Synthesia turns journalists into interactive AI avatars

Artificial intelligence company Synthesia is taking digital avatars beyond scripted videos, allowing users to create virtual versions of themselves that can listen, respond and interact with other people.

The company recently created an interactive digital avatar of a journalist, trained specifically on one of the journalist’s published stories. The avatar was designed to answer questions about the article while refusing to respond to topics outside the material used to train it.

The experiment offers a glimpse into how AI-generated digital twins could be used in journalism, public relations, corporate communication and training.

Synthesia, originally founded in the United Kingdom and now operating internationally, specialises in AI-generated video and digital avatars. The company reached a $4bn valuation earlier this year and reported more than $100m in annual recurring revenue last year.

Its technology is already used by enterprises to produce training and communication videos featuring AI avatars. The company has also developed Roleplay Sessions, which allows employees to practise scenarios such as sales presentations with an interactive AI avatar that responds to their answers and evaluates their performance.

Creating the journalist’s digital twin required a short recording session at Synthesia’s New York office. The process involved taking multiple photographs and recording about two minutes of the journalist’s voice.

The company then produced several versions of the avatar, including scripted versions that simply deliver supplied text and interactive versions capable of listening and responding.

The interactive avatar was trained using a specific published story. Synthesia built the system using several AI components, including voice-to-text, language-processing, text-to-voice and video-generation models.

When a user speaks to the avatar, voice-to-text technology converts the question into text. A language model processes the request and determines an appropriate response, which is converted back into speech before Synthesia’s video technology animates the digital figure.

Synthesia also allows enterprise customers to combine its technology with models from other AI companies, including Cartesia, ElevenLabs, Google and OpenAI. Customers can either host the systems on their own cloud infrastructure or use Synthesia’s hosting.

The company currently operates across three broad product areas: conventional AI-generated videos, interactive avatar experiences through its Sessions platform, and an API that allows developers and businesses to integrate Synthesia’s video and voice technology into other products.

The journalist’s avatar was deliberately limited in scope. When asked questions unrelated to the article it had been trained on, it redirected the conversation rather than generating an answer.

That limitation illustrates one of the differences between a controlled digital twin and a general-purpose AI chatbot. A deterministic avatar can be restricted to approved information, while a less constrained system could generate responses beyond the material originally supplied by its creator.

The technology has also raised questions about how digital avatars could change journalism.

AI presenters could potentially be used to explain stories, answer audience questions or provide information when a journalist is unavailable. In corporate settings, digital versions of executives, trainers or communications staff could respond to routine questions without requiring the real person to be present.

But the use of AI avatars in journalism also raises questions about authenticity and trust.

Journalism relies heavily on the relationship between reporters and their audiences, particularly when journalists are reporting sensitive or contested issues. An avatar that looks and sounds like a journalist could make it difficult for audiences to distinguish between content personally delivered by the journalist and content generated automatically.

The technology also creates new possibilities for impersonation and misuse if digital likenesses and voices are reproduced without proper consent.

Synthesia requires users to consent to the creation of their avatars. In the experiment, the journalist was required to provide consent before the company captured the images and voice recording needed to create the digital twin.

For the journalist involved, the experience offered both fascination and discomfort. Friends described the avatars as interesting but somewhat unsettling, while family members tested the interactive version by asking questions about personal experiences that had not been included in its training material.

The avatar repeatedly redirected them to the article instead.

Synthesia’s experiment points to a broader shift in how people may interact with AI-generated versions of real individuals. Rather than simply watching an avatar read a prepared script, users could eventually hold conversations with digital representations of journalists, executives, educators and other professionals.

Whether audiences will accept such interactions, particularly in news, remains uncertain.

For now, the technology is being positioned largely as a tool for corporate communication, training and interactive experiences. But as digital twins become more sophisticated, questions around consent, disclosure, accountability and trust are likely to become increasingly important.

The technology may allow people to be virtually present in more places at once. The harder question may be whether audiences will trust the digital version as much as the human being it represents.