Illustration: Allie Carl/Axios
"Hi, mom. Hi, dad." The face in the video was mine, and the voice was mine, too. But I hadn't spoken the words my parents heard.
What's happening: The video was the product of a company named HeyGen, which allows anyone to create a personal "deepfake" — an AI-generated video double capable of reciting virtually anything you type into a text field.
Why it matters: Experts have been sounding alarms about deepfakes for years, but for better — and worse — the capability is here now and it requires no specialized skill.
How it works: To create a personalized avatar, customers have to send HeyGen a two-minute video of themselves speaking into a camera (your smartphone is fine) — along with another video giving consent for the company to do its thing.
- HeyGen returns a digital avatar that you can use to generate videos by typing the words you want to speak into a text box. A content filter blocks explicit or violent content.
- In addition to creating avatars of customer's own images, HeyGen offers a range of ready-to-use, generic-faced avatars and voices (they come in various genders and races).
Here's what the result looks like:
Zoom out: There is of course plenty to worry about with such a tool, from scams and fraud to political misinformation.
- But HeyGen is attempting to harness the utility of "good" deepfakes as a quicker, cheaper and easier alternative to recording everything from customized marketing to instructional videos.
- "We want to build a generative video engine to replace cameras, and make it work for everyone to freely create content," said Joshua Xu, HeyGen's CEO, who spent six years at Snapchat before launching his startup at the end of 2020.
The good: HeyGen's recreation of me was convincing — or convincing and creepy, according to the friends and family I shared it with.
- In making the video to accompany this review, I was able to generate a short segment using my avatar in less than five minutes, all without having to put on makeup, grab a tripod or fuss with lighting.
The bad: HeyGen isn't always good at properly pronouncing names. For example, I had to change my name to "Eena Freed" to get my avatar to say it correctly.
- The lip movements are good, but not flawless, and it took a couple tries to get the video just right. One time, for example, my gaze wasn't right at the camera and the resulting video was less convincing.
Between the lines: While HeyGen requires video consent for its avatars, the technology has clearly evolved to the point where deepfakes won't require much expertise.
- My general assumption is that anything that can be done using AI today with safeguards will very quickly become available without such protections.
And even HeyGen's approach raises some questions.
- A colleague, for example, suggested their tool could make it easier to create work-related videos using AI, rather than having to film each one fresh.
- That might be a convenient time-saver — but who would own the rights to my avatar? I definitely don't want to live in a world where employers can create videos of employees saying whatever the company wants them to say.
Practicalities: HeyGen offers a range of subscription options from around $50 to $150 per month that offer a set amount of credits, each good for one minute of video. The higher-end packages offer the ability to create longer and higher resolution videos.
- The average cost is around $3 per minute, while setting up a personal avatar currently costs a flat $199. (A simpler avatar based on just a photo is free, while there's also a $1,000 pro avatar option that requires the use of a studio and green screen.)
- A free trial offers one credit per day, but only with HeyGen's standard avatars or the talking photo.
Our thought bubble: HeyGen's tool is impressive but probably doesn't make sense for casual messages to friends, particularly with the price tag. (Why not just record a selfie video?)
- It could be useful for other purposes, though, such as personalizing professional sales messages without recording something fresh for each customer.
- Such messages might be received as intriguing novelties if they're labeled as AI-generated (which they probably should be). Once the novelty wears off, though, they might end up feeling more spammy than authentic.