TacticAI: what DeepMind's corner-kick assistant means for football

In 2024, Google DeepMind and Liverpool FC published TacticAI in Nature Communications — an AI assistant for one of football's most rehearsed, least solved moments: the corner kick. It's the clearest public example yet of AI moving from describing football to advising on it. Here's what it does and where it fits.

The problem: corners are hard to optimise

A corner is a set piece with a dozen moving bodies, layered runs and blocks, and a tiny window. Coaches drill them endlessly, but the design space is enormous and the feedback is sparse — a Premier League match produces only around ten corners, so learning "what works" from results alone is painfully slow.

What TacticAI actually does

TacticAI treats a corner as a graph: each player is a node, and the relationships between them — distances, marking assignments, runs — are the connections. That structure lets it answer three questions:

1. Prediction — who is most likely to receive the ball, and will there be a shot?
2. Retrieval — which past corners most resemble this one? (Useful for "how have teams beaten this setup before?")
3. Generation — how could we reposition our players to improve the outcome?

The technique is geometric deep learning with graph neural networks. A neat detail: the model exploits the fact that a corner is roughly mirror-symmetric — a setup on the left touchline is the same problem flipped — which effectively multiplies its limited training data.

Did it work?

The results are what made it notable:

• On finding the most relevant historical corner, it hit 63% top-1 accuracy against a 33% baseline — roughly double.
• In a blind study, Liverpool's own football experts preferred TacticAI's suggestions to the real setups about 90% of the time, and rated its generated corners as largely indistinguishable from real ones.

That's not a lab toy being polite to itself; that's domain experts at an elite club choosing the model's ideas.

Why it matters — and what it isn't

TacticAI is a preview of AI as a tactics sparring partner: a system that surfaces options a busy staff might miss, instantly, grounded in thousands of past examples. It doesn't take the decision. DeepMind and Liverpool were explicit that coach and player judgement stays central. The AI widens the menu; humans still order.

It's also worth being precise about scope, because "DeepMind cracked football" headlines blur it. TacticAI optimises one set piece. It doesn't predict league tables, price markets, or tell you who wins on Saturday.

Where outcome models fit alongside it

TacticAI answers "what should we try?". A different branch of football AI answers "how likely is each result?" — and the two shouldn't be judged by the same yardstick.

That outcome question is what Momus Modal works on. Instead of optimising a corner, a quant model reads a whole fixture into a probability for every scoreline — the correct-score grid — and from it the 1X2, over/under and both-teams-to-score, plus where those numbers diverge from the market. It's analysis, not tactics and not tips: what's likely, and why.

The takeaway

TacticAI showed that AI can genuinely help design football, not just measure it — and that experts will listen when the suggestions are good. It's narrow, it's honest about its limits, and it's a strong signal of where the game is heading: humans making better decisions, faster, with a model in the room.

Curious what the outcome side of that looks like in practice? See how a match reads on the desk.