By Sensible Stats · Published
Revised
In Liverpool’s 2019 Champions League semi-final comeback against Barcelona, Trent Alexander-Arnold took a corner before the defence was ready and Divock Origi scored. Google DeepMind opens its explanation of TacticAI with that moment. It is a useful place to start, because the cleverest part of a corner can happen before anyone jumps.
A corner looks like a delivery problem: put the ball somewhere dangerous and introduce it to a tall colleague. But the more interesting question is who gets to that space, who follows them, and who is looking the wrong way when the ball arrives. TacticAI asks whether a computer can help a coach see a better arrangement. The answer is promising. The harder question is how we would know that the arrangement actually works.
References:[2] Zhe Wang and Petar Veličković
How does TacticAI analyse a corner?
Developed by Google DeepMind with Liverpool FC, TacticAI represents players and their relationships as a graph. It can predict the first receiver, estimate whether a shot follows, retrieve similar corners and suggest adjustments. Its architecture also handles reflected versions of the pitch, rather than having to learn each orientation independently. These are the capabilities described in DeepMind’s technical explainer.
Here is an illustrative coaching problem, not a reconstruction of a studied corner. An attacker starts near the penalty spot, another runs towards the near post, and a third stays deeper. Follow only the ball and the second player looks central to the move. Follow the defenders and the first player may be the important one: their movement could occupy the marker who would otherwise challenge at the near post.
Change that starting position and several things might change together. The marker may follow, hand over the runner or hold the zone. The goalkeeper may get a clearer route. The deeper attacker may become the first-contact target instead. There is no sensible way to judge the movement while pretending the other players are furniture. Furniture, to its credit, rarely changes marking systems at half-time.
References:[2] Zhe Wang and Petar Veličković
What does TacticAI’s 90% result actually mean?
The peer-reviewed Nature Communications study used 7,176 historical Premier League corners, randomly divided into 80% training and 20% test data. Its separate expert evaluation involved five Liverpool specialists: three data scientists, a video analyst and a coaching assistant. They preferred the proposed adjustments over the original setups in 90% of their comparisons. That is a result about expert preference, not a percentage of corners converted into goals.
The distinction changes the story. A coach finding an option useful is meaningful evidence for an assistant designed to help coaches. It does not answer whether players execute the move successfully, whether an opponent adapts, or whether the team scores more often. A diagram can win a meeting without winning a header.
The random historical split also leaves a later-season test unanswered. Our reading is that a stronger practical evaluation would reserve future matches, then examine actual implementation. That is a proposed next test, not something this paper reports having completed. The coaching collaboration gives the work relevance; it does not remove the need for evidence beyond the collaborators.
References:[1] Nature Communications
A useful routine must survive the opposition
Return to our illustrative near-post run. Suppose a proposed adjustment gives the attacker a little more space in the model. A coach still needs to ask what the defender is likely to do once the routine becomes familiar. Does the move depend on surprise? Does it require a delivery that the taker can reproduce? If the runner arrives early, does the opportunity disappear?
Those questions suggest a practical role for an assistant: generate alternatives, then make the football staff argue about them. One option may create room for the intended target but expose the team after a clearance. Another may be easier to repeat but less effective against a particular marking scheme. We would want the trade-offs explained before describing either as the best corner.
This is also where viewing a single freeze-frame can mislead. A starting arrangement is a plan, not its execution. To assess our hypothetical routine, we would want to follow the movement through first contact and the second ball, including the defenders left outside the box. The most photogenic part of a set piece is not necessarily the part that decides its cost.
References:[1] Nature Communications[2] Zhe Wang and Petar Veličković
What would convince us?
Our proposed evaluation would begin by recording the recommendation before the corner, along with the reason a coach chose or rejected it. Otherwise, successful routines can acquire rather flattering explanations afterwards. We would keep rejected suggestions too: a system that is useful only when its less persuasive advice quietly disappears is difficult to evaluate.
Next, we would separate first contact, shots, chance quality and goals. Each answers a different question. More shots would not automatically mean better chances; more goals over a short run could reflect finishing as well as the routine. Report the number of corners behind every comparison and the uncertainty around the result. One excellent afternoon is a story. It is not yet a stable rate.
Finally, compare performances against different opponents and delivery types, and account for who is actually on the pitch. In an ideal trial, staff could compare planned alternatives prospectively. Where that is impractical, an observational comparison would need to acknowledge selection: coaches may choose an assisted routine precisely when the matchup already looks favourable. None of these evaluation requirements is satisfied merely by producing a convincing animation.
References:[1] Nature Communications
Watch the player who does not get the ball
For the next corner you watch, try following one runner through the replay without looking at the delivery. Who moves with them? Who takes over? Where does that leave the intended target? Then follow the players outside the box. You are looking for a repeatable mechanism, not trying to infer the coach’s entire plan from one successful finish.
That is the useful promise we take from TacticAI: a way to ask more precise questions about relationships that a ball-focused replay can obscure. Its published results justify interest in coaching assistance. They do not establish a universal recipe for scoring, or an automatic improvement to match forecasts.
At Sensible Stats, we would investigate such a forecasting input only with suitable licensed data, a clear explanation and testing on untouched matches. We have not added one here. For now, the computer has earned a place in the corner meeting. Someone still has to take the corner, and the opposition has regrettably retained the right to object.
References:[1] Nature Communications[2] Zhe Wang and Petar Veličković
Sources and research status
- TacticAI: an AI assistant for football tactics
Zhe Wang, Petar Veličković and colleagues · Nature Communications · · Peer reviewed
- Zhe Wang and Petar Veličković, “TacticAI: an AI assistant for football tactics”, Google DeepMind, 19 March 2024. Developer explanation; not an independent evaluation.
Research findings are attributed to the sources above. Illustrative examples, proposed evaluations and interpretations are Sensible Stats analysis.