When AI shows up in the dugout: an enterprise risk read on MLB's third-tab ban
MLB banned in-game AI assistants after a third of teams ran a hidden third tab on league-issued iPads. The story is a clean read on how regulated industries respond when AI arrives faster than the rulebook.
The story, in one paragraph
MLB recently told its clubs to stop using AI assistants mid-game. The trigger was not a cheating scandal in the usual sense. It was a workaround. Every dugout carries a league-issued iPad for stats and video. A handful of teams, reportedly starting with the Marlins, built a custom third tab that piped the live game into an AI assistant. Pitchers were getting suggested pitch calls between innings. Substitution decisions were getting second opinions from a chatbot with a longer memory than any human coach. By the time the league caught on, at least seven teams were using it. The rule the league had on the books did not cover it, so MLB issued a memo, gave everyone a month, and killed the third tab entirely.
For enterprise risk and AI governance teams, the story is worth more than the headline. It is a clean illustration of how a regulated industry responds when a generative AI capability arrives faster than the rulebook that governs the setting where it shows up. The mechanics of that response are what boards, regulators, and audit committees will eventually expect from any organization that operates inside a regulated perimeter.
Why the workaround mattered more than the cheat
Read the original reporting closely and the interesting move is not the AI itself. It is the routing. The teams did not bolt on a new device. They did not ask the league for permission to bring a phone into the dugout. They used the iPad the league already issued, opened a third tab the league did not think to scan for, and let the AI run inside an authorized surface. That is the same shape that shows up whenever a new capability meets an existing permission boundary: the people who want the capability do not push the boundary outward, they tunnel through whatever is already there.
The same pattern is visible inside enterprises. A sanctioned productivity tool gets a browser extension that wraps a model around it. A sanctioned file-share gets a script that summarises whatever lands in it. A sanctioned CRM gets a sidebar that drafts every outbound message. Each of those is a third tab on an iPad the company already handed out. None of them, individually, broke a rule. Together, they routinely outrun the controls that govern the underlying surface.
The governance question that follows is not whether the rule book is correct. The governance question is whether the rule book is even looking in the right places. Most enterprise AI policies still assume the AI shows up as a new tool, with a new login, on a new device. The MLB story is a reminder that it does not. It shows up as a tab.
What a regulator actually does when the rulebook lags
MLB's response sequence is a useful case study for any compliance officer. The league did not run a forensics investigation. It did not impose fines or revoke draft picks. It sent a memo, named the surface, gave teams a one-month runway, and closed the surface. That is the minimum-viable regulator move when the underlying conduct was not technically illegal at the moment it occurred. It is also, almost exactly, what the CFPB did with auto-lending add-ons in the early 2010s, what the FDA does with off-label deployment of cleared devices, and what financial regulators have done repeatedly with structured products that fit the letter of existing rules while doing things the rule drafters did not anticipate.
The pattern is consistent: name the surface, set a horizon, close the surface, then write the rule that would have caught it. Enterprise AI programs that watch that pattern can plan for it. The companies that get caught by it are the ones whose AI posture was organized around the rulebook they had on January 1 instead of around the surface their teams were actually using by March 1.
The accountability boundary that did not exist
The cleanest detail in the MLB story is the quote attributed to a league executive: "gotta stop the cheating before there's cheating." It is the kind of line that reads as a throwaway and is actually the whole point. The AI was not generating fraudulent signals. It was not spoofing anything. It was, however, sitting inside an authorized surface and producing advice that competed with the human authority structure the league had spent a century defining. The cheating, in the executive's framing, was the slow erosion of where the human authority lived, not any specific act inside a pitch.
That distinction matters in enterprise settings, because most current AI governance is built around the wrong axis. The frameworks audit whether the model produced a correct output. They do not yet routinely audit whether the human authority structure that surrounds the model is still doing the work it was designed to do. A loan officer whose decision is upstream of an AI summary is a different accountability actor than a loan officer whose decision is downstream of an AI summary. The regulators who eventually catch up to this will not be drawing lines around outputs. They will be drawing lines around authority.
For boards, the implication is that AI governance work has to extend past model risk and into the authority architecture that surrounds the model. Who can override whom, on what surface, with what logging, is now a first-class governance question rather than an HR question.
What enterprise teams should take from this
Three operational moves follow from the dugout story, and they map cleanly onto regulated industries outside sports.
First, audit the tabs. Most enterprise AI exposure right now lives inside tools the company already approved. Scan the sanctioned stack for embedded model surfaces the original approval did not anticipate. Browser extensions, third-party plugin marketplaces, sanctioned-tool sidebar apps, and BYO model integrations through approved APIs are all third tabs in spirit. Treat them as such. The probability that a generative model is already running inside one of your authorized surfaces is much higher than the probability that it is running on a shadow device.
Second, set the regulator's runway before you need it. Decide, internally, the maximum acceptable lag between an AI capability appearing inside an authorized surface and the company having a written policy that governs it. A 30-day window is a reasonable starting point for most organizations. A 90-day window is what most regulators will tolerate. Anything longer is what gets you caught flat-footed when the memo arrives.
Third, separate model risk from authority risk in the governance committee. Model risk is the easier conversation. Authority risk is the conversation about whether the humans downstream of the model still own the decision, the audit trail, and the consequences. The two require different stakeholders, different metrics, and different reporting cadences. Combining them tends to produce a governance posture that handles neither well.
The MLB story will fade from the news cycle within a week. The pattern it illustrates, an AI capability arriving inside an authorized surface faster than the rulebook that governs that surface, is the pattern enterprise AI governance programs will be working through for the next several years. The companies that read it correctly now will be the ones that look composed when the memo arrives for them.
Book a strategy consultation to map these patterns to your own regulated AI rollout.