When a Football Data Pipeline Misreads a Rock Obituary

Three in the morning in Tokyo. I opened the report sent back by the content analysis pipeline, and the first line woke me up: "Domain: Football. Level: Deep Professional Analysis." I read through nineteen information points, looking for a single player's name. There was none. A team. None. A minute of play, an expected goals figure, a decisive pass. Nothing at all. Instead there was the story of the death of Sass Jordan, the Canadian rock singer and former Canadian Idol judge.

I sat still in front of the screen. Not because I was shocked by the death of an artist. But because I was staring straight at a machine confident enough not to know what it was talking about. And worse, that machine sat inside a pipeline that I and many colleagues still trusted to deliver judgements about football.

I have spent eighteen years covering sport. Eight Olympics, eight World Cups, the Giro d'Italia and the Tour de France. I have seen data save a piece of writing and data kill one. But I had never seen a serious football analysis system analyse an obituary.

To understand what happened, you have to understand how sports content pipelines operate this decade. A modern sports newsroom no longer writes every piece by hand. It runs a production line: source collection, topic tagging, domain classification, then dispatch into specialised analytical modules. Football has a tactical and transfer module. Track and field has a biomechanics module. Swimming has a training-cycle module. Every module assumes the input data is already on the right subject before it starts working.

The problem sits at the tagging step. Most classifiers work by counting keyword and entity signals. If an article contains enough proper nouns, place names, and abstract nouns, the algorithm assigns it to the nearest domain. For a rock obituary mentioning international tours, large stages, music awards and a television programme, the classifier can catch the wrong signals and drag it into an unrelated bin. In this specific case, it dragged it into football.

But the notable thing is not the mislabelling. Errors are everywhere. The notable thing is how the system responded after mislabelling: it did not stop.

A pipeline is engineered never to return an empty result. This is the implicit principle of almost every large-scale content production system. No module is programmed to say "I do not have enough information." They are programmed to always return an answer, because an empty answer is treated as an operational failure, and operational failures are not something anyone wants to report upward.

The result is that when a rock obituary falls into the football module, the module does not raise an error. It does exactly what it was taught: it fills every blank. Tactical sophistication: insufficient information. Club financial structure: insufficient information. Public-opinion cycle: insufficient information. Each blank is filled with a null conclusion, and the whole report becomes a building of nine floors of neatly arranged blanks, looking highly professional, highly organised, highly credible.

This is the paradox of analytical automation. The more blanks that are validly filled, the more credible the report looks. But formal credibility is inversely proportional to content value. A perfectly formatted empty report is more dangerous than a sloppily formatted empty report, because it is easier to read as real.

I have seen the same thing in my own trade, only at a smaller scale. Years ago, at a major tournament, an automated statistics system tagged a counter-attack as "successful" simply because it ended in a shot. The shot went wide. But the algorithm had counted it as a chance. A week later, a coach quoted that very number in a press conference, and nobody in the room checked the footage again. Data has a voice, and I have been shouted at by it.

The rock obituary story is no different in essence. Only in scale. Here, the thing mislabelled is not a single passage of play but an entire subject. An article about a human being's death was dragged into a pipeline designed to talk about goals, transfers and injuries.

Look at the structure of the report to see it more clearly. It has nine analytical dimensions, from tactics, finance and results to governance, dressing-room, risk, media, and industry transmission. Eight of the nine dimensions return null results. Only one dimension - media narrative and expectations - has real content, and that content analyses an obituary as a media product. In other words, the system confessed its own failure in eight parts out of nine, yet still presented the whole thing as a complete report, complete with an information-value rating table and a recommendation list.

This leads to a question I consider central to every modern sports data problem: when should a system be allowed to stay silent?

In football, we have a very good and very hated silence mechanism: VAR. When the video referee does not have clear enough evidence to overturn the on-field decision, he keeps the decision. He chooses silence. The whole stadium can scream, but the mechanism holds. That is a correct design: under conditions of insufficient data, preserving the original state is the safest choice.

Content analysis pipelines do the opposite. They have no mechanism for preserving the decision. They must always issue a fresh verdict. And so, when the input data is on the wrong subject, they do not preserve the empty state. They manufacture a new verdict out of nothing.

There is a deeper layer. The sports data industry is built on the assumption that everything can be measured, and everything measurable can be analysed. That assumption holds for football. But it is imposed on things that do not belong to football. When you hold a data hammer, everything starts to look like a nail.

An obituary is not a nail. It is a media event belonging to culture, music, and human life. Dragging it into the football module is not merely a technical error. It is the expression of something larger: the expansion of numerical thinking into every corner of life, including corners where numbers have nothing to say.

This is where I want to push back on my own first reflex.

Reading the report, the natural reaction is to blame the algorithm. But algorithms do not create themselves. They are the product of decisions made by people. Someone decided the system must always return a result. Someone decided an empty report is a failure. Someone decided speed and volume matter more than subject accuracy.

In other words, the real error is not in the machine. It is in the belief that a machine can replace editorial judgement. And that belief does not come from the engineering room. It comes from the meeting room, where performance metrics are set, where budgets are cut, where people need output to justify the system's value to leadership.

I once worked in a newsroom whose budget was cut by forty percent during a cancelled season. I know the feeling of having to produce at any cost. I also know that it is precisely in such conditions that people are most tempted to trade accuracy for output. That is when errors like this are born.

So when I saw a rock obituary fall into the football pipeline, I did not immediately think about fixing the algorithm. I thought about fixing the question we are asking. The right question is not "how do we make the system classify more accurately," but "why is the system not allowed to say it does not know."

Another blind spot: we tend to believe misclassification is a small matter, fixed by simply re-tagging. But a misclassification at the input layer propagates down every subsequent layer. A mislabelled subject gets analysed wrongly, then summarised wrongly, then published wrongly. And once published, it becomes part of public memory. Nobody goes back to fix memory.

In football, we have a term for this: the phantom goal. A goal awarded even though the ball never crossed the line. It exists in the match record, in history, in the memory of fans. Fixing it afterwards is almost impossible. Content misclassification is the same: once an obituary is processed as a football event, it has existed wrongly within the system, and everything built on it is wrong by extension.

There is one thing I have learned after many years of working with data: the greatest value of a system lies not in what it can say, but in what it dares not say. A mature analysis pipeline is one that knows how to stay silent when asked the wrong question.

Data has a voice, and I have been shouted at by it. But I have also learned that silence is an answer too. And sometimes, the most honest answer a machine can give is to admit it is standing in the wrong room. With Sass Jordan's obituary, the machine stood in the wrong room for nineteen information points without ever knowing. The reader would never have known either, if someone had not chosen to sit down and read it to the end.

I read it to the end. And I wrote down what I saw, because that is the job of anyone in this trade: not to believe the machine, but to check whether it is standing in the right room.