Victor Riparbelli

Deepfakes as a service: that’s Synthesia’s business model. But it’s not what it sounds like. The London-based AI company offers businesses a platform to turn any document or webpage into an engaging video hosted by a realistic AI avatar (resembling either you the user, or an actor whose likeness Synthesia has paid to license) that can change its expressions, body language, and intonations based on the text it’s reading. In June, the company launched a service allowing customers to create an avatar of themselves using a laptop webcam in as little as 5 minutes. Victor Riparbelli, Synthesia’s CEO, explains that most humans learn more easily from watching a video than they do from reading. AI-generated video is glitchy and not ready for prime-time, he says, but Synthesia is targeting use cases where viewers don’t tend to care. “Our product is not made for making the next Super Bowl ad,” Riparbelli says. “It’s much more made for [clients who say], ‘We have a help-center article, make a quick video that explains the content.’”

Synthesia is also battling against the seedier ways that people are using AI-generated video. Political deepfakes, showing election candidates saying things they never said, have surfaced in elections in India and Slovakia so far this year, although with fewer adverse effects than some had anticipated. More devastatingly, the technology has been used to produce an epidemic of sexualized harassment of women and girls, their likenesses superimposed onto explicit videos without their consent. Synthesia lobbied the U.K. government to ban the creation and sharing of sexually explicit deepfakes–a law that came into effect in April. On its own platform, Synthesia requires users to upload a verification video of themselves reading a script, to ensure that individuals can’t make AI avatars of other people. “I would hope most people agree that non-consensual deepfakes, nobody wants that,” says Riparbelli. “And I think it’s very positive that we’re making policies against that.”

Buy your copy of the TIME100 AI issue here