Meta said to build AI version of Zuckerberg for staff interaction

Meta is building an AI version of CEO Mark Zuckerberg that could talk with employees and give feedback, as the company expands its AI work, people familiar with the matter said.

The project is still early, and Zuckerberg is helping train and test a photorealistic 3D character.

The character is trained on images, his voice, mannerisms, and publicly available statements, as well as his recent thinking on company strategy, one person said.

The effort is separate from an internal CEO agent for Zuckerberg and follows Meta’s earlier push into AI characters, which later drew child safety concerns and, since January, has seen the company limit teen access.

🔗 Source: Financial Times

🧠 Food for thought

Implications, context, and why it matters.

An AI Zuckerberg needs more than conversation. Agentic skills matter

Meta ties the effort to Zuckerbergs idea of personal superintelligence, and it has put tens of billions of dollars into that push over the past year 1.
The system would need a set of dependable procedures, called agentic skills, to handle multi step work such as giving employee feedback 2.
Zuckerberg has taken part in training and testing the avatar, which suggests Meta treats it as a hands on strategic bet 1.
Studies find that human curated skills raise AI agent success rates by an average of 16.2 percentage points, while skills generated by the agent itself lower performance on average 2.

The AI CEO clone could widen executive reach while adding security risk

If the experiment works, AI avatars could help executives communicate at scale and could open a new category of enterprise software.
Reusable skills also create attack targets, including attempts to steal credentials or pull out data through exfiltration, which means unauthorized removal of sensitive data 2.
A model trained on company strategy would draw attention from corporate espionage groups because it can hold valuable internal context.
Researchers describe threats such as poisoned skill retrieval, where an attacker tricks an agent into running a harmful routine, which changes the shape of insider risk 2.