How Gamedai Calls A Play Without Making It Up

The problem nobody talks about

A lot of writing about AI in sports assumes the only thing that matters is making the computer sound human. That part is mostly solved. The voices are good enough now that you would not always be able to tell. The real problem, and the one that has kept live AI sports broadcasts out of consumer hands for years, is honesty.

A computer that has read a lot of internet text is also a computer that has read a thousand wrong tweets, a thousand fake play descriptions, a thousand parody recaps. Ask it "what just happened in the Eagles game?" without a real live feed, and it will confidently invent a touchdown. The Stanford AI Index reported that even the best AI models still make things up between three and seventeen percent of the time on open-ended factual questions. Over a single NFL game, with two hundred plus plays, that translates to six to thirty-four wrong calls per broadcast. Listeners would pull their earbuds out by halftime.

The fix is not a smarter AI. The fix is to wrap the AI in a system that does not let it lie.

The library of football knowledge

The first thing the system needs is a structured library of everything that matters in football. People in the field call this a "knowledge graph," which is just a fancy way of saying an organized library a computer can read fast. What goes in it:

• Who. Every team, every player, every coach, every stadium, every league, every fantasy roster.
• What. Every play, every score, every drive, every penalty, every injury, every timeout.
• How. Every formation, every defense, every pre-snap motion, every personnel grouping.
• Why. The football concepts an announcer would use to explain a play. Cover 2, RPO, the run-pass option, the run game, the screen game.
• Before. Ten years of NFL play-by-play, organized so that a current third-and-two on the opponent's 35 can find its nearest historical cousins in a few milliseconds.

The data inside the library comes from three sources fans already know. nflfastR provides the live play-by-play feed. PFF charts formation, coverage, and pressure. Big Data Bowl 2026 added player-tracking data, frame by frame, that did not exist five years ago. Together, they cover what happened, what was charted about it, and where everybody on the field actually was.

Step one: get the facts straight

The first thing the system does, every play, is figure out the truth of what just happened. It reads the live feed and outputs a tidy record: play number, down and distance, formation, defense, who carried the ball, what the outcome was.

This step is strict. It does not guess. It does not interpret. It uses fixed rules, the kind a rookie scout could write down in a notebook. If this step is wrong, everything that follows is wrong. So we do not let any creative AI touch it. The rules are the rules.

Step two: look it up in the library

With the facts of the play in hand, the system looks the play up in the football library. It pulls back a short bundle of useful context. How often does this team pass on third and two from this part of the field? What does the league do, on average? What defense was the opponent in, and what tends to happen when a quarterback faces that defense?

That bundle is small on purpose, because the next step has to be fast. We aim for the library lookup to take around two to four tenths of a second on a warm system. Anything slower and the broadcast lag gets noticeable.

Step three: the voice writes the sentence

Now the actual AI voice gets involved. It is given three things: the facts of the play (from step one), the context bundle (from step two), and your preferences (which voice you picked, how deep you want the football talk, which players are on your fantasy roster). With those, it writes two to four sentences of broadcast copy.

Importantly, the voice does not get to make up new facts. It can comment on the play. It can compare it to history. It can react. But every number it uses, every player name it mentions, every outcome it states, has to come from the bundle the library handed it. Anything outside the bundle is off limits.

Step four: the safety check

The last step before the broadcast goes into your ear is the safety check. It reads what the voice wrote and runs through a checklist:

• Does the sentence say anything about the play that does not match the facts from step one? Wrong player, wrong yardage, wrong down?
• Does the sentence use a number that is not in the library lookup? (We do not let it cite a statistic the library did not give it.)
• Does the sentence stay in the right voice? Coach Bill cannot suddenly use a phrase that does not fit Coach Bill.
• Are there any banned phrases, or any words that would falsely imply we are an official NFL partner?

If the sentence fails any of those, the broadcast falls back to a simpler factual call written from the facts directly. The listener might hear a slightly drier call. They never hear a wrong one. That is the right trade.

The clock budget

The end-to-end goal is fifteen to twenty-five secondsfrom snap to your ear. Roughly, that breaks down as:

• About 5 seconds for the live data feed to land.
• About 3 seconds for the facts step.
• About 2 seconds to look up the library.
• About 3 seconds for the voice to write the sentence.
• About 3 to 5 seconds to turn the text into speech and deliver it to your phone or smart glasses.

We could go faster by skipping the safety check or using a smaller voice model. Both would cost accuracy. The fifteen-to-twenty-five window is the right balance for a live broadcast you can trust.

Why this approach works across sports

The five steps are the same for every sport. What changes is the football library underneath. A cricket bowler does not throw Cover 2. A Formula 1 car does not run a screen pass. A basketball pick-and-roll has nothing to do with a soccer 4-3-3. Every sport needs its own structured library of what matters in that sport.

That is what makes going wide possible. Once one station is live, standing up the next one is a question of building the library. That is real work, but it is bounded work. We do not have to rebuild the system every time. The infrastructure travels. The knowledge does not. The roadmap tracks each sister station and its target launch date.

What this opens up beyond the radio

The interactive sports radio is the first thing we build on top of this. It is not the last. Once you have a real-time, organized, honest library of every play in NFL history, a bunch of products become possible:

• Fantasy highlight reels that build themselves in seconds after the game ends, custom to your roster.
• Coach-mode broadcasts for amateur teams that want NFL-grade tactical commentary on their own film.
• Conversational sports research for betting and analysis, grounded in real play data rather than internet guesswork.
• Open data for developers, scheduled to be available in 2027.
• Creator stations on rAIdio, the platform we open after the gamedai launch. Anyone who wants to run their own sports broadcast can use the same plumbing.

The radio is what fans hear first. The honest, organized library underneath is what makes everything else possible.

The point

The thing that wins the next decade of sports media is not the biggest AI model. It is the most honest, lowest-delay, widest sports library, and the discipline to let the AI only describe what the library confirms. Gamedai's five-step pipeline is one version of that discipline. The broadcast you will hear on August 7, 2026 is what comes out the other end.

Join the waitlist. If you want a more technical walkthrough than this one, the how-it-works page goes deeper.