Could artificial intelligence have changed the Chargers’ fortune?
Based on projected outcomes, bots are already capable of saying the team should’ve gone for several fourth downs as it failed to protect a 27-point lead against Jacksonville, giving up potential points en route to a 31-30 wild card round defeat. But could more advanced computers go further, suggesting plays that could have averted those fourth downs altogether, or offering a defense that would have stopped the Jaguars on a key fourth-and-one?
It’s a natural question after AI researchers’ recent astonishing demonstrations.
OpenAI’s ChatGPT tool broke through last year by replying to any prompt with human-like responses. Within seconds, it could generate an AP-English level essay about The Metamorphosis, produce meaningful sets of code, or compose a sonnet about laundry.
Before ChatGPT, OpenAI released DALL-E, which created digital images from text prompts in photorealistic or more artistic styles. People are already using the tools to create children’s books, sermons, and pickup lines. How long could it be before AI is creating zone blitzes, too?
Microsoft is now reportedly in talks with OpenAI about a $10 billion investment that could value the startup at close to $30 billion. Alongside investors, academic researchers have pounced on the advances. Some are trying to prompt AI to create new protein-shapes that might treat cancer or prevent pandemics. Surely developing some new wide receiver route combinations would be simpler than that. Machine learning models could also help in player evaluation.
“Humans tend to overvalue things that look good to the eye but don’t necessarily matter to winning,” Zelus Analytics co-founder Luke Bornn said. “The computer is sort of agnostic to that.”
AI has already had success in the realm of games. In 2017, a Google-backed computer program famously beat the No. 1 ranked Go player in the world. In November, Meta announced it had an AI algorithm (codenamed CICERO) that could beat humans at Diplomacy, a board game infamous for the strategic and social skills needed to outwit competitors. DARPA has evidently worked on a similar project.
“It feels like we are very, very, very, very, very far away from that in football,” said Kevin Meers, a former member of the Cleveland Browns and Dallas Cowboys analytics departments who now leads the NFL product at Zelus.
Why aren’t we hearing about computer generated playbooks, with witty names like LombardAI? The simple answer is that football is far more complicated than even the most complex of strategy games.
Many of today’s AI models are built on piles of previous human work. ChatGPT, for instance, figured out what words often go together in different contexts after “reading” over 300 billion of them. DALL-E, meanwhile, was trained on 12 million pictures, “learning” how descriptions and images were generally paired together. Cicero had over 125,000 Diplomacy games to study. In contrast, there have only been about 16,000 NFL games played, not to mention how the game has evolved over that time.
Other AI actually trains against similar models, running countless simulations to find strategies that humans might not. Once again, that process isn’t transferable to football. Madden may occasionally look like real-life, but it’s not that good.
Player positional tracking, which would be necessary both for building a large data set and for simulating action, has improved significantly in recent years. But there is still a lot of room to grow, moving from noting coordinates to tracking people’s skeletons for more exact records.
Even after all that, you get to the obvious issue that NFL players are not chess pieces. Coaches also have to keep in mind dozens of men’s physical and emotional states, as well as their opponents’. Investing years into football-coaching machines isn’t a current priority either, especially with teams continuing to kick on key fourth downs.
“We’re still trying to break through on the stuff that we already know but we need to drive adoption for,” Meers said.
Ironically, that’s one area where AI may be able to help already. Current systems like ChatGPT are built to explain things, after all. Cicero is even partially built to persuade. The analytics community has long struggled to break through to sporting traditionalists; maybe a computer could make all the difference.
“When I was using analytics to create the XFL rulebook, 95% of my job wasn’t actually data science. It was sales,” says Sam Schwartzstein, who is now Prime Video Sports’ analytics expert. “There’s a barrier in how people from a data science standpoint communicate with people from a softer skills standpoint, and [ChatGPT] can help bridge the gap. That to me is really exciting.”
But Schwartzstein isn’t holding out hope for AI creating the perfect playbook—or play-caller. “Football, you cannot solve it,” he said, returning to the Jaguars game for proof.
After hours of high-flying, quick-passing action, Jacksonville coach Doug Pederson changed strategies with the game on the line. On 4th-and-1, the man who helped make the inventive Philly Special famous turned to the T formation, a creation from 1882 that was largely abandoned at the pro level by the 1950s. It was actually offensive line coach Phil Rauscher that came up with the play.
It turned into a 25-yard run that set up the game-winning field goal. How could a computer have predicted that?
“That’s what’s beautiful about football,” Schwartzstein concluded.
Chargers fans may feel differently.