Synthesia adds real emotions to AI video avatars

If you have ever had to sit through a corporate compliance video featuring an AI avatar, you know the pain. The body stays completely still, the eyes stare blankly into the camera, and the mouth flaps around like a poorly dubbed cartoon.

AI video unicorn Synthesia wants to fix that. On Thursday, the company rolled out what it calls Expressive Avatars, pitching them as the world's first video avatars completely generated by AI models rather than stitched together from pre-recorded footage.

Moving past the Frankenstein method

Traditional avatar generation has mostly relied on a digital Frankenstein approach. Studios record a human actor performing dozens of phonemes and expressions, and the software cuts and pastes those snippets together to match a script. The result is almost always robotic.

Synthesia claims its new avatars are generated dynamically on the fly using large pretrained models. Instead of manually stitching video assets, the neural network learns multimodal distributions of real facial movements to recreate how humans naturally emote.

Company CEO Victor Riparbelli admits the technology is not quite 100% there yet, but claims it will be mind-blowing by the end of the year. He noted that deep learning networks are uniquely suited for capturing the subtle twitches of facial muscles that humans cannot easily describe. The next big technical hurdle on his roadmap is hands, which he bluntly called super hard.

Double down on the B2B playbook

Unlike OpenAI, which chases massive consumer buzz alongside enterprise deals, Synthesia is playing a focused game similar to Perplexity. It only cares about corporate use cases like employee onboarding, training, and sales demos.

That enterprise focus has already paid off. Ahead of this release, 200,000 creators used Synthesia's 225 legacy avatars to produce over 18 million videos across roughly 130 languages. Big-ticket enterprise clients include Zoom, the BBC, and DuPont.

Staying inside corporate walls also helps the company manage deepfake risks. Synthesia's avatars have previously been exploited to produce propaganda in Venezuela and fake news stories on social media. In response, the startup recently tightened its content rules, expanded its safety team, and started testing C2PA content credentials to cryptographically tag generated clips.

Synthesia hit a $1 billion valuation back in June 2023 after pocketing a $90 million round. Riparbelli says there are no immediate fundraising plans, though generating high-fidelity video on the fly is notoriously compute-heavy.

Why it matters

We think Synthesia's decision to narrow its scope is a very smart survival strategy in a market threatened by generalist giants like Sora. Instead of trying to generate fantasy worlds from scratch, it is perfecting a boring but highly lucrative workflow: replacing expensive corporate video shoots.

For enterprise teams, adding believable emotion and accurate lip-sync across 130 languages means training and marketing content might finally hold someone's attention without costing tens of thousands of dollars in studio fees.

However, teams should approach the launch with measured expectations. While facial micro-expressions are improving, the lack of natural hand and body movement still leaves these avatars flirting with the uncanny valley. A head that smiles naturally while sitting on a rigid mannequin body will still distract viewers.

There is also the inescapable security liability. The more persuasive these avatars become, the easier it is for rogue actors to impersonate executives or spread internal disinformation. Adopting C2PA credentials is a solid technical step, but enterprise risk teams will need strict verification protocols before giving synthetic presenters access to sensitive communications.