In a conference room at YouTube headquarters in San Bruno, California, CEO Neal Mohan chuckles. A week earlier, OpenAI had unceremoniously announced it was shutting down Sora, its popular app for creating AI-generated video clips. Once seen as a flagship product and the future of AI video — it attracted a $1 billion investment from Disney — the sudden shuttering had rippled through the AI industry like a bombshell. “Oh boy,” Mohan tells Forbes when asked about the shutdown. “Well, I was as surprised to hear about it as maybe you were.”

YouTube is the indisputable king of online video. With 2.7 billion users, it's also hooked into one of the biggest, most important AI companies in the world: Google. That makes Sora's shut down, from arch-rival OpenAI, something of a blessing for a business navigating the rapidly shifting AI landscape. One less rival — YouTube Shorts released its own version of Sora’s most viral feature in April, which lets users create digital avatars of themselves — but also a harbinger of how fraught it is to generate, host and share AI videos.

For more than a decade, YouTube has faced its fair share of scourges, like accusations of radicalizing users or harming their mental well-being. But AI is a radically different challenge. It has the power to transform the site completely, from how people make content to what they consume. Mohan doesn’t downplay it. “This is a profound paradigm shift, and the technology is going to dramatically change how things are done,” he says.

The AI explosion, first and foremost, means more content — and more money for YouTube’s $60 billion annual revenue business. AI is already supercharging what creators can make, bringing down production costs and unlocking new ideas and business prospects. How-to videos, a mainstay of YouTube, can now be generated with a few simple prompts. And AI is also revolutionizing how quickly and cheaply marketers make the ads that are the site’s economic engine. There are now an estimated 29 billion videos in total on the platform, according to a January report from the research firm Omdia, with accelerating growth driven by factors like AI-generated video and the popularity of Shorts.

At the same time, AI means spammers flooding the zone with unbelievable efficiency. There’s the menace of deepfakes, which have already become a problem: Last year, an AI-generated version of Nvidia CEO Jensen Huang hawking a cryptocurrency scam during a keynote racked up more views on YouTube than the actual event. YouTube is also facing a growing mountain of slop, according to a November report from video editing company Kapwing, which estimated that more than 20% of the content the YouTube Shorts algorithm was showing to new users was AI-generated. In response to the study, YouTube terminated several channels that violated its spam policies. “This standalone and unverified study is not an accurate representation of what’s on our platforms,” a YouTube spokesperson said in a statement. “When users first join YouTube, they see a wide variety of content as they start to express their interests, which helps tune their feeds.” In Forbes’ own testing, after playing 200 YouTube Shorts videos on an already-established account, 17.5% of them were AI-generated.

“What happens when someone sees Mickey Mouse spitting a Kendrick Lamar line that Disney may not want?”

If the company lets low quality AI video flourish, viewers could get fed up. So as AI proliferates across YouTube, the company has a treacherous tightrope to balance — allowing AI to give it the glut of content that serves as its lifeblood, while preserving the human authenticity that’s made it one of the most popular sites in the world since it launched in 2005. “Nobody wants a feed of just AI slop,” Mohan says. But at the same time, the goal is “also letting the amazing creativity that will come from AI shine through. I think that's not a trivial problem,” he says.

One former YouTube executive is optimistic that slop won’t win out. “The reality is, the algorithm is so strong and consumers will vote with their hours that the slop will get deprioritized anyway,” they say. “Even though there may be an initial period that looks so much like garbage.”

Right now, YouTube is in the messy middle. In fighting slop, the company has a lot of constituencies to keep happy, including the creators, music studios and media companies that provide much of its most popular content. Resistance to AI has been particularly pointed for the human artists whose work has been scraped and stolen to train models, and for the powerful holders of their copyrights. “What happens when someone sees Mickey Mouse spitting a Kendrick Lamar line that Disney may not want?” the former YouTube executive says.

Have a tip? Contact reporter Richard Nieva at rnieva@forbes.com or RNieva.26 on Signal.

It’s a dynamic that Mohan is acutely aware of. YouTube is “a company that faces the creative industry every single day: Hollywood creators, music and media companies, et cetera. And I do think it gives us a unique vantage point when it comes to AI and its impact on creativity,” he says. “The heart of YouTube will always remain human.”

YouTube was a distinctly human endeavor from the start. The first video ever uploaded on the platform, in 2005, is an awkward dispatch from the San Diego Zoo by cofounder Jawad Karim, who explained why elephant trunks are cool. Karim cofounded the site with Steve Chen and Chad Hurley, who originally envisioned the platform as a dating site. That concept quickly failed, but the founders knew they were onto something by making it easier to upload and host videos. They sold the startup to Google a year later for $1.65 billion, where it grew to become the second most-visited website in the world, behind only Google itself.

One of YouTube’s primary solutions to the AI slop problem is labeling. Much AI-generated content is so lifelike that the casual viewer might not realize they’re not looking at something real. Last month, YouTube said it would prominently label “meaningfully AI altered or generated” videos, using internal tools to sniff them out even if they weren’t disclosed by the creator.

It’ll also have to actively police copyright and harmful content, a familiar challenge for the company. In 2017, YouTube Kids, a version of the site specifically for children, came under fire when disturbing content slipped past its safety filters, like one video of Paw Patrol characters dying, and another of Nick Jr. characters at a strip club. AI has the potential to make those kinds of videos easier and faster to create. Some creators are already targeting AI slop at babies.

As much as YouTube wants to preserve the site’s human heart, it’s rolled out a steady stream of new AI features. Last year, it launched an Ask button in the YouTube video player that lets users ask questions about a video, like how to plan an itinerary around the sites mentioned in a Croatian travel guide video. A new search tool lets users find videos as if they were prompting a language model, similar to Google’s AI mode.

But the AI avatar feature, similar to Sora’s Cameo tool, has far greater potential to change YouTube’s soul. Launched in April as part of YouTube Shorts, the company’s short-form video rival to Tikok and Instagram Reels, the feature lets users create digital clones of themselves to star in AI-generated videos. Think: You winning the Super Bowl, or walking on the moon, or playing the banjo on your own made-up 1980s cable access show. YouTube declined to share specific stats on the feature’s popularity.

For now, YouTube’s AI Avatar feature is limited. People can only create and control avatars of themselves. But it points to a future where YouTube is populated by fake people in fake situations. In the meantime, YouTube has developed a security tool called Likeness Detection, which will police the platform for unauthorized AI versions of yourself. (You need to upload a photo of yourself to enable it, but YouTube vows it won’t use it for any other purpose.)

“I like being able to be in front of the camera and talk with my audience, and I'm not sure if an AI could replicate that.”

YouTube’s other AI features are aimed at creators, like one product called Ask Studio that uses Google’s Gemini AI model to help creators script their videos, dub or translate audio, and come up with ideas for future videos. But while those tools can help with the creative process, some creators worry about AI diminishing their artistic role.

Brooke Ashley Hall, who runs a channel focused on her family called the Beverly Halls with 11.4 million subscribers, says she probably wouldn’t use the AI avatar feature to digitally clone herself for her channel. “I like being able to be in front of the camera and talk with my audience, and I'm not sure if an AI could replicate that,” she says. But she often uses AI to generate images of herself and her family for thumbnail photos, as well as to crunch analytics, brainstorm video ideas and produce special effects. “I don't feel like it's going to replace all creators,” she says. “I feel like it's going to replace the creators that don't integrate it.”

Meanwhile, creators are figuring out what role they want to actively play in training AI. They can opt-in to YouTube sharing their videos with AI labs and other third parties to train new models. They don’t have to allow it, but if they do there’s no compensation. Today, YouTube says about a million have opted in, still a small slice of the estimated 69 million active creators on the site, according to the research firm Social Blade. (YouTube declined to disclose how many creators are on the platform.) Some creators have posted on Reddit that AI companies are reaching out to them to license their video data, some offering up to $100,000 for every 1,000 hours of video. Hall says she’s “considered” licensing her video data, and would “probably” do it. “There should be some kind of gain for us as the original creators of the content,” she says.

As YouTube grapples with the promise and peril of AI, at the end of the day, whatever encourages more content and more time spent watching it will win out, the former YouTube executive says. “Ultimately, YouTube, as a platform, cares about watch time,” a former executive says. “They don't really care as much — or they care secondarily — what that time is spent on.”

More from Forbes

ForbesCustodians Are Crypto's Boring Backbone. Now They're Taking OverForbesInside Donald Trump’s Billion-Dollar Golf EmpireBy Brett KnightForbes$1,000,000,000,000: What $1 Trillion Means These DaysBy Matt DurotForbesHow SoFi Went From Fintech Darling To Wall Street PariahBy Jeff KauflinForbesForget Elon’s Data Centers In Space. This Startup Wants To Float Them At SeaBy Alan Ohnsman