TacticAI
Last reviewed
Jun 3, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,639 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
7 citations
Review status
Source-backed
Revision
v1 · 1,639 words
Add missing citations, update stale details, or suggest a clearer explanation.
TacticAI is an artificial intelligence system that gives football (soccer) coaches tactical advice on corner kicks. It was built by Google DeepMind in collaboration with Liverpool FC and described in a paper titled "TacticAI: an AI assistant for football tactics," published in the journal Nature Communications on 19 March 2024.[1][2] The system models the players on the pitch as a graph and uses geometric deep learning to predict how a corner is likely to unfold, then generates alternative player setups that a coach can weigh up. In a blind study, five football experts from Liverpool preferred TacticAI's suggested setups to the ones actually used in matches in 90% of the cases they reviewed.[1][3]
Corner kicks are an unusually good target for this kind of analysis, and the reasons are worth spelling out. Most of a football match is fluid and continuous, with players reacting to each other in ways that are hard to break into discrete, repeatable events. A corner is different. Play stops, both teams take up planned positions, and the routine that follows is rehearsed on the training ground rather than improvised. That makes corners one of the few moments in the game that resemble a set play in a sport like American football, and it makes them tractable for a model.[2][4]
They also matter on the scoreboard. Corners are among the most reliable routes to a goal in open competition, so even a small improvement in how a team attacks or defends them can change results over a season. The catch, as the DeepMind researchers noted, is that there are only around ten corners per Premier League match, which means the amount of high-quality labelled data available to train a model is limited compared with the millions of examples that power image or language systems.[2] A central design goal of TacticAI was to squeeze strong performance out of that relatively small dataset.
The project did not appear from nowhere. It grew out of a multi-year research relationship between Google DeepMind and Liverpool that began around 2021. Earlier outputs from the same collaboration included a 2021 study on using AI to analyse penalty kicks and the 2022 "Graph Imputer," a method for estimating the movements of players who drift off camera and out of the tracking data.[2][5] TacticAI brought those strands together into a single, more complete system aimed at one well-defined phase of play.
Liverpool FC supplied both the data and the human expertise that made the work possible. The club provided the tracking and event data used to train and test the models, and its staff took part in evaluating the results. The named lead authors of the paper were Zhe Wang and Petar Veličković of Google DeepMind, with Karl Tuyls and DeepMind chief executive Demis Hassabis among the larger group of contributors that included Liverpool personnel.[1][3]
The partnership reflected a wider trend of elite clubs hunting for analytical advantages. As Veličković put it in interviews around the launch, top clubs are always searching for an edge, and techniques like these were likely to become part of modern football going forward.[3] The collaboration was framed as research rather than a finished commercial product, and the published work focused on demonstrating that the approach was sound and that its recommendations struck practitioners as useful.
The technical heart of TacticAI is a choice about how to represent a corner kick. Instead of treating the pitch as an image or the play as a sequence of numbers, the system turns each corner into a graph. Every player on the field becomes a node, and the relationships between players become edges connecting those nodes. Each node carries features describing that player at the moment the corner is taken, including position, velocity, and simple physical attributes such as height and weight, along with which team they are on and whether they are the goalkeeper.[2][6]
A graph neural network then processes this structure through message passing, a procedure in which each node repeatedly updates its own representation by gathering information from its neighbours. Over several rounds, this lets the model build up an understanding of the whole setup: who is marking whom, where the space is, and which attacker is best placed to meet the ball.[6]
What sets TacticAI apart from a generic graph network is its use of geometric deep learning to exploit a symmetry of the problem. A corner routine works the same way whether it is mirrored left-to-right or top-to-bottom across the pitch, so the underlying tactics are unchanged under those reflections. TacticAI bakes this in using group-equivariant convolutions over the dihedral group of horizontal and vertical reflections, effectively letting the model learn from four reflected views of every corner at once and treating mirror-image situations consistently.[2][6] This is the trick that lets the system generalise well despite the scarcity of training examples.
TacticAI is organised around three tasks, two predictive and one generative.[1][2]
| Component | Type | What it does |
|---|---|---|
| Receiver prediction | Predictive (node classification) | Estimates which player is most likely to make first contact with the ball |
| Shot prediction | Predictive (graph classification) | Estimates whether the corner will lead to a shot on goal |
| Guided generation | Generative | Recommends adjustments to player positions and movements to raise or lower the chance of a shot |
The first two components answer questions about a corner as it stands. Given the starting setup, who gets on the end of the delivery, and does a shot result? On the benchmark data, the receiver model reached a top-three accuracy of about 0.78, meaning the actual receiver was usually among its three most likely candidates, and the shot model reached an F1 score of roughly 0.64.[6]
The third component is where the system moves from analysis to advice. Using a generative model, TacticAI proposes small, realistic changes to how players are arranged, for example nudging an attacker's run or shifting a defender's starting spot, and predicts how those changes would alter the likelihood of a shot. A coach can sample several variations of a routine and pick the one with the most favourable predicted outcome, whether the aim is to create a chance when attacking or to smother one when defending.[2][7] The model deliberately adjusts one team while holding the other fixed, which keeps the suggestions interpretable but is also a simplification of how both sides would adapt in a real match.[6]
Because there is no way to replay a Premier League season with AI-designed corners, the team validated TacticAI through a study with human experts rather than on-pitch results. Five Liverpool FC specialists took part: three data scientists, one video analyst, and one coaching assistant.[2][6]
The experts were tested in two ways. In one test, they were shown a mix of real corner setups and ones generated by TacticAI and asked to tell them apart. They largely could not, scoring close to chance, which indicated that the system's suggestions looked like plausible, realistic football rather than artefacts of a model.[6] In the second test, they compared TacticAI's recommended setups against the setups actually used. Across 50 corner situations, the raters judged the system's suggestion to be at least as good as the real one in 45 cases, or 90%, a result the authors reported as statistically significant.[1][6]
That 90% figure is the headline number most widely quoted, and it is worth reading precisely. It is a measure of expert preference in a blind review, not a measured increase in goals scored or conceded. The study showed that knowledgeable practitioners found the recommendations credible and often preferable, which is a meaningful endorsement, but it is a different claim from a proven on-field performance gain.
TacticAI was presented as a proof of concept for a particular way of using AI in elite sport: not to replace coaches, but to give them a tool for exploring tactical options faster and more systematically than they could by hand. The researchers suggested the same graph-based approach could extend beyond corners to other set pieces and, in principle, to other team sports where modelling the relationships between players is useful and tracking data exists.[3]
Commentators were also careful to mark the limits. A corner is a prearranged routine, and an assistant like this cannot make split-second decisions during open play, such as whether to take a quick corner to catch an opponent off guard.[4] There is also an arms-race dynamic: once attacking teams adopt AI-optimised routines, defending teams can use the same kind of analysis to counter them, so any advantage may erode as the tools spread.[4] Even so, the work was widely seen as a credible demonstration that geometric deep learning could produce tactical advice that professionals take seriously, and as a sign of how analytical methods are moving from the back office onto the training ground.