Brian Roberg's Blog

Pitch Tipping and Awareness of AI in Baseball

Someday we're going to look back on the 2025 Major League Baseball (MLB) season as the first one in which AI began to visibly affect how the game is played on the field. (It's been influential off the field as a tool for analytics for years.) As of now, though, very few people seem to have noticed.

I saw it in two ways, one of which has gotten a lot more attention than the other.

If you followed baseball this year, you may have noticed there was a lot of talk about pitch tipping. (For the uninitiated: a pitcher "tips" his pitches when he inadvertently gives subtle clues to the other team about which pitch he's about to throw.) In past years, there might be a few stories a year about pitch tipping. This year, it seemed like there was a new story every week. The difference was notable enough that The Athletic wrote an article about it (paywall).

That story examined many angles on the phenomenon, but did not mention AI. Now if pitch tipping were the only unusual event suggestive of AI's influence this year, I probably wouldn't write about it. (Pitch-tipping is a phenomenon that tends to attract attention through simple hysteria, sometimes without any actual pitch-tipping occurring at all.) But there's another data point about a different part of the game that makes me think that there's a pattern.

This season the New York Mets set a new MLB record for stolen base success rate (89.1%) over a full season. Moreover, in a stretch of games lasting from mid-June to mid-August, the Mets tied the all-time major league record for consecutive successful stolen bases by stealing 39 straight bases without being thrown out. To understand the magnitude of that streak: if you treat the league-average stolen base rate this year (77.7%) as a probability, the odds of a team stealing successfully 39 times in a row by luck is approximately 1 in 19,064. (Stats from Baseball Reference).

As striking as those records are, the manner in which the Mets stole all those bases is even more remarkable. Traditionally, the best base-stealers have two qualities: they are exceptionally fast runners, and they are lightning-quick in their reaction after the pitcher begins his motion toward home plate. The Mets' base-stealers exhibited neither of these qualities. As a team the Mets ranked near the bottom of the league in sprint speed, and some of their most prolific stealers were notably slow runners. But even more significantly, the Mets adopted a nearly uniform strategy of breaking for the next base before the pitcher committed to throwing home—an approach that conventional wisdom regards as recklessness rather than strategy. But the Mets somehow did it 39 times in a row without getting caught.

This FanGraphs post by Davy Andrews, written during the stolen base streak, does a good job of showing how early the Mets were running. You might also notice the Mets' TV announcers' incredulity that the team kept getting away with it. Andrews went deeper than the Athletic article did on the question of how the Mets became such a base-stealing machine, mentioning "analysts crunching the numbers on when pitchers do and don’t throw over, and coaches and video room personnel crushing tape to pick up on tells and tendencies." But he stopped short of mentioning AI as a contributing factor.

I believe that AI explains both the resurgence of pitch-tipping and the Mets' stolen-base success. I don't think I'm revealing any big secret in saying so, but I am pointing out that there is not yet an awareness of this among baseball fans or even those who write about baseball for a general audience.

All MLB teams have analytics departments, and AI is increasingly part of how those departments do their work. 20 of the 30 MLB teams had active AI or machine learning job postings on their websites as of a few weeks ago (according to a web research project I did using Claude). Of course, teams do not talk publicly about what their analytics departments are doing—that would nullify the competitive edge they're seeking from the analytics in the first place. But there is every reason to expect that teams are using AI to analyze video of past baseball games looking for actionable information.

In the case of pitch-tipping, it's easy to picture how this would work. Teams have easy access to datasets that collate video of pitches being delivered with highly detailed data about ball movement (and also a label showing the traditional name of the pitch: fastball, curveball, slider, etc). In other words, a team can easily feed thousands of video clips of a particular pitcher into a machine learning system, labeling each clip by pitch type. It would be a fairly straightforward machine learning task to examine video of a pitcher's motions before and during delivery, for example to compare his movements before he throws a fastball to his movements before he throws a changeup. Any pattern the AI finds that is visible to the batter in real-time is actionable intelligence.

(Note that this is entirely within the rules. Any technology is allowed as long as it's only used to analyze past events. In this sense, using AI to analyze video is just an elaboration on the time-honored practice of studying game tapes. What's forbidden is using technology to observe and act on events in real time, as the Houston Astros infamously did in 2017.)

The same general idea would apply to base stealing, though in this case (unlike with pitch data) a team would have to do more groundwork on their own. But the general shape of it is not mysterious:

  1. Assemble a database of every clip of a given pitcher when there's a runner on first base.
  2. Label each clip according to what happened (perhaps "threw to first," "pitched after looking at the runner," and (most valuably) "pitched without looking at the runner").
  3. Use machine learning to find ways the pitcher subtly indicates that he's not going to look at the runner again before he pitches.
  4. Set an MLB record for stolen base success rate.

A system like this would not be low-hanging fruit like a pitch-tipping detection system. It would require more work by analysts to put together, and probably greater machine learning expertise to design the pattern recognition system. So it stands to reason that the Mets, whose financial resources outstrip all but a few competitors, would be trail-blazers in this application of AI.

These two phenomena from the past season seem to me to be clear evidence that AI is already changing how the game is played. I think it's too early to judge whether this will be good for the game overall, but I do see two particular ways that it will affect people in the game.

First, it will change the way that teams value baseball expertise (or what more generally might be called "domain expertise"). In the past, there was a mystique to skills like detecting pitch-tipping. That kind of pattern recognition was the domain of the graybeard bench coach, the baseball lifer who could diagnose your swing at a glance while he told stories of his playing days on the Brooklyn Dodgers. Now teams are developing computer systems far more capable at recognizing patterns than the most venerable baseball whisperer.

(An interesting case study of this dynamic is playing out right now with the Mets. The public face of their base-stealing success was their first-base coach, Antoan Richardson. He was one of only a few coaches they invited back for 2026 after the team's disappointing season this year. But in a plot twist, Richardson declined to return and has instead joined the Mets' despised rival, the Atlanta Braves. Some Mets fans despaired at this, thinking that the Braves will now wield Richardson's supposed supernatural ability to read pitchers against the Mets. While Richardson certainly deserves credit as an excellent liaison between the Mets' analytics department and the players, I suspect that the Mets will still be capable base-stealers next year. Perhaps the more interesting question is whether opposing teams will be able to coach their pitchers not to telegraph when they're not going to throw over to first.)

AI technology also seems likely to disproportionately affect one other group of people: pitchers. Because the pitcher is the one who initiates every play in a baseball game, he is uniquely subject to the kind of micro-scrutiny I've described here. Perhaps it's possible to gather intelligence a pitcher could use based on how a batter stands in the box, for example, but my guess is that the scope for actionable information is always going to be more favorable for batters than pitchers. In coming years this will put more and more pressure on pitchers, which is hard because these days it seems pitchers face enough of a challenge just keeping their elbows intact.

Will AI be good for baseball overall? I don't know. If it tilts the balance of the game in favor of batters over pitchers, perhaps some adjustment will need to be made to re-balance. (How about calling the strike zone according to the rule book?) But I'm fairly confident that AI-driven analytics is not the gravest threat the game faces today.