Mind Matters Natural and Artificial Intelligence News and Analysis
blue-and-green-turf-meets-futuristic-stadium-lights-as-footb-932319723-stockpack-adobe_stock
Blue and green turf meets futuristic stadium lights as football players clash, pixelated projectiles in flight, technology merging with old-school gridiron glory.

Do Fantasy Sports Tell Us Something About Artificial Intelligence?

My biggest takeaway from my own involvement is how well fantasy football illuminates some weaknesses of artificial intelligence (AI)
Share
Facebook
Twitter
LinkedIn
Flipboard
Print
Email

More than 60 million Americans play fantasy sports, including more than 30 million playing fantasy football. Now I do too.

One of my sons manages multiple fantasy team in several sports, including football, baseball, and hockey. After several years of resistance, I joined an ESPN fantasy football league that my son set up for family and friends a few years ago.

The rules are straightforward. There are nine positions on a fantasy football team: one quarterback, two running backs, two wide receivers, one tight end, one flex player (who is generally a running back, wide receiver, or tight end), a defense/special team, and a kicker. Fantasy teams harvest fantasy points in each game based on several performance metrics, including rushing yards, passing yards, touchdowns, and fumbles.

The snake draft

At the start of the season, there is a 16-round snake draft, with the draft order reversed each round, so that the team manager who chooses last in any round chooses first in the next round.

Teams play head-to-head each week with each team’s manager choosing the nine players to use that week. As the season progresses, team managers can replace players on their rosters through trades or picking up players who are not on any team.

Enough background. On to the games. Fantasy has been surprisingly fun, even addictive. I now have a tangible stake in the outcomes. On the other hand, I am mostly interested in how individual players are doing and now pay far less attention to teams. Instead of checking how the Buffalo Bills are doing, I check how their quarterback, Josh Allen, is doing.

Fantasy football has also strongly reinforced some of my favorite theories. I have often written about the role of luck in our lives and, in particular, about the paradox of luck and skill—in competitions among highly skilled opponents, victory is often determined by luck. Fantasy managers are repeatedly buffeted by gusts of fortune and misfortune. A fumbled ball will sometimes bounce into the hands of an opposing player; other times not. Officials will sometimes call a holding penalty; other times, not. The list is very long. My fantasy team got 110 points last week. The week before, it got 210 points.

Injuries are bad luck, too

I didn’t appreciate how frequently football players get injured until I played fantasy. The San Francisco 49ers’ star running back, Christian McCaffrey, for example, was the top fantasy football player in 2023 and was the first pick in essentially every 2024 fantasy draft. But he was a late scratch for the 49ers’ first game of the season due to Achilles tendinitis and sat out the entire first half of the season. Fortunately, I didn’t have the number one pick in our league.

Regression toward the mean has been confirmed over and over as the top performers each week generally do not perform as well the week after. This regression is not due to complacency. Their stellar performances usually involve some fortunate luck that cannot be counted on to persist.

Weaknesses of artificial intelligence (AI)

My biggest takeaway, though, is how well fantasy football illuminates some weaknesses of artificial intelligence (AI). During the pre-season draft, team managers see a list of available players (and relevant statistics) sorted according to an ESPN AI algorithm’s recommendations for who to draft next. Managers can choose the players they like or use an auto-draft function that selects the algorithm’s recommendations. At the end of the draft, the AI algorithm ranks the teams, from first to worst. The algorithm can also be used to select the 9 players to start each week.

Last year, one of the teams in our league was fully AI, allowing AI to draft the team and choose each week’s lineup. The algorithm’s preseason prediction was that this team would finish first. It finished last. At the time, I thought this was an amusing aberration. But this season, too, an almost fully AI team was predicted to finish first and is currently in last place, while the human-controlled team the algorithm predicted to finish last is tied for first. The correlation between the algorithm’s pre-season predictions and current standings is –0.43. Yep, that’s a negative sign in front of the 0.43.

I know, I know. Data is not the plural of anecdote, but these anecdotes got me thinking about why AI is so bad at fantasy football—and it confirms some of AI’s fatal flaws outside of fantasy.

AI models are really good at finding patterns in data but they are really bad at deciding whether these spotted patterns can be used to make reliable predictions. The labels attached to the data might as well be randomly selected letters because, not understanding what words mean, AI algorithms have no way of distinguishing between correlation and causation.

While AI algorithms can calculate and sort useful individual statistics like average fantasy points per game for each player in each position, they struggle mightily to deal with intercorrelations among player performances. Football is a team sport and individual player successes and failures depend on the performance of each player’s teammates and the opposing team’s players.

During the initial fantasy draft, there are many considerations that AI algorithms overlook; for example, good managers do (and AI algorithms do not) move quickly for positions with thin talent and wait for positions with deep talent.

Similarly, in a 10-team league, Manager A might be drafting 9th in the third round and 2nd in the fourth round and need another running back and wide receiver. If Manager B, who has the 10th pick in the third round and 1st pick in the fourth round, has already selected two wide receivers, Manager A can choose a running back in the third round, confident that Manager B will not pick a third wide receiver at this point. AI algorithms don’t take such considerations into account. They simply advise Manager A to choose a running back or wide receiver based on which player has the higher projected fantasy points.

Good managers also think about interrelationships that algorithms ignore when deciding who to play and who to sit each week. For example, if an elite quarterback is injured and his backup is poor to mediocre, that’s usually bad for the team’s wide receivers and good for the running backs. Similarly, if a team’s key offensive linemen are injured, that’s usually bad for the quarterback and good for the tight ends.

Much of modern life is like this—a complicated web of interrelationships. A single action has far-ranging consequences that might be understood by humans but not by computer algorithms that have no way of knowing how the data they input and output relate to the real world that humans inhabit.

Computers are undeniably useful, bordering on indispensable for narrowly defined tasks. They can recall facts that we forget. They can make lightning-fast, error-free calculations that we might bungle. They don’t get tired or bored like we do. However, the claim that computers are intelligent in any meaningful sense of the word is still a fantasy—which is why they should not be trusted for fantasy sports or for real-world decisions that have serious consequences.


Gary N. Smith

Senior Fellow, Walter Bradley Center for Natural and Artificial Intelligence
Gary N. Smith is the Fletcher Jones Professor of Economics at Pomona College. His research on financial markets statistical reasoning, and artificial intelligence, often involves stock market anomalies, statistical fallacies, and the misuse of data have been widely cited. He is the author of dozens of research articles and 16 books, most recently, The Power of Modern Value Investing: Beyond Indexing, Algos, and Alpha, co-authored with Margaret Smith (Palgrave Macmillan, 2023).

Do Fantasy Sports Tell Us Something About Artificial Intelligence?