The Financial Times recently reported that an AI system designed to profit from football betting had lost its shirt. The headline was amusing. The underlying story is not—it reveals a category error about what can, and cannot, be predicted.
The failure was not a software bug. It was not a data shortage. It was not even a modeling mistake in any correctable sense. It was something more fundamental: a collision between a genuinely complex system and tools that, however sophisticated, are structurally misaligned with it.
Understanding why AI cannot reliably predict football results illuminates something deeper—both about the limits of artificial intelligence and about why billions of people watch the game.
The Structure of the Game Amplifies Contingency
Prediction works best when systems are decomposable—when they can be broken into parts, analyzed independently, and reassembled into a faithful representation of the whole. Baseball, whatever its other merits, approximates such a system. A pitch, a swing, a fielded ball: discrete events, nearly independent, accumulating toward outcomes that stabilize statistically over time. The law of large numbers does its quiet work. Models converge.
Football is something else.
It is a continuous, nonlinear, many-body system: twenty-two interacting agents moving through space in real time, generating dependencies at every moment. There is no natural reset, no clean decomposition into independent trials. The sequence of play matters more than any aggregate. A substitution alters positioning; positioning alters pressure; pressure alters decisions; and a single altered decision may produce the decisive moment.
Cause and effect are entangled. Football is, in the language of complexity science, a coupled dynamical system—and such systems resist the reductive logic that underlies most predictive modeling.
The Low-Scoring Problem
Two structural features intensify this resistance.
The first is scoring scarcity. Basketball produces over a hundred scoring events per game; randomness averages out, and the better team typically prevails. Baseball generates dozens of at-bats; statistics stabilize. Football produces two or three goals on a good day, sometimes none at all.
With so few decisive events, variance does not average out—it dominates.
This is not a superficial observation; it has a precise mathematical consequence. Even a perfect model—one that captured every training session, tactical pattern, and physiological condition—would still produce predictions with irreducible error. A single deflection, a moment of individual brilliance, or a goalkeeper’s mistake can determine the outcome.
The uncertainty is not a failure of the model. It is a property of the system.
The second feature compounds this: sensitivity to singular events. Football operates near criticality—a regime at the boundary between order and disorder, where small perturbations can produce disproportionately large effects. A slight deflection, a mistimed step, an unexpected burst of creativity—these are not statistical noise to be averaged away. They are structurally irreducible.
In such systems, what appears as an “upset” is not an anomaly. It is an expected feature.
There is an additional, often overlooked feature. Football’s rules institutionalize imprecision. American football ends when the clock reaches zero. So does basketball. Football does not. It ends when the referee decides it ends. Time is not a continuous variable but a discretized approximation—ninety minutes plus an indeterminate quantity of “added time,” estimated rather than calculated. Even the temporal boundary of the game resists exact specification. The system does not merely produce uncertainty; it encodes it.
Metrics Without Causality
The failure of AI prediction in football is not for lack of data. Modern football generates extraordinary volumes of information: tracking data, pass networks, pressing intensity, expected goals.
We have more data than ever, but not more causality.
Most available metrics are descriptive rather than generative. They record what happened; they do not capture the mechanisms that produced it. Possession is not threat. Shot volume is not shot quality. Even expected goals approximates likelihood without capturing the unfolding sequence that determines whether a chance becomes a goal.
In a coupled dynamical system, sequence carries information that aggregates discard.
There is a deeper issue. AI systems excel when patterns are stable, when training data is representative of the future, and when problems can be decomposed into tractable parts. Football violates all three conditions. Tactics evolve within a match. Context shifts continuously. And the decisive moments—the creative acts that shape outcomes—are precisely those least represented in historical data.
You can model tendencies. You cannot foresee singular inventions.
No Single Optimum
In many high-performance domains, optimization leads to convergence. The NBA trends toward height and wingspan. Elite tennis players exhibit increasingly similar builds.
Football resists this.
Lionel Messi, at 5’7″, uses a low center of gravity to change direction at speed. Cristiano Ronaldo, taller and more powerful, dominates in the air. Erling Haaland, built like an industrial striker, organizes his game around positional efficiency. Garrincha, with asymmetrical legs, became one of the most unplayable dribblers in history.
These are not variations on a single optimum. They are distinct local solutions in a rugged landscape.
In this sense, football resembles an evolutionary system: variation persists because the environment—defined by opponents—is constantly changing. Strategies co-evolve. There is no fixed peak, only shifting terrain.
Why This Makes the Game Irresistible
Up to this point, the argument has been analytical. But it leads to an aesthetic conclusion.
A game whose outcomes were tightly determined by underlying quality would be predictable—and, ultimately, less compelling. What sustains football’s global appeal is precisely the irreducibility of its uncertainty. Every match preserves the genuine possibility that the outcome will not follow from the inputs.
Underdogs do not win by exception. They win by structure.
There is something deeper here. Football resembles life more closely than most sports. Outcomes hinge on small moments. Structure and randomness coexist. There is no single optimal path. Success depends not only on accumulation but on timing—on when something happens, not just what happens.
And occasionally, the system is bent.
Diego Maradona’s second goal against England in 1986 was not a probability. It was an event. It was something the system, as constituted, could not fully anticipate.
This is why such moments endure. They are not just achievements; they are singularities.
The Broader Lesson
Jorge Luis Borges, in The Lottery in Babylon, imagined a society in which chance is not an occasional disturbance but a governing principle—“an intensification of chance, a periodic infusion of chaos into the cosmos.” What begins as a game becomes a structure that permeates reality itself.
Football is not a lottery. Skill, preparation, and structure matter profoundly. But it shares something essential with Borges’s vision: chance is not an external defect to be eliminated. It is an internal feature that shapes outcomes.
The AI betting failure is, in the end, a small story about a larger point. We live in a moment when artificial intelligence is associated with an expanding promise of foresight—markets, epidemics, political outcomes, human behavior. Football offers a counterexample of unusual clarity.
In systems governed by sparse outcomes, nonlinear interactions, sensitivity to singular events—and even rules that embed imprecision at their boundaries—the aspiration to full prediction is not merely unmet. It is misplaced.
Football does not fail prediction; it exposes its limits.
The game defeats prediction. That is not a problem. It is the point.
