Meta is training a photorealistic, AI-powered 3D version of chief executive Mark Zuckerberg to field questions from employees on his behalf, the Financial Times reported Monday, citing four people familiar with the project. The avatar is being built on Zuckerberg's mannerisms, voice, tone, publicly available statements, and his recent thinking on Meta's strategy, with the CEO himself personally training and testing it inside the company's newly formed Superintelligence Labs. The effort, which Meta hopes to eventually extend to creators through its existing AI Studio product, arrives as the $1.6 trillion social media company pushes employees toward AI adoption and plans to spend up to $135 billion on AI infrastructure this year.

Key Takeaways

AI-generated summary, reviewed by an editor. More on our AI guidelines.

What the bot is being trained on

The character is being taught how Zuckerberg speaks, how he carries himself, and, according to the FT, what he currently believes about where the company is going. That last piece matters. Trained only on speeches and earnings calls, an avatar recites. Fed recent strategic thinking, it starts to manage.

Meta's Superintelligence Labs, which was stood up after the Llama 4 benchmark mess, has been building a catalog of photorealistic 3D characters for real-time interaction. The Zuckerberg version is the one the team has been told to prioritize. Scaling the technology is hard. Real-time photorealism eats compute, and latency kills the illusion the second a reply feels late. Meta acquired voice-AI startups PlayAI and WaveForms last year to shore up the speech side of the stack.

The project is early. Nobody has publicly seen it work.

Two Zuckerbergs, different jobs

This is not the AI Mark Zuckerberg you may have already heard about. A separate effort, first reported by the Wall Street Journal in March, involves what Meta calls a "CEO agent." That one is a personal assistant designed to pull up information on demand, summarize activity across the company's roughly 78,000 employees, and cut out the layers of management a question usually passes through before landing in front of the real CEO.

One Zuckerberg talks to staff. The other lets the staff layer thin out.

Zuckerberg is spending between five and ten hours a week writing code for Meta's AI projects and dropping into engineering reviews, the FT says. He is, in other words, acting less like a chief executive and more like a tech lead, and asking his engineering organization to build a simulation of him that can cover the hours he is not in the room.

Employees are nervous

Inside Meta, product managers have been asked to complete an "AI skills baseline exercise" that includes technical system design and something the company calls "vibe coding." Participation is framed as optional and diagnostic. Staff are reading it as a proficiency test. Some worry it is a shortlist for the next round of cuts.

Those fears are not abstract. Meta is preparing workforce reductions of as much as 20% in the coming weeks, according to the Wall Street Journal's earlier reporting, with the company openly tying the cuts to AI infrastructure spending. An avatar that can answer employee questions while a product manager waits on a skills review does not feel neutral. It feels like a rehearsal.

The Muse Spark overhang

All of this is landing on top of a miss. On April 8, Meta Superintelligence Labs released Muse Spark, its first model and the first test of the lab's hundreds-of-millions-of-dollars hiring spree. A Meta executive told Bloomberg the model would not match OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini in several areas. Meta's own release called it "an early data point on our trajectory."

Yann LeCun, the former Meta AI chief who left the company amid the Llama 4 drama, told the FT in January that after the benchmark results "were fudged a little bit," Zuckerberg "was really upset and basically lost confidence in everyone." That is the boss now being modeled, voice and tone and strategic instincts, into a character designed to hand out feedback.

A crowded mirror

Meta is not the only company doing this. Uber's Dara Khosrowshahi told a podcaster earlier this year that his own staff had quietly spun up an AI clone of him without asking. Last week, YouTube shipped an avatar tool of its own, letting any user mint a lookalike that speaks and moves in their voice. And AI Studio, Meta's own consumer product, already lets Instagram creators bolt a digital version of themselves onto their DMs.

The difference is the direction. You do not choose to speak to the AI Zuckerberg. The AI Zuckerberg is what shows up when you ask to speak to Zuckerberg.

History matters here. Remember the 2022 Horizon Worlds screenshot? Cartoon Zuckerberg. Dead eyes. Cartoon Eiffel Tower. That image still gets dragged out of the archive every time Meta ships anything avatar-shaped, and for good reason. The man once described himself, in his own words, as "like the most awkward person." Now Meta wants its engineers to compress that same awkward person into a 3D rig, a voice stack, a strategic memory, and a feedback button on an internal tool.

The real Zuckerberg will keep coding. The simulated one will take the meetings.

Frequently Asked Questions

What did the Financial Times actually report?

The FT reported Monday that Meta is training a photorealistic, AI-powered 3D version of CEO Mark Zuckerberg to engage with employees on his behalf. Four people familiar with the project told the newspaper that the avatar is being built on Zuckerberg's mannerisms, voice, tone, public statements, and his recent thinking on company strategy. The project is in early stages and runs inside Meta's Superintelligence Labs.

Is this the same as the 'CEO agent' project?

No. The Wall Street Journal reported in March that Zuckerberg is also building a separate 'CEO agent' designed to help him retrieve information and cut through management layers inside Meta's roughly 78,000-person workforce. The Zuckerberg avatar is a different effort: a photorealistic 3D character meant to talk to employees, not an operational tool for the CEO himself.

Why are Meta employees nervous about it?

Meta is preparing workforce reductions of as much as 20%, partly to offset AI infrastructure costs. Product managers have been asked to complete an 'AI skills baseline exercise' that includes a technical system design test and a 'vibe coding' session. Staff describe the exercise as diagnostic on paper and a proficiency test in practice, landing right as the company builds a simulation of its CEO.

How much is Meta spending on AI?

Meta has told investors it plans to spend up to $135 billion this year on AI infrastructure, part of what Zuckerberg calls a push toward 'personal superintelligence.' The company has also hired aggressively for its Superintelligence Labs unit and acquired two voice-AI startups, PlayAI in July 2025 and WaveForms in August 2025, to support more realistic speech synthesis.

Has Meta's new AI lab shipped anything yet?

Yes, but the results are mixed. On April 8, Meta Superintelligence Labs released Muse Spark, its first model. A Meta executive told Bloomberg the model could not match OpenAI's ChatGPT, Anthropic's Claude, or Google's Gemini in some areas, and Meta's own announcement described it as an early data point. The Zuckerberg avatar is being built by the same unit.

AI-generated summary, reviewed by an editor. More on our AI guidelines.

Arm Abandoned Neutrality. Its Biggest Customers Cheered.
The Implicator
Moltbot Punched Through Every Security Wall. Attackers Followed.
The Implicator
Anthropic Chases OpenAI Into Healthcare. The Graveyard Is Crowded.
The Implicator
AI News

San Francisco

Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: [email protected]