YouTube’s AI life-or-death moment: the CEO made a bold decision

How should the world’s largest video website deal with the incoming wave of AI-generated content, while also leveraging AI to amplify the real human creativity that originally helped power the platform’s rise? CEO Neal Mohan is trying to strike a balance between the two.

In a conference room at YouTube headquarters in San Bruno, California, Neal Mohan couldn’t help but laugh. A week earlier, OpenAI unexpectedly announced the shutdown of its popular AI video short-form generation tool, Sora. The product had been seen as a benchmark in AI video and a direction for the industry’s future development, and it also received a $1 billion investment from Disney. Its sudden shutdown was like a bombshell, triggering ripple effects across the entire AI industry. When asked about it, Mohan told Forbes, “Oh, honestly, like everyone else, I was very surprised when I heard the news.”

YouTube is indisputably the global leader in online video, with 2.7 billion users, backed by a top-tier AI enterprise with global scale and far-reaching influence: Google. For YouTube—looking to find opportunities in an AI landscape that changes by the minute—OpenAI shutting down Sora is, to some extent, a good thing. On one hand, the market loses a competitor: YouTube Shorts rolled out in April a tool designed to match Sora’s breakout feature, enabling users to generate custom digital virtual personas. On the other hand, the incident also serves as a warning, signaling that risks may be hidden throughout the full process of AI video—from production and storage to distribution and sharing.

For more than ten years, YouTube has faced all kinds of thorny problems, such as accusations that it intensifies users’ extreme beliefs and harms users’ mental health. But the challenges brought by AI are fundamentally different: they will rewrite the platform’s ecosystem entirely, with no exception across everything from content creation methods to users’ choices of what to watch. Mohan didn’t downplay the impact of this shift: “This is a profound paradigm shift, and this technology will significantly change how the industry operates today.”

The most direct change brought by an AI technology surge is an explosion in content volume, creating more revenue for YouTube, which has annual revenue of $60 billion.

AI greatly lowers the barriers for creators, cuts production costs, and gives rise to new creative ideas and business prospects. Tutorial videos are a core content category for YouTube; now, all it takes is entering a few simple prompts to generate them. Advertising is YouTube’s core revenue pillar, and AI also boosts marketers’ efficiency in making ads while significantly lowering costs. A report released by market research firm Omdia in January showed that YouTube’s current total video count is about 29 billion, and the booming AI-generated video and short-form video segments continue to drive content numbers to grow at a rapid pace.

At the same time, AI enables spammers to flood the platform with junk information at astonishing speed.

The threat posed by deepfake content has already become a problem. Last year, an AI-synthesized video surfaced in which Nvidia CEO Jensen Huang hawked a crypto scam through a keynote speech. That video even surpassed the playback count of the original footage from the real event on YouTube. Video editing platform Kapwing released a report in November last year stating that YouTube is facing a problem of being flooded with low-quality AI content. The report estimated that among the content pushed by short-form algorithms to new users, more than two out of every three is AI-generated. In response to the report, a YouTube spokesperson issued a statement: “This standalone, unverified report cannot accurately reflect the full picture of platform content. When new users just sign up, the platform will recommend a variety of content to help users discover their interests, thereby adjusting the subsequent recommendations feed.” Forbes also conducted its own test: using a long-established mature account to play 200 YouTube Shorts back-to-back, it found that 17.5% were AI-generated content.

If the platform allows low-quality AI videos to run rampant, viewers will eventually grow tired. As AI content pours heavily into YouTube, the platform is walking a dangerous tightrope: it needs to rely on AI to produce content continuously—that is the foundation for the platform’s survival—but it also must preserve the authentic feel of real human original work. It is this authenticity that helped YouTube rise quickly to become one of the most popular platforms worldwide since it launched in 2005. Mohan said: “No one wants to scroll through a screen full of cheap, shoddy AI content. But at the same time, we want high-quality ideas generated by AI to be fully shown. That is far from easy.”

A former YouTube executive who is optimistic believes that low-quality AI content ultimately cannot occupy the mainstream: “In fact, the algorithmic mechanisms are quite robust. Viewers effectively vote with watch time, and low-quality content will ultimately be given lower recommendation weight—it's just that in the short term, the platform will be filled with a lot of rough content.”

Right now, YouTube is in an awkward situation where it’s stuck between a rock and a hard place.

While cracking down on low-quality AI content, the platform also has to consider the feelings of various related groups, including creators that support the majority of the platform’s popular content, record labels, and media organizations. Many real human creators strongly resist AI technology. Their original works are scraped and stolen to train AI models, and copyright holders have also raised fierce objections. The former executive mentioned above said: “What if someone makes an AI video of a Mickey Mouse cover of Kendrick Lamar, and that’s something Disney doesn’t want—how would you even handle that?”

Mohan is fully aware of how complex this situation is. He said: “YouTube deals with the creative industry every day—connecting with Hollywood creators, record labels, and major media groups, and so on. That also gives us a unique perspective when thinking about AI’s impact on the creative industry. YouTube’s core is always human creation.”

From the beginning, YouTube has relied on real human originality. The first video uploaded to the platform in 2005 was a somewhat awkward short clip shot by co-founder Jawed Karim at the San Diego Zoo, in which he talked about the wonders of an elephant’s trunk. Karim, along with Steve Chen and Chad Hurley, founded the platform together. Initially, the three imagined building it as a dating site, but that idea quickly failed. They then realized that simplifying video uploads and video hosting contained a huge opportunity. A year after founding the company, the three sold it to Google for $1.65B. Since then, YouTube has grown into the second-highest-traffic website in the world, second only to Google’s homepage.

One of YouTube’s main solutions for addressing low-quality AI content is content labeling.

Many AI-generated videos are so convincing that ordinary viewers struggle to tell that the content is not real footage. In May, YouTube announced that the platform will put clear labels on videos that are “substantially modified by AI or fully generated by AI.” Even if creators don’t proactively declare it, internal detection tools can identify such content and apply labels.

The platform also has to actively police copyright infringement and harmful content, which is a long-standing challenge for YouTube. In 2017, YouTube Kids, the platform’s child-only section, was widely criticized after large amounts of disturbing and inappropriate content bypassed safety filtering mechanisms, such as scenes showing a Paw Patrol character dying, or Nickelodeon Kids characters appearing in strip clubs, and more. AI technology will significantly lower the barrier for producing such bad videos and shorten production cycles, and there are already creators who specifically produce low-quality AI content aimed at infants and toddlers.

Even as YouTube zealously protects the core of real human creation, it continues rolling out multiple AI features. Last year, the player introduced a “Ask” button, allowing users to ask questions about the video content—for example, planning an itinerary of attractions based on a Croatia travel guide video. There is also a brand-new search tool that lets users input prompts to search for videos, similar to how they would operate a large language model—functionality akin to Google’s AI search mode.

But an AI virtual persona feature similar to Cameo from Sora is even more likely to fundamentally change YouTube’s core.

The feature launched in April as a companion to YouTube Shorts, which is the platform’s short-form product designed to compete with TikTok and Instagram Reels. Users can generate their own digital avatars, letting the avatars star in various AI-generated videos. A few examples: winning a Super Bowl and walking around the moon, or playing a banjo in your own fictional 1980s cable TV program. YouTube refused to disclose specific data on how popular the feature is.

At present, YouTube’s AI virtual persona feature still has many limitations: users can only generate and control their own digital avatars. But the feature suggests that in the future, the platform will be filled with virtual human figures situated in fictional scenes. Meanwhile, YouTube has developed a safety tool called Likeness Detection to patrol and remove content where users’ unauthorized AI virtual personas are generated. Users can enable the feature by uploading a photo of themselves, and YouTube promises not to use those photos for any other purpose.

YouTube has also launched multiple AI features for creators. One tool called Ask Studio connects to Google’s Gemini large model. It can help creators write video scripts, voiceovers, translate audio, and generate ideas for future video concepts. While these tools can assist the creative process, many creators worry that AI will weaken their role as the central artistic driver.

Brooke Ashley Hall runs a family-focused channel, “The Beverly Halls,” with 11.4 million subscribers. She said she’s “very likely” not to use AI digital stand-ins to appear in her channel’s videos: “I like talking to the audience standing in front of the camera, and I’m not sure AI can replicate that kind of real interaction.” But she often uses AI to generate images of herself and her family to use as video thumbnails, while also relying on AI to analyze back-end data, brainstorm topics, and produce special effects. “AI won’t replace all creators—it will only replace creators who aren’t willing to proactively use AI tools.”

At the same time, creators are thinking about what role they want to play in AI model training.

Creators can choose whether to authorize YouTube to provide their videos to AI labs and other third parties for training new models; they can also refuse authorization. But even if they agree, they receive no compensation. YouTube says that currently around 1 million creators have opted to authorize. Meanwhile, research firm Social Blade estimates that there are about 69 million active creators on the platform, meaning the proportion of those who authorize remains small (YouTube refuses to disclose the platform’s total number of creators). Many creators have posted on Reddit forums that multiple AI companies have already contacted them to negotiate authorization for video materials, with some companies offering as much as $100k per 1,000 hours of video. Hall said she “has considered” authorizing her own video materials and that she will “probably” agree. “As an original content creator, I should receive appropriate compensation.”

The former YouTube executive mentioned earlier said the platform is balancing the opportunities and risks brought by AI, but ultimately anything that can produce more content and extend users’ viewing time will have the advantage. “At the end of the day, as a platform, what YouTube cares about most is watch time,” the former executive said. “What the user is watching—actually the platform doesn’t care that much. At most, it’s secondary.”