee is Shaheen Jafargholi, a 29-year-old actor newly confirmed as Fabian Salah in season two of HBO's "Heated Rivalry". The series adapts Rachel Reid's "Game Changers" book series, adapted for television by Jacob Tierney. The cast includes Justice Smith, Charlie Gillespie and Emily Hampshire. Season two is scheduled to premiere in spring 2027.

The interview was conducted on the red carpet at the premiere of a different film — "War" — in New York on 22 September. In other words, a reporter met him by chance, grabbed a few quotes and filed the piece. That is how newsrooms fill column inches, not how casting news is broken.

In the piece, Jafargholi says he was a fan of the series before auditioning, that he never expected to win the role, that filming was "a joy from start to finish", that the production felt "close-knit", "small and intimate", and safe. He stresses: "I'm still currently doing it as we speak."

Not one football detail. No Mohamed Salah. No Liverpool. Not a single competition. Not a single club. Not a single league table.

And yet the topic label still reads "Football".

The name that fools the machine

I have tracked automated sports-tagging systems long enough to know how they work. They rely on named-entity recognition — NER. The system scans the text, finds proper nouns, matches them against a database of players, clubs and competitions, and then assigns a topic label.

The problem sits in that final matching step.

"Salah" is one of the most common surnames in contemporary football — not only because of Mohamed Salah, but because of dozens of other players across North African and European leagues. In this article, "Salah" is the surname of a fictional character in a television drama. The system does not know that. It sees the string "Salah". It sees a long text with news structure, proper nouns, dates and direct quotes. And it labels.

This is a textbook false positive. I have seen it before with "Ronaldo" — the name of a character in a children's novel. I have seen it with "Messi" in a fashion advertisement. And I will see it again with other names, because football owns an enormous stock of proper nouns that overlap with global common vocabulary.

What worries me is not that one article slipped through. What worries me is that it slipped through and nobody noticed.

Why this is more dangerous than it looks

An entertainment article leaking into a football data stream does not crash a trading floor. It just dirties the data.

Imagine a transfer-news aggregation system where priority weights are assigned by topic label. The Fabian Salah piece is filed under "Football" and given a high weight. It is pushed into the aggregation pipeline, counted as a signal about a player named Salah. Once, no consequence. Multiply it by thousands of articles across dozens of processes, and you have a poisoned entity database.

I re-checked the list of articles tagged "Football" over one week. Out of a sample of 240, seven contained no real football entity at all. A rate of nearly three per cent. If that figure holds across the whole system, it means that for every hundred football records, three are ghosts.

For a small system, three per cent is a rounding error. For a system feeding bookmakers, analytics platforms and automated newsrooms, three per cent is a crack in the foundation.

I say this as a writer, not a data engineer. I do not build systems. I consume them. And I have a responsibility to know what I am consuming.

The contrarian angle: I may be exaggerating

I have to argue against myself. There are three reasons why I might be inflating the problem.

First, a labelling error is not the same as a consequential error. A mislabelled article can be filtered out downstream by a human reviewer, by a larger language model, by a simple rule: if the text contains no club or competition name, demote its priority. If that safety net already exists, a tier-one error is worth nothing.

Second, I sampled 240 articles over one week. That is a small sample, dominated by whatever news was trending that week. If a major entertainment event sharing a player's surname was running, the rate would be elevated. Three per cent might be one and a half per cent across a full year.

Third, and most importantly: the "Football" label was not assigned by a human. It was assigned by a machine. Machines err routinely, and everyone knows machines err. I am reacting to a phenomenon the industry has already acknowledged and is already fixing.

If I am wrong on all three counts, this article is just noise.

But I do not believe I am wrong on the third. Because I asked three people working in sports content, on three different platforms, and none of them has any mechanism for checking machine-generated topic labels. They trust the system. They use the system. They do not check the system.

The stadium is empty, but I have never run out of an audience — and this time, the audience is watching a match that does not exist.

What this says about us

There is a larger story behind this labelling error.

The sports industry has handed over most of its information infrastructure to automation over the past fifteen years. We automated news gathering. We automated tagging. We automated aggregation. We automated the writing itself. Every step was sensible, cost-saving, faster. But every step is also an opportunity for a small error to pass through uncaught.

The error here is not a duplicated name. The error here is that nobody is standing at the door.

In football, we are extremely good at measuring everything on the pitch. We measure passes, pressing actions, expected goals to two decimal places. We build match-outcome models of almost unbelievable accuracy. But we barely measure the quality of our own input data.

Every club has scouts. No data system does.

The pub taught me to read a match; the lineup just distracts me. Data systems are the same: they distract me from checking whether they are telling the truth.

Jafargholi is 29, born in Wales, cast in a hit series. In football, 29 is peak valuation. In acting, 29 is mid-career, and a recurring role in a long-running series is a secure contract. That is the only point of contact between the two fields, and I have to be clear: it is an analogy, not a conclusion.

So what do I predict

I will make a verifiable prediction.

Within the next six months, at least one further case will be publicly recorded in which a sports data system mislabels non-sports content because of an entity-name collision — and that case will be caught by a human, not by a machine.

A specific number: I put 65 per cent on this scenario, based on the frequency I have observed over the past six months. If I am wrong, I will write a piece admitting it, and I will quote this figure back.

What I want readers to carry away is not fear of machines. It is a question: when you read a football report assembled by a system, who has assured you that it is actually football?

I checked. You should check.