Just days before this year’s Paris Olympics kicked off, my colleague Alex Kirshner argued that the Games were a real now-or-never moment for Peacock. The 4-year-old NBC streaming service has lagged behind its more established rivals in terms of popularity, customer satisfaction and market share. So naturally, NBC has been touting plenty of goodies in the months leading up to the opening ceremony, including “multiview” livestreaming, special correspondent Snoop Dogg and NFL-style “Gold Zone” coverage.
Perhaps the most questionable offering, however, was the use of artificial intelligence: a virtual likeness of veteran sportscaster Al Michaels was created with his permission, trained to sound exactly like the iconic commentator, and engineered to provide individual fans with personalized “daily Olympic Games summaries.”
There was certainly some skepticism about Peacock’s ability to provide a proper recreation of the popular sports reporter at scale, especially as the much-hyped generative AI boom continues to deliver spotty results. (Not to mention that viewers of Google’s Olympics AI ad were so pissed that the company eventually pulled it from the schedule.) But Michaels himself was impressed with NBC’s early results, as he told Vanity Fair’s Tom Kludt in his real voice: “Frankly, it was amazing. It was incredible. And it was a little bit scary.” Last week, The Washington Post’s Shira Ovide also called the “Al Michaels” recaps “surprisingly good.” So, as a dedicated Olympics viewer, I knew I had to not only evaluate this for myself, but find out. Why it worked so well if the rumors were actually true. And, reader: They were true.
From my laptop (the feature isn’t available in the TV app), I accessed Peacock’s “Your Daily Olympics Recap” every morning for the past week. To set it up, I had to enter my name, select the three sports categories I was interested in, and then the two “themes” for the clips I most wanted to rewatch daily (like “viral and trending moments”). I was a little disappointed to later find out that I couldn’t change those customizations if I wanted to, but I understood that kind of limitation on a brand-new application that “hundreds of thousands of users” have signed up for, as Peacock’s president told Fast Company. (Incidentally, the viewership numbers Peacock has been pulling for the Games this year are setting crazy records.) The recaps, appropriately, appear in a URL tagged “voice-ai.”
The segment, sponsored by Microsoft’s “AI companion” Copilot, contained between 14 and 18 segments, averaging one minute in length, bookended by an introduction and a farewell from “Al Michaels.” And I have to say, I was amazed at how solid, fluid and smooth they usually were.
The first sports category would appear with three floating clips of the day’s biggest events – produced by the NBC team and pulled here with the help of AI – as “Al Michaels” quickly and pointedly commented on everything from Malaysian badminton victories to a single athlete’s knee injury. This would then give way to a series of individual, full-screen highlight reels of each sport and game – no AI voices here, just moments and moods. Then another sports category would follow, and finally a collection of “Must See” or “Viral” moments would float by, succinctly contextualized by the voice bot, before signing off with a preview of what would be on NBC’s primetime that evening.
There were many entertaining aspects of this program: The voice of Al Michaels not You can sound like a tinny robot, watch your favorite Olympics events in full screen, and flip back and forth through the series at will. Plus, you’ll get a brand new device at the start of each day that’ll catch you up on everything you missed the day before.
I did spot a few issues here and there, particularly with the auto-generated subtitles (anyone know a “Noah Lyons”? Or someone with “tie-tie parents” instead of “Thai parents”?). I also found “Michaels” to be a little flatter than I expected on occasion—for example, when I was commenting on the France-Egypt soccer match, the Michaels bot almost muttered that the game had “a much happier ending” for France than for Egypt. (A similar circumstance was when it noted that Germany’s basketball competition provided “another exciting” development.) While the clips were mostly exciting, they could also be disappointing; a “must-see” highlight on swimming just showed the competitors diving in, while for others breaking records in climbing or track and field, I saw maybe 30 seconds of actual action and then another 30 seconds of their celebration, with no other real sporting highlights.
Still, I was impressed by how quickly and seamlessly these entire presentations came together day after day, drawing from full hours of various sports to give you a concise recap the next day. From Peacock’s own marketing, I knew that generative AI was “used both to recreate Al Michaels’ voice, trained (with his consent) on his previous appearances at NBC, and to create the commentary for the intros and overviews of the clips using the AI voice.” There’s also an element of human review at the end for final cleanup and editing. But I just wanted to learn How Here, AI was fully utilized for all parts of the equation: the commentary, the clips, the design.
I reached out to the folks at NBC to find out. They told me that right at the end of each day’s Olympic rounds (so around 4 or 5 p.m. ET), a large language model goes to work synthesizing all sorts of information available on Peacock—schedule, lineups, winners, losers—about the day’s games and outputting short summaries for the Al Michaels bot. “We work with a large language model for the text by using a type of ‘prompt chaining,’ where we break a very complex task down into a series of smaller, connected steps to produce more accurate, higher-quality output,” John Jelley, senior vice president of product and UX at Peacock, told me. “The key inputs include the metadata of the events – what sport it is, who the athlete is – and incorporating information from the subtitles.” Human editors then adapt the generated text to what the real Al Michaels would spontaneously say, while also correcting the text for typos – particularly regarding the names of the athletes.
Al Michaels’ voice also uses a different LLM, one that is not only trained on Michael’s previous broadcasts, but also “enhanced” with the names and word pronunciations to be used in this year’s Olympics. The resulting output is also subject to human review.
“Our Peacock product and data science teams continue to optimize the sound of the voice to make it as close to Al Michaels’ as possible by adjusting several variables: breathing, tone, speed and intonation,” Jelley said.
The clips come from a daily library created by NBC’s “Highlights Factory” in Connecticut, which consists of human videographers who cut out the best moments for use in NBC’s external media (TV lineup, YouTube, social media, Peacock Reels, etc.). While an AI system also helps select the appropriate clips for the daily recaps based on matching metadata, human reviewers run a scan to make sure the predicted clips are appropriate for the purpose of a particular sport or category. These editors at every step of the process are drawn from across NBC Sports, working to make sure their AI won’t fail in public like others’. After all, the machines will always be as fallible as the humans who make them work. If an Al Michaels bot that sounds as good as this one still requires as much manpower to look like a disembodied version of the real man, it’s hard to imagine a “fully” automated host gracing the Los Angeles Olympics.
In a way, this is an Olympic team effort in itself and an excellent demonstration of the possibilities of how generative AI can actually be used for a Additive and efficient use case – not one designed to displace anyone, least of all the great Al Michaels himself. That is not to say that all Olympic AI has been so benevolent; just ask the Parisian tourists who have to deal with the sophisticated AI-driven surveillance systems that are installed all over the city and will probably stay there after the Olympics. But at least at home, you could do a lot worse than robotic AI Michael’s.
