clock menu more-arrow no yes mobile

Filed under:

Mark Zuckerberg’s not-so-secret plan to join the AI race

Why Meta is cautiously following the hype.

A photo of Mark Zuckerberg on screens. Michael Nagle/Bloomberg via Getty Images
Shirin Ghaffary is a senior Vox correspondent covering the social media industry. Previously, Ghaffary worked at BuzzFeed News, the San Francisco Chronicle, and TechCrunch.

At a high-profile White House summit on AI last Thursday, CEOs from some of the most important technology companies in the world sat together in a rare joint meeting. Vice President Kamala Harris and several senior administration officials hosted the event. President Joe Biden even dropped by. It was a critical moment for the future of AI technology, but one company was notably left out: Meta.

While OpenAI, Google, and Microsoft were all present, the owner of Facebook, Instagram, and WhatsApp was reportedly not invited to the meeting. An unnamed White House official told CNN that Meta was left out because the meeting “was focused on companies currently leading in the space,” particularly “on the consumer facing product side.”

It’s true that Meta doesn’t have a hit AI-powered app like OpenAI’s ChatGPT, which has surpassed records for being one of the fastest-growing consumer app products of all time. And it also doesn’t have an imminent splashy product announcement like Google, which is expected to reveal plans to put AI into its core search engine later this week.

But the White House snub belies the reality that Meta has been building powerful AI tools of its own, without getting as much attention for it.

Zuckerberg is starting to speak up about his company’s AI plans, though. During Meta’s strong April earnings call a week before the White House summit, Zuckerberg mentioned “AI” at least 27 times in the span of an hour, sharing some of the most detailed plans yet about how to integrate the technology more deeply into Facebook, WhatsApp, Instagram, and other Meta apps. The CEO said generative AI is “literally going to touch every single one” of Meta’s products, and made a case that his company was “no longer behind” on some areas of AI research.

While this sort of thing could be seen as Zuckerberg buying into the AI hype, it also seems like he’s drawing attention to his company’s long history of investing in the technology. For years, Meta has employed a world-class AI research team that, as recently as last month, has been publishing industry-changing research, and is taking a notably different, open-source approach that could give it an advantage over its peers. Recently, the company has been pushing to commercialize that research by assembling a new top-level product team focused on getting generative AI into Meta’s core apps.

Even though we can’t see it, Meta has, for years, used AI to recommend posts in our feeds, moderate content, and target ads behind the scenes in Instagram and Facebook. Now, Zuckerberg is trying to incorporate AI more visibly into his company’s products. He recently announced plans to put AI chatbots that can act as customer service representatives into WhatsApp and Messenger. And while Meta hasn’t announced specific plans yet, it’s easy to see a world in which AI could write a new Facebook post based on your interests, or create an AI avatar for you to post to Instagram. Indeed, some of Meta’s competitors are already experimenting with the technology. LinkedIn is putting AI-generated articles in people’s feeds and TikTok is suggesting AI-generated profile pictures.

The implications of Meta integrating AI more deeply into its core products could fundamentally change the way people make and consume content on the world’s largest social media apps, bringing us closer to what some have called a “semiautomated social network” — in which AI can create its own posts or even create entire accounts.

“Our single largest investment is in advancing AI and building it into every one of our products,” Zuckerberg wrote in a blog post in March. “We have the infrastructure to do this at unprecedented scale and I think the experiences it enables will be amazing.”

But Meta faces some serious challenges in becoming a more visible leader on AI. Given the company’s years of brand baggage around the spread of harmful misinformation, polarizing political content, and data leaks, Meta needs to be especially careful in figuring out how to put AI into its products without spooking users. Complicating matters further, the company was also slow in securing the right kind of chips needed to scale AI projects, leaving it playing catch-up to its competitors on a key piece of hardware.

Meanwhile, Zuckerberg’s very public pivot toward the metaverse a year and a half ago makes it hard for him to refocus on AI without being accused of “ditching” his original metaverse vision, though he argues that the two concepts are related. An even bigger problem: If Meta’s AI-powered tools generate biased, incorrect, or emotionally loaded content — as previous Google’s Bard and Microsoft’s BingGPT have already done — the company will surely attract even more controversy.

Meta’s next move in the AI race could impact how its billions of users communicate and interact with its flagship products, like Facebook and Instagram. It will also determine whether or not the company maintains its grasp on the social media landscape for the next decade or falls behind competitors, like TikTok.

Meta’s long but boring history with AI

You may not realize it, but AI plays a crucial role in what posts, videos, pictures, or even ads you’re seeing when you log in to Facebook or Instagram.

In earlier days, Facebook and Instagram mainly showed people content from accounts they follow. But now, around 40 percent of all content people see on Instagram is recommended by AI, and 20 percent on Instagram and Facebook combined. Those numbers, which Zuckerberg cited in Meta’s latest earning call, have been growing in recent years as the company has changed its strategy and started showing people suggested content — especially short-form videos, like Reels — outside of their direct social network, like TikTok does.

Meta and Google both have a “PR deficit,” said Jim Lecinski, a professor of marketing at Northwestern University Kellogg School of Management, who was a VP at Google before leaving the company in 2018.

“While they’ve been using AI behind the scenes, they’ve not been talking about it,” Lecinski said, “and perhaps, then, not being given credit for having pioneered, in many respects, the use of AI and machine learning specifically.”

On the business side, Meta also uses AI to help run its main line of business: selling ads. Meta says it’s been improving its AI tools that let advertisers target ads. Because of better AI, the company says it’s improved its “monetization efficiency,” or how much the company makes off of ads they sell on Reels, by 30 percent on Instagram and over 40 percent on Facebook quarter over quarter.

Still, while Facebook and Instagram heavily use AI, Meta’s mega apps don’t allow people to create new content, like text or images, the way that ChatGPT and other viral generative AI apps do. So why hasn’t Meta created its own splashy generative AI app and pushed it to its billions of users yet?

A person takes a selfie in front of a Meta sign.
Meta’s headquarters in Menlo Park, California.
Liu Guanguan/China News Service via Getty Images

The potential reputation risk for Meta is a big factor. More so than companies like Google or Microsoft, Meta has long been criticized for being a platform where hate speech and misinformation can run rampant. It’s not surprising that the company doesn’t want to hastily create a tool that could spread even more toxic or incorrect content.

Meta already tried that, actually. Last November, the company released an experimental AI tool, called Galactica, trained on scientific papers that could summarize academic knowledge. But users quickly found that, with little prompting, Galactica could spit out factually incorrect and racist information. Within three days, Meta took the tool down.

“It was a bit of a train wreck, overhyped and easily co-opted into generating misinformation,” said Gary Marcus, an entrepreneur and emeritus professor of psychology and neural science at NYU.

But Galactica is also just one project and likely a lesson learned.

Another factor that’s holding Meta up has to do with the hardware necessary to run powerful AI systems. An internal memo posted by Meta’s new head of infrastructure in September said that the company has “a significant gap in our tooling, workflows, and processes when it comes to developing for AI” and needs “to invest heavily here,” according to a recent Reuters report. Zuckerberg said on Meta’s April earnings call, however, that the company is “no longer behind” in building out its AI infrastructure.

Meta’s open source AI research strategy

One of Meta’s biggest strengths in AI is the caliber of its research department, which many experts say is competitive with industry peers like Google and OpenAI. Meta has been staffing up its AI research team since 2013, including by hiring NYU professor Yann LeCun to lead the department as its chief AI scientist.

A lot of the work that Meta does to integrate AI into its social media apps is secretive. No one outside the company knows exactly what the recommendation algorithms powering your Facebook feed look like, for instance. But its AI research department is relatively open.

The department, called FAIR, regularly publishes publicly available AI research papers, while other AI companies, like Google and OpenAI, are becoming less open due to concerns about competition. Zuckerberg has said that by doing this, Meta can take the lead in setting the industry standard for how AI products are developed, and let outside developers better integrate with Meta’s ecosystem.

In February, Meta’s research department shared a large-language model, called LLaMA, with researchers. This technology let academics create their own AI chatbots and code-generating tools with fewer computational resources than other models. Researchers like Eric Wallace, a computer science PhD student at UC Berkeley, called it “unprecedented.”

“This was a very exciting release,” said Wallace, who used LLaMA to program a chatbot as part of a research project with other academics. “There were so many questions that we wanted to tackle even years ago that we can maybe start to study because this model is out.”

LLaMA’s capabilities are currently behind OpenAI’s latest GPT4 model as well as Google’s Bard, but the model still represents a major step forward for the AI research community. That’s in large part due to researchers’ ability to refine the underlying code to their liking, Wallace said. It’s also a lot easier to operate LLaMA without beefy computers that other models require to run; the model is so lightweight that people can even load it on their phones. A major caveat is that Meta has made LLaMA available only for the purposes of research and not for commercial use. A version of LLaMA was leaked online earlier this year, however, making it more public than Meta originally intended.

Even some employees at Google think Meta’s strategy to make its AI research more open makes sense.

“The uncomfortable truth is, we aren’t positioned to win this arms race and neither is OpenAI,” one Google engineer wrote in an internal memo that was made public in May. “While we’ve been squabbling, a third faction has been quietly eating our lunch. I’m talking, of course, about open source.”

It’s still early to say if Meta’s open source strategy will beat OpenAI and Google, but clearly it’s making an impact.

Zuckerberg wants it all

Lately, many people have noticed how Zuckerberg seems to be saying “AI” a lot more than “metaverse.” It’s only been 18 months since Zuckerberg changed the name of his company from Facebook to Meta and made the metaverse — a series of virtual worlds where people can socialize, work, and play — the north star of the company. But the tech CEO wants people to know that he’s committed to both ideas.

“A narrative has developed that we’re somehow moving away from focusing on the metaverse vision, so I just want to say up front that that’s not accurate,” Zuckerberg said on April’s earnings call. “We’ve been focusing on AI and the metaverse, and we will continue to.”

A Meta employee demoing the Meta Quest Pro VR headset.
Robyn Beck/AFP via Getty Images

Zuckerberg envisions a world in which the metaverse and AI are intertwined. He said that the two concepts are “related,” and that AI can be used for “understanding the physical world and blending it with digital objects” and “being able to procedurally generate worlds.”

So what does that all mean?

Right now, it takes computing power, graphics, and coding knowledge to design a new avatar or virtual world in the metaverse. In the future, Meta hopes, this can get easier thanks to generative AI. The company has demoed early tools that let people create virtual environments by describing what they want to see. You might simply say, “I want to see a palm tree over there,” and poof, a 3D tree appears in your living room. The company has also been building hyper-realistic avatars that could make AI chatbots look and feel more like real people.

The problem is, not enough people want to be in the metaverse right now. AR/VR headsets aren’t nearly as widely used as cell phones or computers right now. Reports indicate even those who do use Meta’s Quest headsets don’t use them for very long. But by Zuckerberg’s own estimation, it could take a decade to build the technologies that will enable the metaverse’s full potential.

Meanwhile, AI is already reshaping our present reality. Everyone from Joe Biden to high schoolers have used ChatGPT, while AI-generated Drake songs and fake Pope images have captured the world’s attention.

That’s why Zuckerberg wants generative AI in Meta’s core apps more quickly, and saying that the technology is integral to Meta’s future. At the same time, if the new top-level product team working on integrating generative AI into Instagram and Facebook moves too quickly and has a misstep around misinformation or privacy, the whole strategy could backfire.

That may be why one of the most tangible and detailed examples Zuckerberg has given around using generative AI in its apps in the near-term is relatively basic. The new product simply lets businesses employ AI chatbots to respond to customer service questions.

“Once you light up the ability for tens of millions of small businesses to have AI agents acting on their behalf, you’ll have way more businesses that can afford to have someone engaging in chat with customers,” said Zuckerberg on April’s earnings call.

Giving small businesses easy customer support is a pretty measured and banal start to Meta’s AI-powered future. And it’s easy to see the appeal for Meta. It’s relatively uncontroversial, and it could make some money.

But what will likely determine the success of Meta’s AI future is whether it can build a hit consumer AI tool like ChatGPT. And while Meta might seem far from that right now, it’s also still far too soon to count it out of the AI revolution.