He Predicted the Metaverse Would Fail. He Has a Warning about AI.
Five years and $80 billion later, Meta is giving up on the metaverse.
Last week, the company announced it was shutting down Horizon Worlds, its virtual reality social platform, before partially reversing course the next day. Users can still access it, but Meta will stop adding new content and shift its focus to mobile.
Translation: We've sunk enough money into a virtual world nobody showed up to. We're going all in on AI.
Let's start with the most entertaining aspect of this story: Mark Zuckerberg's spectacular miscalculation. When he unveiled Horizon Worlds in 2021 and renamed the company Meta, he predicted over a billion people would one day spend part of every day immersed in virtual worlds. He dedicated 15,000 employees and $83 billion to bring his vision to life. At its peak, Horizon Worlds attracted only 200,000 monthly users. In 2025 alone, Reality Labs — the division behind the project — lost $19.2 billion.
Now, as satisfying as it is to watch Mark with egg on his face, analysts have noted the metaverse failure is likely a blip in the bigger picture. Meta will spend $135 billion on AI in 2026 alone. Its AI-powered Ray-Ban smart glasses tripled in sales last year. The company's overall revenue is projected to grow this year, and Meta is still considered one of the strongest stocks to own.
Nevertheless, when a goliath like Meta missteps so dramatically, it's worth digging a bit deeper to understand exactly why. There are obvious answers: expensive headsets, a market flooded with cheaper entertainment, the AI hype cast VR aside.
But Meta may have also fundamentally misjudged the generation it was building for. Gen Z is the loneliest generation on record — constantly on their phones, distrustful of institutions, and deeply isolated. Yet, paradoxically, they are hungrier for real-world connection than older generations. Young people want something the metaverse was never going to give them.
I had the chance to pressure-test this theory with Ethan Zuckerman, a professor of public policy at UMass Amherst and director of a university initiative working to reimagine the internet as a tool for civic engagement.
Ethan is an icon in tech policy circles. In 1994, at 21, he left grad school to become an early employee at Tripod.com, one of the first personal publishing websites. It was there that he inadvertently invented the pop-up ad while trying to help an advertiser whose banners kept appearing next to images of anal sex. The story has since become a piece of internet founding lore, but a source of genuine remorse for Ethan. He has called the ad-supported business model "the Internet's Original Sin," arguing it was the foundational mistake that corrupted everything that followed.
Tripod sold for $58 million in 1998, and Ethan has spent the past 28 years building a better internet. He founded a nonprofit that sent IT specialists to work in West Africa, and a citizen journalism network now operating in 150 countries. He led the MIT Center for Civic Media for nine years before resigning in protest over the Media Lab's ties to Jeffrey Epstein. And in 2024, he sued Meta to establish that users have the right to modify their own social media feeds — a case that has since been dismissed.
He also predicted the metaverse's failure. Few people are better positioned to make sense of this development than Ethan, who has spent three decades at the intersection of technology and civic life. I sat down with him to ask why the metaverse failed, and what it might portend for AI. He stopped short of drawing a direct causal line between the metaverse's collapse and the surge in Gen Z civic engagement — but suggested that Gen Z's deep suspicion of social media and Mark Zuckerberg personally may have played a role.
The conversation below has been edited for length and clarity.
There are platforms like Roblox and Fortnite that still attract hundreds of millions of users doing something that looks superficially similar to what Meta was attempting. So clearly there's still an appetite for virtual social spaces. Why do you think Horizon Worlds failed so spectacularly? Was it just a bad product, or does it reveal something deeper about what Meta specifically got wrong?
I think we can zoom out and say there's the hard metaverse vision and the easy metaverse vision. The hard metaverse finds its way back to Neal Stephenson's Snow Crash. The vision is immersive — goggles, maybe a suit. An open creative space that serves as a full alternative to flat two-dimensional computer space. That's the version everybody seeks to build because it's the real metaverse.
Then there are lesser versions. Second Life sought to be immersive without requiring goggles, and retained the ability to build and create — but at such a high degree of difficulty that it appealed only to very specialized populations. Never more than hundreds of thousands of users. Then there are Roblox and Fortnite. Really all you have there is three-dimensional social space. I actually think the most successful metaverse anyone has created is shared Minecraft servers. You're giving up photoreal, you're giving up lifelike avatars — but you are creating a game space that people can interact in, and it works.
What Zuckerberg got wrong was he bought the hard metaverse vision hook, line, and sinker. He concluded that because he had good hardware — and I actually think Oculus is very good hardware — he would be able to make this normal and ordinary with millions of users in a short amount of time. And here's the thing: if Meta can't do it, it probably can't be done. They had the ability to put this in front of three billion users. They had hardware that actually works. If they can't get it done, you have to take seriously the idea that this is just a bad idea.
There's been research showing a spike among Gen Z in volunteering and community service — and at the same time, it's one of the most isolated generations on record. Do you think the metaverse's failure is related to this hunger for real-world belonging, or are they just two unrelated phenomena happening in parallel?
I certainly don't want to say they're unrelated. Let me add some anecdotal evidence from our lab. We've been trying to build new social networks. We backed away from some of those because building local community networks is really hard. People would say, why you? Nobody knows who you are. So we ended up working on music instead. A friend of mine built a social network called Freq. We did piles of user testing with college radio folks. As soon as we said "social media," they were out.
This lines up with a lot of other experiences. I teach a class called Fixing Social Media. The students universally believe social media is broken and bad for them. They all conclude: I have a bad, addictive relationship with social media. To a person.
So I think it's very possible that Zuckerberg — who people already resented for turning Instagram into a space for advertisers and influencers — was never going to be the one to create the next great social space. But beyond Zuckerberg personally, it could have been some combination of Gen Z saying, I'm suspicious of anything that wants me to socialize online, and Millennial parents saying, no way am I letting my child do that.
You've called the ad-supported internet "the original sin." Now massive investment is flowing into AI, built by the same companies with the same business models. Are we just repeating the same mistake?
Yes. I break the critique of social media into three categories. First: it's bad for us individually — addictive, bad for body image, performative. Second: it's bad for society — polarizing, it spreads misinformation, atomizes people rather than creating an interactive public. The third, which I find maybe the most persuasive: media can make it very easy for us to be bad to each other.
I want to apply what we've learned to AI. The first thing: the attention economy is very dangerous. If a company like OpenAI continues selling ads in its free product, they immediately have an incentive to capture as much of your time as possible — which creates incentives to be sycophantic, to reinforce you, to become your friend or partner in ways that may be deeply unhealthy. Second: compiling personality profiles. We have no comprehensive privacy legislation. People are interacting with AI in very personal, very intimate ways — revealing things they might not reveal to anyone else. That sets up a clear path toward harm.
Then there's personalization. AIs are amazing echo chamber tools. They can very easily put you in a situation where you go down rabbit holes and never encounter opposing points of view. And AI outputs can be far more persuasive than a social media feed.
There's also a research problem: the ways we know how to study social media depend on that information being mostly public. AI conversations are inherently private. There's no good way for a journalist or scholar to say, give me the last million conversations. So studying these things is going to force us to rely on sanitized platform data, sock puppets, or data donation — all of which have real limitations.
Do you use AI in your own work? How do you think about using it without reproducing the very dynamics you've spent 25 years pushing back against?
I use it a lot, in the spirit of trying to figure out what it's actually good at. Coding agents have recently gotten very good — the ability to say, here's a pile of data, here's what I need, go build it. I'm doing my share of that.
But I have a column coming out this month about a research project where AI told us we'd found something fascinating and wrote a ten-page report on it. When we dug in carefully, it was total nonsense — in a fairly subtle way. The AI had done something poorly, and when we posed that result back to it, it generated bogus data, which led to a bogus analysis.
There's also something called RAG [retrieval augmented generation — constrains AI responses to a specific set of source documents, preventing it from drawing on its broader, and sometimes unreliable, training data]. The best example is Google's NotebookLM. You load it with documents and ask questions, but answers can only come from those documents. When I loaded some of my own research papers into it and asked if a frog is an amphibian, it said: "I'm sorry, these papers are about random sampling strategies for YouTube. There's nothing here that would tell me if a frog is an amphibian." That's exactly what you want.
As for anxieties — they're through the roof. But they're not preventing me from trying to figure out how to use these tools effectively.
If the metaverse's collapse is some sort of verdict — the market rejecting a hollow vision of human connection — is there something you'd want to see over the next decade as a counter-verdict?
I have always felt that XR — extended reality — is much more interesting than VR. VR has always felt dystopian to me: the world is awful, I want nothing to do with it, so I'll fence myself away from it. XR, where you're out in the world and add a data layer on top of it, feels very interesting. A bad data layer is walking down the street and being told that woman is attractive and single, here's her dating profile. A good data layer: that building was built in 1870, it's owned by so-and-so, it's zoned for this. I'm an urbanism geek — that would be amazing.
My wife and I are part of a board game club. Almost every Wednesday night we go to a community center in North Adams and play board games with a whole group of people. When we need a rule clarification, we pull out our phones and check Board Game Geek. But we're not spending a lot of time on them, because board gaming is an analog activity where you're face to face with people. We've worked out a norms framework around how that data layer comes in and out.
That negotiation is the one we're going to have. My money's on XR, not VR. Data layers laid over physical reality, within strenuous norms that we're still figuring out — how to do this in a way that enhances rather than replaces real human connection.
Discover more from Kliger's Corner
Subscribe to get the latest posts sent to your email.