An Earthquake Alarm Ringing Inside the Football Data Room

I was sitting in my small apartment in Lyon, in the 7th arrondissement, when the clock on my phone clicked to 6:12 in the morning. The habit of an investigative sports journalist is not to open the scores first, but to open the system log. Every morning I have three scripts running automatically: one filters transfer feeds from Europe, one tracks movements in Asian betting odds, and one records every news headline my aggregation system has scanned in the past 24 hours. That third file is the one I read most carefully. It is like a gatekeeper's notebook: even the irrelevant lines stay there, waiting to be cross-checked. That morning, among the familiar names — a Premier League full-back negotiating a renewal, a Serie A striker suspended for two matches — one line made me stop. It was tagged football in the classification column. The content had nothing to do with football: “Will the seismic alert sound change this September 19?” The original headline was in Spanish, the source was Mexico, the publication date fell within the period of the second national drill of 2026. Of the 11 information points the system extracted, there was not a single club, not a single player, not a single coach, not a single competition. I sat still for about two minutes, coffee cup still in hand. This is the kind of error modern newsrooms describe with a very technical name: a data-pipeline classification error. But it is not just a technical error. It is a signal. Three harmless data points, stitched together, drew a map of money flowing into a village with no football pitch. A headline about an earthquake, a misapplied classification label, and a stream of data flowing into a football feed — three fragments that look harmless in isolation, but placed side by side, they sketch a portrait of an entire ecosystem operating without anyone checking it. The incident sounds small. A mislabeled story — what is there to say amid millions of sports articles every day? But in ten years of watching this industry, I have learned one thing: the biggest errors never begin with a big scandal. They begin with tiny specks of dust inside the machine, specks no one bothers to wipe away, until the whole apparatus makes a strange sound. And that strange sound, to me, was the seismic alert ringing in the wrong place. Misclassification of content is not new. What caught my attention was its systematic nature. If an algorithm labels an earthquake article as football, the question is no longer “why did this one go wrong,” but “how many other articles have gone wrong in the same way, and how many of them have we read without ever knowing?” --- CONTEXT --- To understand how a headline about an earthquake can slip into a football data room, one must understand how large the sports data industry has swelled over the past decade. Ten years ago, when I was a first-year journalism student in Lyon, an investigative sports reporter worked mainly with a phone, an address book, and a notebook. They called, met people, took notes, then cross-checked two or three sources before writing. Today, an investigative sports reporter works with at least four automated data sources running in parallel, each pulling thousands of lines a day. One source delivers situational data down to the individual pass. One delivers transfer data updated by the minute. One scans betting odds across dozens of bookmakers. And one aggregates news, sweeping the world's online press to detect relevant keywords. That last source — the news aggregator — is where things begin to spiral out of control. Aggregation is not an act of reading comprehension; it is an act of pattern matching. The system does not “understand” what the article is about. It counts words, matches them against a keyword list, and decides on a label. If an article contains enough words like “scenario,” “protocol,” “response,” “drill,” and “exercise,” a thinly trained classifier can jump straight to the conclusion: this is a piece about sports tactics. Look again at the structure of the original article about Mexico's earthquake. It has a named subject — President Claudia Sheinbaum. It has authorities. It has a scheduled event. It has five regional scenarios for the drill. It has response protocols. It has the numbers 80 million cell phones and 23 thousand loudspeakers. To a machine that only counts keywords, this is a perfect feast: a subject, scenarios, protocols, and large-scale figures. Every signal resembles a sports tactical analysis. Only one tiny detail differs: the subject is not a manager, the scenarios are not tactical formations, and the loudspeakers are not stadium stands. This is the nature of the problem. The modern sports data industry runs on the principle of speed, and speed always beats accuracy. In a world where a correct transfer story can be worth hundreds of thousands of euros to a newsroom, and a wrong one can cost an entire outlet its credibility, the pressure always tilts toward publishing first and verifying later. Data pipelines are designed not to miss anything. They are not designed to understand. And when you design a system not to miss anything but not to understand, that system becomes ever better at catching more things, including things it should never have caught. --- WHAT THE DATA PACKAGE ACTUALLY SAYS --- Before blaming the algorithm, I want to do something few newsrooms bother to do: read every information point in the original article again, place them side by side, and see what real picture emerges. The first information point concerns the existence of the second national drill of 2026. The second specifies the exact date and time. The third mentions whether the familiar alert sound might be kept or changed. The fourth and fifth record the President's role in the morning press conference. The sixth addresses consulting experts before a decision. The seventh describes five regional scenarios for the drill. The eighth and ninth are two numbers: 80 million phones and 23 thousand loudspeakers. The tenth and eleventh concern response procedures and reaction protocols. If you are a pattern-recognition system with no semantics, what do you see? You see an event with a date and time, a powerful subject, “five scenarios,” “protocols,” and two large numbers. In the vocabulary of a sports machine, all of these are strong signals: a calendared event, a key figure, a tactical system, and scale. The machine did exactly what it was taught. The problem is that it was never taught to distinguish an earthquake drill from a football match. This is not the first time I have encountered this type of error. In 2026, when Ligue 1 was cancelled mid-season because of the pandemic, I spent three weeks cross-checking every line of Olympique Lyonnais' emergency financial report against the financial regulator's files. I found a brokerage fee transferred to a Luxembourg company established exactly two months earlier, whose director shared a name with the agent of a substitute player. The data-analysis skills I honed in 2026 — when I rewatched all 48 World Cup group-stage matches and manually logged 1,247 refereeing decisions — helped me recognize the typical shell-company structure behind that harmless-looking number. The lesson from those episodes is simple: data never lies, but the labels stuck onto data can. A correct number placed in the wrong context leads to a wrong conclusion, and a wrong conclusion multiplied across hundreds of articles becomes a false truth the whole industry believes. That is why I never let a data line pass without asking where it came from, who labeled it, and who benefits if that label is wrong. --- WHERE THE MACHINE BREAKS --- To find the culprit, you must trace the data flow upstream. First is the collection layer. News-scanning tools work by sending requests to news sites, downloading headlines and excerpts, and storing them in a raw database. This layer is almost never wrong. It does not judge; it only copies. The problem begins at the second layer. The second layer is extraction. Here the system tries to identify entities: who, what, where, when. An article about Mexico's earthquake will be extracted into entities such as “Claudia Sheinbaum,” “Mexico City,” “national drill,” “September 19.” All correct. The problem is that the extraction layer has no job of deciding which field the article belongs to. It only picks out the pieces. And nothing in that entity list says on its own, “this is not football.” The third layer is classification. This is where the murder happens. The classifier is asked to assign a domain label to the article. If the classifier was trained on a dataset in which sports articles often contain words like “scenario,” “tactics,” “protocol,” “response,” “drill,” “lineup,” then it will learn that the appearance of these words is a signal to assign a sports label. And an article about an earthquake drill is full of them. Recall what you have seen in tactical press conferences before a big match. A manager talks about the “scenario” for handling the opponent. An assistant talks about the pressing “protocol.” The press writes about a team's “response” after conceding. The vocabulary of tactical football and the vocabulary of disaster response overlap alarmingly, differing only in context. And context is precisely what a simple classifier cannot grasp. This is the crux. In esports, players' win rates are public, but investors' win rates are not. The same logic applies here: we know very well which algorithm mislabeled the article, but we know almost nothing about who designed, trained, and sold that classifier. Responsibility is pushed toward the machine, while profit flows toward the people. --- THE GROWING DUST --- One mislabeled article is one speck of dust. But this speck does not stand alone. In my review, I found a troubling pattern. Articles about natural disasters, national drills, and government press conferences are frequently mislabeled as sports content, especially when they contain large-scale numbers and scenario-flavored keywords. The reason is fairly clear: sports is one of the most heavily labeled domains, because the demand for sports content is endless year-round. From transfers to previews, from stats to rumors, there is a colossal demand that biases the system toward labeling anything that looks like an “event” as sports. When everything can be sold as sports content, the boundary between sports and the rest of the world begins to blur. A politician speaking about a national drill can be turned by the system into a “power figure applying pressure.” A disaster-response protocol can be turned into a “tactical system.” A number about loudspeakers can be turned into “stadium capacity.” All of these are meaningless transformations, but they are produced systematically, at industrial speed. The danger lies here: once these specks slip into the pipeline, they do not disappear. They spread to downstream layers. A mislabeled article can become input for a summarization model, and that model will generate a complete sports report about an event that never existed. It can become raw material for a betting feed if that feed is designed to update with the news stream. It can become a “market signal” that drives an investor's decision. Dirty data does not stay put. It spreads. --- WHAT I KEPT FROM 2026 --- In 2026, when I was still a first-year student, I spent an entire month rewatching all 48 group-stage matches of the World Cup in Russia. After the France-Australia match on June 16, 2026, I noticed something odd: in matches with large odds gaps, most penalties went to the underdog. I manually logged 1,247 refereeing decisions, cross-checked them against open data from statistics providers and five Asian bookmakers. The result showed that one referee had 78% of his foul-flag decisions tilt toward the weaker team, 2.5 standard deviations from the mean. My article was published on a faculty blog and taken down after 48 hours. But I kept the entire spreadsheet. And that spreadsheet taught me the most important lesson in this profession: never conclude first and then go looking for evidence. You start from raw data, you build a process, you record every step, you accept that the result may contradict your initial hypothesis. Once you allow yourself to move in the opposite direction, you are no longer an investigator. You are a copywriter for your own bias. That lesson applies wholly to today's story. When I see an earthquake article sitting in a football feed, my first reaction might be outrage. But the right reaction is curiosity. Why this article? Why now? Which system did this, and how long has it been doing it? If I merely get angry, I will not find answers. If I ask questions, I will find the path of the error, and from the path of the error, I will find the path of responsibility. From World Cup 2026 to now, I have drawn another lesson: referees can read spreadsheets too. In football, people can claim that errors are random. But when you place 1,247 decisions on a spreadsheet, randomness disappears. What remains is a pattern, and a pattern always has an owner. This is true of referees, true of transfers, and equally true of the machines that label data. A single error can be an accident. A pattern of errors is a design decision. And a design decision always has a person responsible. --- FOLLOW THE MONEY, NOT THE HEADLINES --- People often think investigative sports journalism means watching football and writing impressions. It does not. Most of the real work happens in spreadsheets, in contracts, in the small footnotes at the bottom of financial reports. The transfer market never lies if you are willing to read the broker-fee column instead of the player-price column. The same logic applies to the sports data industry. You can ask: who pays for this data pipeline? Who sells it? Who buys it? Who is responsible when it goes wrong? The answers to the first three are usually easy to find. The answer to the fourth is usually very hard, because no one wants to take responsibility for a machine they merely “operate.” But if you dig deep enough, you find a network. Sports data providers are often private companies operating across borders, signing contracts with newsrooms, bookmakers, and social platforms. The supply chain is very long, and each link blames the one before it. The bookmaker says the data came from a partner. The partner says the data came from a provider. The provider says the error came from the classifier. The classifier says the error came from the training data. And the training data is checked by no one. This is why I say the problem is not merely one mislabeled article. The problem is an entire ecosystem built on the principle of shifting responsibility. When no one is responsible, anything can happen and nothing gets fixed. The seismic alarm will ring in the football feed today, in the betting feed tomorrow, and in an AI-generated content model the day after. Nothing self-corrects. A system only fixes itself when someone forces it to. --- THE CONTRARIAN ANGLE --- Now let me say something my colleagues often do not want to hear, but which I believe needs saying. Automation is not the enemy. The data pipeline itself is not the villain. Without them, a reporter like me could not track thousands of matches, thousands of contracts, thousands of movements across the world. The truth is that without automation, I could not have spotted the 2026 referee pattern, could not have found the Luxembourg brokerage fee in 2026, could not have cross-checked 1,247 decisions in a month. Tools harm no one. How people use tools harms. In other words, the right response is not to smash the machine. The right response is to force the machine to be transparent. Force it to show why it applied a label. Force it to allow re-checking. Force its operators to be accountable when it fails. This is why I never join debates about whether AI will replace journalists. That question is wrong. The right question is: when AI mislabels a million articles, who signs the bulletin that goes out? If no one signs it, it is not journalism. It is a data stream, and a data stream has no conscience. In sports, we are used to the feeling that the score is always correct. But a score is only correct when the referee blows the whistle correctly. And the referee, as I learned in 2026, does not always blow the whistle correctly. The same holds for the data-labeling machine. It is only correct when someone checks it. I know some will say: a mislabeled article — what harm does it do? It affects no one. But I have seen the harm with my own eyes. I have seen a wrong article lead to a transfer rumor, and that rumor move the share price of a listed club. I have seen a wrong report cause a sponsor to pull a contract. I have seen a tiny data error multiplied into a “truth” a whole community believed for weeks. Specks of dust never stay put. They only wait for the wind. --- WHEN DIRTY DATA MEETS AI --- What makes today's story more topical than the usual error is the context. Over the past two years, large language models have become an indispensable part of sports content production. They summarize, they translate, they generate headlines, they produce short bulletins. And they do so at a speed no human can match. Picture the chain of events. A wrong article about a “power figure” and a “national scenario” slips into the pipeline. A language model reads it, summarizes it, and generates a very plausible sports bulletin about “a national tactical drill.” That bulletin is then used as a source by another outlet, which generates the next bulletin. After three loops, the original truth has vanished entirely. What remains is a confident, fluent block of text, full of subjects and predicates, with not a shred of factual basis. This is the worst kind of distortion, and also the hardest to detect. Human distortion usually leaves traces — a bias, a relationship, a motive. Machine distortion leaves no traces, because it has no motive. It is simply a probabilistic outcome. And precisely because it leaves no traces, it spreads faster, deeper, and is harder to undo than any false truth created by humans. In financial investigation, we have a term for this kind of distortion: laundering, but laundering done with semantics. You push a false piece of information through many layers of intermediaries so that it loses its origin. When it surfaces at the final layer, no one knows where it came from, and therefore no one is responsible for it. This is why I believe football data pipelines should not be judged solely by the volume of articles they produce, but by their traceability. A pipeline that cannot be traced is a pipeline that cannot be trusted, no matter how fast it runs. --- THE PRICE OF SPEED --- When I was young, I used to think speed was everything. In a competitive world, whoever publishes first wins. But I have changed, and I remember the exact moment I changed. It was the autumn of 2026, when I received a scanned sponsorship contract. I could have published the story immediately. I could have had a big headline ahead of every colleague. But I did not. I spent ten more days verifying. I cross-checked every signing party, every subsidiary, every endpoint of the money flow. I found that the recipient company had been established just two months before signing, registered an address but had no actual office. I confirmed it with three independent sources. When the article appeared, the conclusion was rock solid. And the story made an impact, triggering an internal audit and forcing the relevant organization to admit that most of the costs had no verifying documents. I tell that story not to praise myself. I tell it to prove a point: in this profession, patiently cross-verifying three sources is not slowness. It is precaution. People choose speed because they believe opportunity waits for no one. But real opportunity, in investigative work, only comes to those with evidence. A story published early but wrong costs you far more than it earns. For the sports data industry, this lesson matters even more. Because a wrong data pipeline is not an isolated mistake. It is a mistake multiplied automatically, exponentially, every second. If each journalist verifies slowly by half a day and correctly, then a pipeline that never verifies produces thousands of errors a day. Speed, in this case, is not an advantage. It is an unpaid debt. --- LESSONS FROM COVID-19 --- The pandemic taught the world a lesson most of us had forgotten: systems that look solid can collapse within weeks. When Ligue 1 was cancelled, clubs were forced to publish emergency financial reports. It was in that very situation that the biggest holes in financial reporting were exposed. COVID-19 closed every stadium in the world, but the holes in financial reports never socially distanced. The story of the mislabeling machine is the same. It only surfaces as a problem when there is pressure. When there is no pressure, the machine runs smoothly, and no one checks. When there is pressure — a national drill, a major event, a moment when everyone needs fast news — the specks accumulated over time begin to fly. The problem is not created during a crisis. The problem is created during normal times and only surfaces during a crisis. This is why I always smile when someone tells me the sports industry is transparent. Sports is one of the most opaque fields in the world, because it wraps itself in an image of purity — of youth, of aspiration, of victory. But behind that image is an industry operating like any other industry, with the same blind spots, the same holes, and the same people who only want to pass responsibility on to someone else. Sports culture is at its most beautiful when viewed from the stands; at its most revolting when viewed from the accounting office. --- THE DUST OF TOMORROW --- If today it is an earthquake article labeled football, what will it be tomorrow? I can predict with fairly high accuracy. It will be an article about the climate crisis labeled “transfer news” because it contains words like “season” and “volatility.” It will be an article about elections labeled “transfers” because it contains “negotiation” and “coalition.” It will be a public-health article labeled “lineup” because it contains “vaccination strategy” and “substitution.” Each individual error will be ignored, because it is too small to draw attention. But combined, they will form a colossal block of noise in the data pipeline. And when that block of noise is absorbed by language models, it becomes part of how the machine “understands” football. This is the most frightening thing: not just wrong data, but wrong data becoming the foundation for further wrong data. In auditing, people call this “error propagation.” A small error at the input layer, if undetected, is multiplied through each processing layer, until the final output is completely wrong. The secret of every good system is not to never make mistakes. The secret is to detect errors at the layer where they are born, before they multiply. For the sports data industry, this means there must be a control layer at the point of entry. An article, before it is labeled football, must pass a basic test: does it contain at least one recognized football entity — a club, a player, a competition, a governing body? If the answer is no, then no matter how many “scenario” and “protocol” keywords it contains, it must not carry the football label. This is a rule so simple that one wonders why it has not been widely adopted. But the answer, as always, lies in motive. Adopting that rule would reduce the number of articles in the pipeline. Fewer articles means less revenue. And in an industry where volume is measured by speed, no one wants to be the first to slow down. --- THE READER'S VIEW --- Let me speak plainly to the readers who are spending their time on these lines. You owe the sports data industry nothing. You have the right to demand truth, and you have the right to be wary of numbers without sources. You do not need to know how the algorithm works. You only need to know one thing: if someone cannot show you the origin of a piece of information, that information cannot be trusted. In ten years of watching this industry, I have realized that most football fans are natural investigators. They remember exactly which player scored at which minute, which match ended with which score, who moved from where to where. They are like archivists of collective memory. Precisely because of this, they are the first to notice when something is wrong. When a bulletin says something absurd, fans sense it before they can prove it. That instinct is their most precious asset. The problem is that modern data systems are trying to blunt that instinct. They deliver too many numbers, too much information, too many streams, until readers no longer have enough time to check. This is the strategy of any system that wants to hide something: pump in so much information that no one can focus on the important detail. Overwhelm is not an accidental mistake. It is a tool. --- THE INSIDER'S VIEW --- And here is the hardest part, the part I think I can speak to as an insider. I know what it feels like for a young reporter sitting in front of a vast data pipeline. I have been there. You are given two hours to write a piece. You have a feed containing everything. You know that if you verify every detail, you will miss the deadline. And you know the editor will not ask “did you verify,” but “is it done.” In that situation, the temptation is to go with the flow, trust the feed, and write. I do not judge those who do. This industry is built to encourage them to. But I want to say one thing to the young people in that position: every time you trust a source-less number, you are gambling your own credibility on something you do not control. Credibility is the only thing you truly own in this profession. It is your working tool. It is your asset. It is all that stays with you after every article has grown old. It took me ten years to build that credibility. I spent hundreds of hours cross-checking documents no one bothered to read. I turned down hot-publish opportunities just to wait for the evidence to be strong enough. I have been wrong, and I have corrected myself in public. There is no shortcut in this profession. No machine can build your credibility for you. A machine can make you faster, or it can make you wrong faster. The choice is yours. --- THE QUESTION POSED TO THE INDUSTRY --- So what should we do with this story? Not shut down data pipelines. Not abandon automation. Not return to the era of notebooks and phone calls. Those tools are obsolete, and trying to return to them would only make us slower without making us more transparent. What we should do is set a new standard: every data label must be traceable. When a pipeline labels an article football, it must record the reason — which keyword triggered it, which entity was recognized, which rule was applied. If it cannot be traced, the label has no value. This is the principle I have applied to myself since 2026: if I cannot show someone how I reached a conclusion, that conclusion does not deserve to be believed. What we should also do is build a culture of re-checking. Not checking out of suspicion of colleagues, but out of respect for readers. A pipeline checked periodically catches errors before they spread. A pipeline never checked accumulates errors until someone finds them, and usually the finder is not the operator. This is why I say every investigation into sports data is also an investigation into the industry's power structure. And the last thing we should do, and the most important, is to remember that behind every line of data is a human being. A mislabeled article is not just a technical error. It can be news about the safety of millions of lives turned into an amusement. It can be a disaster warning downgraded to a short snippet. When we forget that data represents people, we have lost the reason this profession exists. --- CLOSING --- That morning in Lyon, after reading the earthquake line in my football feed, I did one simple thing: I saved the entire log. I did not delete it. I did not publish it. I saved it, because I knew that one day, when someone asks me “how did you know this system was wrong,” I would have evidence to answer. That is how I have worked since 2026. I keep the spreadsheet. I keep the log. I keep the letters, the contracts, the scans no one remembers I have. Not because I want to denounce anyone. But because I believe truth does not need to be created. Truth only needs to be recorded, and with enough records, it reveals itself. The seismic alarm in the football feed today may be just a small error. But to me, it is a reminder. A reminder that the industry I watch has grown too large for anyone to control with the naked eye. And because it is too large, we need larger rules, higher standards, and people willing to say “I am not sure enough” instead of “I will publish it now.” In football, a match is not over until the referee blows the whistle. In the data industry, a data belief is not over until a human checks it one last time. And that whistle must always be blown by a human.