Synthesia builds a journalist avatar that answers questions

A double with a narrow brief

Synthesia first sent Davis a link to an avatar of its head of corporate communications, Aleksandra Voika. It answered common press questions about the company. Davis found that more striking than AI-written PR proposals she had recently been asked about.

The company later invited her to its new New York office, where she agreed to make an avatar. It was Synthesia’s first avatar for a journalist and its first for anyone other than Voika.

The interactive version was trained on Davis’s article about why venture-backed startups are more likely to commit fraud than companies without that backing. Its boundary was clear: it answered questions about that article, not about Davis’s life. Asked what she did before TechCrunch or where she lived in New York, it returned to the piece.

Synthesia made the avatar from a photograph and two minutes of recorded speech. The company also created a personal avatar that can read any supplied text, with versions wearing glasses and without them. The interactive avatar came in those two versions as well.

The interactive system combines several components:

Synthesia uses its own video and voice models, while allowing customers to choose other providers, including Cartesia, ElevenLabs, Google or OpenAI. Customers can host avatars in their own chosen cloud or pay Synthesia to host them.

The company sells three kinds of products: a platform for making and distributing scripted avatar videos; Sessions, where people interact with AI agents in surveys and role-play exercises; and an API that combines Synthesia’s video and voice models with other technology.

The limits are part of the demo

The team took several days to make Davis’s avatars. Testing the personal version, she gave it an ordinary script about autumn arriving in New York, her favourite season. The synthetic voice sounded close enough to hers, and without the hoarseness she had had while recording the sample.

The interactive version unsettled some friends, who said its voice was less like Davis’s and its resemblance weaker than the personal avatar’s. Still, it looked enough like her to feel strange. Her mother called it “amazing”—high praise from her—and tried, with Davis’s father, to ask questions about Davis that they thought only they would know. The avatar kept redirecting them to the article. Afterwards, her mother joked that she could not remember giving birth to two children.

That refusal to wander is a product feature, but it also defines the experience. Davis’s avatar was deterministic: it could say only what it had been taught to answer. A less constrained version, perhaps powered by a chatbot allowed to reason freely, could make an interaction more compelling. It could also make it easier to mistake a simulation for a person.

I think the experiment is most revealing where it falls short. A narrow avatar can demonstrate a use case without pretending to replace its subject. But the same boundary that kept Davis’s parents from getting personal answers also kept the system from feeling like a genuine conversation.

What a journalist can delegate

Davis’s test raises a question beyond whether people will watch news presented by avatars: would they trust an avatar in place of a journalist? One investor told her no. Her own view is that journalism depends on trust, and that this part of the work cannot be handed to AI.

There may be a less ambitious use for digital doubles. Someone returning from a holiday or celebration could leave an avatar available to answer work questions. That is a practical promise, though the announcement leaves open how widely companies in the United States will adopt it.

Davis told an investor she expected Generation Z to be unlikely to get used to digital doubles. They still feel like science fiction made real. She finds avatars less unsettling than humanoid robots, in part because an awkward interaction can be ended by logging off.

But the tension is not simply between a real person and a convincing copy. It is between an avatar that stays safely within its script and one that feels alive enough to invite trust. Synthesia’s current system chooses the first side. Its next challenge is whether customers will want anything else.