Generative AI Allowed the PGA TOUR To Add Written Descriptions to 30,000 Shots A major goal of the generative-AI effort is more-personalized services for fans By Ken Kerschbaumer, Editorial Director Tuesday, March 25, 2025 - 3:40 pm
Print This Story | Subscribe
Story Highlights
Every week, there is more clarity about how generative AI and the cloud offer new ways to create content more efficiently. The most recent example is from the PGA TOUR, which used generative AI to create written narrative copy for the TOURCAST app, allowing fans to not only see 30,000+ shots but also read commentary and details about them.
When the Tour began looking at generative AI, according to Scott Gutterman, SVP, digital and broadcast technologies, PGA TOUR, there was plenty of experimenting, but one question was always asked: what can generative AI do to improve Tour-related content creation and, ultimately, better serve fans, players, and the tournaments themselves?
It's about scaling coverage across the entire event, he says. Typically, our staff can cover 25 or 30 golfers, but, with generative-AI content generation and data collected through ShotLink powered by CDW, we can tell the story of every single player. And it's not in a basic way - like JJ Spaun hit a 300-yard drive and has 125 yards to the hole as we have done for years and years. Now it will add context: he just hit his longest drive of the day or, at 125 yards out, he gets to within 10 ft. of the hole 20% of the time. It effectively generates two things: a fact and the context about each shot. We're excited because it gives us that foundation to tell a story for every single shot.
The use of generative AI allowed the PGA TOUR to add shot descriptions to more than 30,000 shots during THE PLAYERS.
The effort has been two years in the making, and the Tour began working with AI partner AWS, becoming a foundation partner for the AWS Bedrock generative-AI toolset. Bedrock effectively allows you to use almost any generative-AI model and a suite of tools to create these types of experiences, Gutterman explains. We are using Anthropic's Claude models to create these types of experiences. In particular, we're using the Anthropic Claude 3.5 Sonnet and have spent the past year moving away from proofs of concept and instead figuring out to operationalize it.
Moving beyond POCs involved building a set of data-acquisition and context services to take data from ShotLink and apply a rules engine so that the prompt engine knows how to use the data and apply context that makes sense: for example, after a player hits the first tee shot of the day off the first hole, it doesn't write that the player hit the longest drive of the day.
You have the data coming in and the context services that talk to a rules base that includes things like prioritizing what types of things to talk about, he adds. For example, we can tell it to talk about greens in regulations on approach shots every three narratives so that the text doesn't become redundant across all the players. We also teach the model with different ways to talk about a drive so that it isn't described the same way every time or the same way a putt would be described. All of that is fed into the Anthropic cloud model to generate the narratives.
The output narrative from Claude 3.5 Sonnet goes through a validation service to make sure the ShotLink data referred to in the output matches up with what was input into the system (for example, drive distance). The next step is cosine similarity: the system makes sure the text falls within a range of how one would talk about a drive. If it passes those tests, the text is sent to the publishing engine, where additional tests ensure complete accuracy.
During THE PLAYERS Championship, Gutterman says, the accuracy on the 30,000 shots was around 96%, which is where we thought we would be.
The system is currently text-based because the models do not process video, especially live video from an event, Gutterman points out. We are building toward a day when it'll be a combination of live data, live audio, live video and then using a multimodal output to create a video and generate a voice. We are working with AWS on a synthetic voice that can read off the prompts. We'll eventually get there.
Bedrock allows the PGA TOUR to be model-agnostic and find the best model for the task. If future models can do a function at a cheaper cost, the Tour can pivot to it without issues. You'll hear some people talking about a model that can do everything, he notes. What we're seeing is, that is not the case. We want to be able to use Antrophic for this instance but use AWS Nova, which is brand new, for things like image recognition and other work. And there will probably be another model for things like translation.
The Tour continues to build internal tools, such as story-generation tools and the synthetic-voice commentary capabilities. For five years, Gutterman explains, we have not had commentary on the 48 streams of Every Shot Live, It has just been natural sound and stats. Our goal is always to have a human telling the story, but having two commentators across 48 streams all day is cost-prohibitive. With AI, the viewer could turn on commentary the same way they turn on closed captioning. And, if they want it in Spanish, they can have that as well.
Those types of personalized services are the goal for much of the generative-AI effort, he adds. It moves us down the road of hyper-personalization, where a fan can get a story at the end of the day with the best video from their favorite players. And we can start to do things like predictive engagement, where the app already knows what you like and just serves you what you want.
