The next weekend of top-flight European soccer — Premier League, La Liga, Serie A, Bundesliga — will print several hundred match lines at sportsbooks before Friday's first kickoff. Somewhere in that same window, another round of "we asked Grok and GPT and Claude to pick the matches" posts will land on X and Reddit. This is a ritual. It is also, by every serious read of the bet-settlement numbers anyone has published, one that ends with the models losing money.
This piece answers the questions the next round of those posts will not: why LLMs keep failing this test, why xAI's Grok is catching the loudest part of the blame, and who the structure of the experiment actually benefits.
Where did the "AI models are terrible at soccer betting" story come from?
From a growing pile of reproducible experiments, most of them run by developers and journalists rather than labs, in which an LLM is handed a set of upcoming matches, asked to produce a pick or a probability, and then scored against the actual results and the bookmaker's closing odds. The earliest examples were hobbyist notebooks. The recent examples have been built out into weekly columns, newsrooms, and betting-forum threads with real profit-and-loss accounting attached.
The reason the story has a recognisable shape now, rather than just being "developers tinkering," is that enough of these experiments have run across enough leagues for a consistent pattern to emerge. The models do not beat the books. Most of them lose money against closing lines. Some lose faster than others. Grok is the one that keeps getting named in the headlines, but the underlying loss curve is the field's.
Why is xAI's Grok specifically singled out?
Because xAI has been the loudest, publicly, about Grok being well-positioned to reason about sports, prediction markets, and real-time information. When a model is marketed as the one that handles live, time-sensitive reasoning, the soccer-betting test is the one the internet will run on it. It is cheap, it has a clean scoreboard, and the result is binary: did the pick cover or not.
The quieter part of the story is that Grok is not failing in a unique way. It is failing in exactly the way every other LLM fails at this task — overconfidence, poor calibration, a tendency to narrate a match before pricing it — but because xAI has positioned Grok as the informed-bet model, the test gets applied to Grok first, hardest, and loudest. The branding drew the crosshair. The failure mode belongs to the whole category.
Are LLMs structurally bad at probabilistic predictions?
Yes, and the reason is not "they don't watch the matches." The reason is that LLMs are trained to produce the most plausible next token, not the best-calibrated probability. When you ask a model "what is the chance Real Madrid beat Girona at the Bernabéu Saturday," it is not running a simulation over a forward model of football. It is retrieving the linguistic shape of how an analyst would answer that question — and the linguistic shape of confident analysis skews toward round numbers and narrative coherence.
That is not a knowledge problem that better training data fixes. It is an optimization target problem. The model is aiming at the wrong thing. Sportsbook pricing lives or dies on fractional-point calibration. LLM output lives or dies on whether the sentence sounds right. The two objective functions do not compose.
Is this a knowledge problem or a calibration problem?
Calibration, decisively. You can stuff every injury report, every expected-goals model output, every referee tendency, and every weather forecast into the prompt, and the model will still produce badly-calibrated probabilities, because the failure is not in what it knows. It is in how it translates what it knows into a number.
The diagnostic is simple. Take any LLM output that says "Arsenal has a 72% chance of winning this match." Run the same prompt a thousand times with varied phrasing. You will get a distribution of probabilities wide enough to drive a truck through. That variance is the calibration problem on full display. A real probability model does not swing fifteen points on rephrasing. The model is not telling you what it thinks the probability is. It is telling you what a confident sentence about this match looks like, and the numerical tail of that sentence is essentially vibes.
Who makes money when AI gets positioned as a betting oracle?
Not the bettor. That is the whole answer, and it is the part of this story that gets soft-pedalled in most write-ups.
The people who benefit from framing LLMs as betting tools are, in order: sportsbooks, who welcome more retail flow at the closing line and do not care which heuristic brought the user there; affiliate-driven content sites, which can now publish "AI's top picks" as a new flavour of search bait; and the model vendors themselves, who get free press every time a journalist runs the experiment, regardless of outcome. The story "AI picked three matches and they all hit" goes viral. The story "AI picked three matches and they all missed" also goes viral. Either version is oxygen.
The reader running the experiment on themselves with real money is the one subsidising the ecosystem that produces the framing.
Does any model actually beat the sportsbook on soccer?
Not reliably, and the specialized, non-LLM models that come closest are not the ones anyone asks Grok to compete with. Serious football prediction work is done by teams using Poisson-style scoreline models, expected-goals-based simulators, and market-informed blends of the two. These models do not write analysis paragraphs. They produce probabilities, size bets via Kelly fractions, and eat thin margins over thousands of matches.
The gap between those systems and an LLM is not a gap of intelligence. It is a gap of objective function, tooling, and discipline. The specialized model knows it is a bookmaker's adversary. The LLM thinks it is a commentator. Asking an LLM to beat a sportsbook is like asking a sports columnist to run a trading desk: the output shape is wrong before the content ever gets evaluated on the merits.
What would it take for a language model to be useful in this workflow?
A much narrower job description than "pick the winner." The useful version of an LLM in a serious betting workflow is a research assistant — summarising injury news, parsing tactical previews, flagging line movement that does not match public commentary, pulling structured numbers out of unstructured match reports. That is a task LLMs are actually good at, and it is the task a disciplined bettor would pay for.
The useless version is the one where the model is asked to produce the probability directly. That is the version the viral posts test, and it is the version that keeps failing. If the workflow puts a calibrated numerical model downstream of the LLM — LLM extracts features, numerical model prices the match — the combination can be useful. If the LLM is the whole pipeline, the pipeline is losing money by design. The viral experiments are almost always the second version.
What should a builder take away from this?
That "can an LLM do X" is the wrong question when X is a task whose output is a calibrated number. The right question is "where does the LLM fit in a pipeline whose output is a calibrated number." That is a more boring question, and it does not produce viral headlines, but it is the one that separates useful tooling from content.
The sports-betting test is valuable precisely because it has a hard, unambiguous scoreboard. Most LLM use cases do not. When a vendor claims their model is good at a task where calibration matters — medical triage, legal risk scoring, financial forecasting — mentally run the soccer experiment on it. Would it pass? If the answer is "probably not, for the same reason Grok does not," then the claim is marketing, not capability. That is the real use of the soccer story.
What is the one number that should change how you read the next "AI picks the matches" headline?
Zero. As in: the number of publicly-documented, independently-verified experiments in which a general-purpose LLM has produced positive returns against closing sportsbook lines on soccer over a statistically meaningful sample. We have not seen one. If you have, the experiment is either too small to matter, cherry-picked by date range, or scored against opening lines rather than closing lines — which is the difference between a bet that survives contact with the market and a bet that does not.
That zero is what should decide whether you pay for an "AI betting signals" subscription, whether you treat an LLM's match pick as anything more than narrative, and whether the next viral thread of green-tick emoji is worth ten minutes of your attention. It is not. The math is closed.