GenAI Matters in Sports Too
Another Tailwind for Sports Rights Values
Sports is one of the only bright spots in media today. Sports franchise valuations keep rising. NFL ratings are up so far this season. The NBA is basking in a successful rights renewal that saw new bidders emerge. Viewership of this summer’s Paris Olympics marked a huge rebound from the record-low numbers in Tokyo three years ago. More generally, sports are one of the only TV genres that still attract younger viewers.
Meanwhile, the talk of every boardroom and investor meeting is increasingly consumed with GenAI. What, if anything, does GenAI have to do with sports programming? The concept may seem like a contradiction in terms. GenAI is synthetic. Sports features real people doing real things in real life.
As I explain below, there are important applications of GenAI in sports that could enhance the fan experience, maybe boost revenue and, as a result, add another tailwind to the value of sports rights.
Tl;dr:
The implications of GenAI for scripted entertainment and sports programming are very different. For Hollywood, it could prove a disruptive threat; for sports it’s an opportunity.
Since the main appeal of sports is its authenticity, most GenAI use cases in sports programming relate to the creation of derivative or supplementary programming that enhances the sports viewing experience, not replaces it.
These include real-time translation/localization; multi-modal conversion between combinations of text, audio, images and video; video augmented with GenAI, like dynamic real-time graphics and even virtual commentators; and the ability to create personalized feeds.
Another (eventual) opportunity for the leagues is licensing footage for model training, which might be among the most valuable data licensing deals struck so far.
An elephant in the room is the complex rights dynamics between leagues, teams, distribution partners, players associations and players. “AI rights” would need to clarified and codified for much of this to move forward.
There isn’t much forcing the leagues to act, since GenAI is more opportunity than threat. But progressive leagues will figure it out.
This post is sponsored by WSC Sports.
Learn What Generative AI Can Do for You in Sports
LLMs, text-to-image, text-to-audio, text-to-video—how does it impact sports and how can you keep up?
WSC Sports, the pioneer in automation and AI for sports content, is here to help guide the sports industry through the next frontier of generative AI products and solutions. With a sole focus on sports, WSC Sports understands the needs of sports organizations and is dedicated to helping those companies innovate.
Applications of the new area of technology are being developed for leagues, teams, conferences, broadcasters, and more.
Reach out today to learn how you can become a partner in generative AI with WSC Sports and help shape the future of sports for fans around the world.
Contrary to Perception, Sports Rights are Probably Not in a Bubble
Sports rights continue to go up, even as linear TV continues to erode. Understandably, that’s raised questions about whether sports rights are in a “bubble.”
I think the answer is no, especially for top tier rights. I outlined several reasons in Video: Follow the Money and Video: Forecast the Money. Among them: sports is the only sure thing on TV and continues to dramatically outperform entertainment; last year, the U.S. video ecosystem spent about 4X as much on entertainment programming (~$93B) as sports rights (~$24B), so there’s a lot of room for sports to take share of budget; and the recent move across the industry to embrace advertising, most notably by Netflix and Amazon, has increased the number of viable bidders for sports rights. Recent evidence clearly belies a bubble. In the last few months, we’ve seen large new bidders emerge for sports rights, such as Amazon and Comcast step up for NBA rights; Netflix acquire rights to select NFL games; and Roku acquire MLB rights.
We can add one more reason to the list: GenAI may further boost the value of sports rights.
The Fundamental Difference Between GenAI in Scripted Entertainment and Sports
The implications of GenAI for scripted entertainment and sports are quite different.
For the past two years, I’ve been writing why GenAI is a potentially disruptive threat to Hollywood (see here for a recent summary). The rationale is that GenAI will likely lower the barriers to entry for high production value scripted video and increase the supply of competitive (if not necessarily comparable) content from outside Hollywood (both from small teams and independent creators).
One of the chief reasons that Hollywood is susceptible to disruption by synthetic GenAI content is that much of Hollywood content is synthetic to begin with. The social contract between viewers and Hollywood is that the former will suspend their disbelief if the latter tell a sufficiently convincing story. But everyone knows that the scripts are made up; principal photography—the sets, locations, costumes, make-up, camera angles, physical special effects, etc.—is an artful illusion; and, these days, a large proportion of shots are edited or fabricated in post production using computer-generated visual effects (VFX). Over time, as GenAI technology advances and starts to show up on screen, viewers probably won’t even know. The distinction between a synthetic VFX shot added in post and a synthetic GenAI-enabled shot will become indistinguishable. (This will be especially true when AI features are added to popular edit tools, such as Adobe embedding Firefly in Premiere Pro and After Effects.)
TV series and movies are so susceptible to competition from GenAI-enabled content is because they are already substantially synthetic.
By contrast, sports is insulated from the potentially disruptive effects of GenAI because its primary appeal is that it’s real. People love sports for a lot of reasons, but clearly a big part is the (authentic, unpredictable, unscripted) human drama: the personal backstories, the long-held rivalries, the thrill of victory, the agony of defeat and all that. People love the humanity in sports. It’s notable, for instance, that no one wants to watch two computers play each other in a video game, but millions watch esports and Minecraft videos, which is humans playing video games.
No one wants to watch computers play against each other; in sports, the humanity is the point.
GenAI Use Cases May Both Attract and Retain Sports Fans
AI and machine learning (ML) have been used extensively in sports production for years. For the last couple of years, the Bundesliga has used AI to analyze large quantities of data in real time and provide on screen “Match Facts,” like Shot Speed, Keeper Efficiency and Ball Recovery Time. WSC Sports, the sponsor of this post, offers a platform that uses Automatic Content Recognition (ACR) (like crowd noise and commentator volume and pitch) and facial recognition, among other techniques, to automatically cut up a game into clips and rank them, making it easy for the leagues, teams and networks to publish to social or their own sites and channels. Pixellot offers a variety of products for automated streaming/broadcasting of sporting events.
The most promising GenAI use cases in sports are derivative of the source programming, not a replacement for it.
What’s new about GenAI is the “generative” part. As opposed to processing vast quantities of data for analysis, process automation or predictions, GenAI creates entirely new content. Unlike scripted programming, in which GenAI may ultimately replace parts of principal photography, since the main attribute of sports is that it is real, the most promising GenAI use cases are enhancements or derivative of the source programming, not a replacement for it.
With the caveat that all of this is subject to clarity on rights (which I’ll discuss later), GenAI opens up a bunch of cool use cases:
Localization
One of the main goals of large sports leagues (especially U.S. leagues) is to expand globally. Today, by far the most global sport is soccer, followed by basketball and tennis. Other sports have large and fanatical local followings, such as american football and cricket (in India, Pakistan, Australia, England and parts of Africa), but attract little attention outside their core markets. The NBA has been more successful growing internationally than the NFL, but either way both have made global expansion a priority.
An estimated 70-80% of the global population doesn’t speak English.
Language is a key stumbling block. An estimated 70-80% of the global population doesn’t speak English. But it’s labor intensive and costly to write foreign language commentary and record it, especially doing it 20 or more times.
AI-assisted real-time language translation has been around for a while, but it hasn’t been accurate enough for broadcasting. Since large language models (LLMs) understand (or, more precisely, have been trained on) slang and other cultural nuances—and they can be fine tuned with sports-specific language—they make it possible to automatically translate games and highlights, including both commentary and graphics, in real time, in a much more authentic way. (For instance, a few months ago Sinclair started using HeyGen to translate Petko Unfiltered, a show on the Tennis Channel, from English to Spanish.) This doesn’t just cut down on costs, it also makes it possible to reach markets that aren’t large enough to justify the expense of recording new local commentary.
Multimodal Conversion
Multimodal conversion refers to the ability to convert any combination of text, image, audio and video into any other combination of text, image, audio and video.
For instance, GenAI could be used to dynamically convert the transcript of game commentary into a simple article. It could convert an in-depth article into a text-and-image social post for Instagram; a short tweet and GIF for Twitter; a podcast for Spotify (maybe incorporating game commentary and player interviews); or a video for YouTube (or the leagues’, teams’ or broadcasters’ owned channels) by using text-to-speech and pairing key moments with video footage, player reactions, or even AI-generated visualizations of the game (like animated replays or highlight clips).
AI-Augmented Video
GenAI will also make it possible to augment the broadcast itself. As mentioned above, AI already makes it possible to process vast amounts of data and provide real-time stats (like the Bundesliga example). Using GenAI, it will be possible to automate graphical overlays that are relevant to the game, such as displaying player performance, historical comparisons, tactical formations and heatmaps. Whether the leagues want to embrace sports betting is an open question, but there are also a vast number of potential betting applications, like providing probability models for upcoming plays.
It will also be possible to apply style transfer to the game (for example, change it to anime, grainy historical footage or an 8-bit video game). (Which isn’t to say that the leagues, players or fans want this.) Likewise, it will be possible to add virtual commentators, even including fictional characters. In an effort to attract kids, it would be possible to create a highlight show anchored by a cartoon character. Or, it will be possible to do brand integrations, like Homelander giving a halftime report during Amazon’s Thursday Night Football coverage.
Personalization
In theory, it will also be possible to personalize any combination of all of the above.
Sports broadcasters are already experimenting with personalized highlights or commentary based on user behavior and preferences. For the Olympics, NBC offered a personalized recap featuring the (synthetic) voice of Al Michaels. (See below.) NBC estimated that there were about 7 million possible permutations. ESPN also recently said that it is evaluating a personalized version of SportsCenter when it launches its flagship ESPN streaming service next fall. The appeal of, say, a weekly NFL highlight reel personalized to each user’s fantasy football team is obvious. An estimated 50 million people play fantasy sports in the U.S.
Eventually, this personalization could extend beyond a customized feed of clips to include customized commentators, control of the commentary itself (constant or infrequent, emphasis on technical analysis or backstory), language, modality, on screen stats, etc.
Fantasy Simulations
Another idea that sometimes gets floated is fantasy simulations, like the 1995-96 Chicago Bulls playing the 2016-17 Golden State Warriors or maybe Michael Jordan going 1-on-1 against LeBron James. I’m skeptical that this is much more than a novelty for the reason stated above. The appeal of sports, even esports, is about the humanity in it. We’ll see.
Training Rights are Likely Valuable
Another potential opportunity for the leagues is to license training rights for AI video models. (Note that I am referring here to licensing in a way that prohibits generating content that resembles real players and, as a result, producing clearly derivative works.) This is murky territory, because none of the courts, the copyright office or Congress have clarified key legal issues around training AI models (for instance, is training a model that doesn’t create clearly derivative works infringing and require licensing at all?) and, partially as a result, the economic framework for licensing is formative. The leagues probably won’t be the pioneers in figuring this out. But, once the dust settles, it could be lucrative.
Over the last few months, there has been a wave of next-generation video models, starting with Sora and continuing with Google Veo, Luma Labs’ Dream Machines, Runway Gen-3, Kling and Minimax, among others. I haven’t tried all of these, but from what I’ve seen, they’re horrible at depicting sports because they haven’t been trained on enough sports footage.
Below is a comically disturbing attempt at a MMA video created by AI filmmaker Guillaume Hulbert using Luma’s Dream Machine.
And here’s my effort to generate footage of a baseball game using Runway Gen-3. (Here’s the prompt: “A professional baseball game, at night, viewed from the stands. A man is on first base and the batter hits the ball to the shortstop, who makes the play at first base.”) Gen-3 has very little idea what any of this means.
These examples are not meant to criticize Luma, Runway or anyone else. The point is that they will never be able to generate realistic sports scenes unless they are trained on a sufficient amount of relevant footage. This footage could be useful for other training purposes, too, such as creating more realistic sports video games or more complex and nuanced understanding of human biomechanics for future “world models.”
What might those rights be worth? To date, there are few publicly-disclosed examples of video rights holders licensing training rights to AI model providers. (Contrary to some reports, there is no evidence that Lionsgate licensed training rights to Runway as part of their recent agreement.) Most of the media deals so far are by print publishers and, even in those cases, usually terms are not disclosed (Figure 1). However, the News Corp-OpenAI deal was reportedly for $250 million over five years; the deal between Reddit and Google is for $60 million annually; and OpenAI is reportedly paying Axel Springer “tens of millions” of euros.
Figure 1. A Select AI Training Rights Deals
Source: Variety VIP+.
Sports video rights deals would almost certainly be substantially more valuable than these publishing deals because of their scarcity. While all good print publishers have some exclusive content, such as deeply researched stories and opinion pieces, a lot of news is relatively commoditized. For the frontier LLM providers, there is probably a diminishing marginal return to the next training license, which gives them bargaining power over print publishers to some degree. By contrast, the sports leagues exclusively control their respective sports, at least professionally.
A lot of news is relatively commoditized, while the leagues own exclusive rights to their respective sports, at least professionally.
Anthropic co-founder Dario Amodei made news a few months ago by suggesting that the cost to train future frontier models would increase by an order of magnitude in each generation: $100 million, then $1 billion, $10 billion and conceivably $100 billion within a few years. In this context, hundred of millions for training rights is a drop in the bucket.
“AI Rights” Need to be Clarified
So far, I’ve been dancing around the significant questions about rights. For the most part, existing rights structures don’t contemplate “AI rights,” which are not a monolithic thing. To realize many of these applications, the leagues, teams, broadcasters/networks, player associations and players will likely need to clarify and codify AI rights.
Today, rights are usually delineated something like this:
Leagues own the primary rights to most game coverage, especially national rights, and all historical footage.
Depending on the league, individual teams retain some rights for promotional use and sometimes local broadcast rights (albeit often based on guidelines set by the league).
Broadcasters, networks and streamers (let’s call them “distribution partners”) license rights to air games live, including some exclusivity period for replays and highlight packages.
Players’ associations typically negotiate collective image rights for players, usually with some guidelines about what is “collective.” For instance, while it varies by league, often the players associations don’t have jurisdiction when there are fewer than 3-5 players involved.
Players control their own name, image and likeness (NIL) rights, but they often aren’t allowed to use game footage or real uniforms for personal promotion.
GenAI use cases raise a lot of questions. While leagues typically own the rights to game footage and have broad authority how to use it, would players get a share of compensation for training rights? Does incremental “AI revenue” affect salary caps? If the leagues license training rights in a way that prevents generating footage that resembles actual players, does that require player consent or not?
Similarly, do the leagues or distribution partners’ have the right to dub players’ voices? What are the distribution partners’ rights to modify the broadcast with dynamic graphical overlays? Perhaps most important, do the distribution partners need to pay additional fees for these rights? The leagues and distributors have a strong mutual interest in any programming enhancement that brings in new viewers or helps retain existing fans. But to the extent the distribution partners are able to use GenAI to boost revenue, the leagues would probably want to be compensated.
To the extent new GenAI use cases boost distribution partners’ revenues, the leagues would probably expect to be compensated.
As I mentioned, “AI rights” is not a monolithic thing and, given how fast things are changing, it may be tough to establish a framework that contemplates every application. However, at a minimum, new and revised licensing deals and collective bargaining agreements should broadly address:
Training rights
Content analysis rights
Content modification rights
Leagues Shouldn’t Sleep on It
I don’t mean to sugarcoat the challenges here. As mentioned, the rights issues are sticky and there are a lot of constituencies to satisfy. In addition, GenAI is controversial and some of the use cases I described above will raise concerns about displacing some labor.
Nevertheless, for sports, GenAI is more opportunity than threat. There are a lot of ways that GenAI may enhance the fan experience and, as a result, the value of sports for the leagues, players, distribution partners and fans. The flip side is that, unlike Hollywood, which faces a potentially existential threat from GenAI, there is less forcing the leagues’ hands. But the most progressive leagues will figure it out.