Aiodic

Back to the table

102 Creative tool Hg HeyGen Creative Tools

HeyGen

AI avatar video.

What it is

HeyGen is a website that makes videos of people talking without anyone standing in front of a camera. You type a script, pick a presenter, and a few minutes later you get a video of that presenter saying your words, with lips, face and head moving in time with the speech. The presenter can be one of the stock "avatars" the company provides, which are digital versions of paid actors, or it can be a digital copy of you, built from a short recording of yourself. The same tool can take a video you already have, translate what you say into another language, and redo your lip movements so that you appear to be speaking that language in your own voice.

An avatar, in this sense, is a digital stand-in for a person: a moving, speaking image that the software can make say anything. Lip-sync means matching mouth movements to the words being spoken. HeyGen's selling point is doing both convincingly and cheaply enough that a small business, a teacher or a self-published author can use it.

The main things it does, in plain terms (the names of features change often, so treat these as descriptions rather than menu labels):

Under the hood it is deep learning: systems trained on a very large amount of footage of people talking, so they have learned how faces move with each sound. The avatar doesn't replay a recording. It generates new frames for every script, which is why it can say words the real person never said. That is the whole point of the product, and also the whole problem with it.

Where it came from

HeyGen was founded in Shenzhen, China, in 2020 by Joshua Xu, its chief executive, and Wayne Liang, under the name Surreal. According to the South China Morning Post, the company moved to Los Angeles in 2022 and renamed itself Movio, then rebranded as HeyGen in April 2023. The same paper reported in June 2024 that, after scrutiny from American lawmakers over its early links to China, HeyGen had closed its Shenzhen entity at the end of 2023 and asked its Chinese investors, including HongShan (formerly Sequoia China), IDG Capital, Baidu Ventures and ZhenFund, to sell their shares to American investors. The paper said the move also kept the company clear of US limits on buying the advanced chips used to train AI. Today HeyGen describes itself as a Los Angeles company, and the Post reported that its software is not offered in mainland China.

In June 2024 it raised 60 million dollars in a round led by the venture capital firm Benchmark, which Bloomberg reported valued the company at 500 million dollars. At that point HeyGen said it had more than 40,000 paying business customers and had been profitable since the second quarter of 2023. These are the company's own figures. In a paid press release in June 2026, HeyGen said its annual recurring revenue had passed 200 million dollars and that it had over 30 million users; again, that is HeyGen's claim, not an independent audit.

Why it matters

It takes the camera out of making a video. Filming a talking-head video, even a short one, means lighting, a microphone, a tidy room, several takes and some editing. For anyone who is camera-shy, short of time, or making dozens of training clips, that is a real barrier. HeyGen and its competitors let a person record themselves once and then produce new videos by editing a script, the way you would edit a document. For a small business that means product explainers and onboarding videos without a studio. For an author it can mean a short welcome video on a book page that can be updated each time a new book comes out.

It makes one voice speak many languages. Subtitles work, but many viewers skip foreign-language videos, and traditional dubbing with a different actor's voice feels disconnected. Video translation keeps the speaker's own face and a copy of their voice. A course creator who speaks only English can offer lessons in Spanish or Hindi. This builds on AI translation, with the same caveat: machine translation can be wrong, and a fluent-looking video makes a mistranslation more convincing, not less.

It is a deepfake tool with a business model. A deepfake is a video or audio clip that makes a real person appear to say or do something they didn't. HeyGen's core ability is exactly that, done with permission. The company's rules require consent, but rules only bind people who follow them, and journalists have documented HeyGen output being used to impersonate people who never agreed (see the examples below). Everyone who watches online video now has to accept that a person talking to camera may not be a recording of that person at all.

It is competitive and moving fast. HeyGen is one of several companies selling AI presenters. Two of the best-known alternatives are Synthesia, founded in London in 2017 and focused on corporate training videos, and D-ID, founded in Israel in the same year and known for animating photos; each has its own pricing, consent rules and strengths, and none is simply "the best". General video generators such as Sora, Veo and Kling can also now make people talk, which blurs the line further.

Real-world examples

Video translation goes viral, and a pop star speaks Mandarin (2023)

On 7 September 2023 HeyGen announced on X that its video translation feature was live. Its demonstration showed the tech YouTuber Marques Brownlee speaking Spanish and Apple's Tim Cook speaking with an Indian voice, each with lips re-synced. The tech news site The Decoder tested it within days and was impressed by the lip-sync, while noting that clips had to be between 30 seconds and 5 minutes long, only one face could be on screen, and some voices sounded slightly robotic. Clips made with the tool spread quickly on social media.

Within weeks the obvious misuse appeared. In October 2023, clips of Taylor Swift apparently speaking fluent Mandarin on a talk show went viral on Chinese social media; The China Project reported that one had passed six million views and that they were made with HeyGen's translation tool. They were built from real interview footage by unknown users, and there is no public sign that Swift was involved. Viewers were split between amazement and alarm, with some pointing out how easily the same trick could serve scammers. It showed both sides of the product in one clip: impressive translation, applied to a real person who hadn't asked for it.

Milei at Davos, in English (January 2024)

On 17 January 2024 Argentina's president, Javier Milei, gave a speech in Spanish at the World Economic Forum in Davos. Aaron Slodov, the co-founder of an American manufacturing start-up, posted a version translated into English with HeyGen, with Milei's own voice and matching lip movements. According to The National, the clip passed 75 million views on X. Many viewers who would never have watched a Spanish-language speech watched this one. It is a good example of the legitimate promise: the words were Milei's own, and the translation opened his speech to people who would never have sat through it in Spanish. It also shows how little it takes to make a head of state appear to speak a language they didn't use, which is why clear labelling matters so much.

Stolen faces selling pills and propaganda (2024)

In March 2024 The Washington Post reported on women whose faces and voices had been taken from their own online videos and turned into adverts they never agreed to. One, the creator Michel Janse, found herself in an advert for erectile dysfunction pills, made from a YouTube video she had posted about her divorce. The Post reported that people making such ads used tools including HeyGen and ElevenLabs to clone voices and re-sync lips.

The same year, Voice of America reported on Olga Loiek, a Ukrainian student at the University of Pennsylvania, who found dozens of videos on Chinese platforms in which a copy of her face, under names such as "Natasha", spoke Mandarin as a Russian woman, praised China-Russia friendship and sold Russian goods. VOA reported that some of these videos on Bilibili carried HeyGen's logo, and that a tutorial there showed how to make one with HeyGen using a clip of her. HeyGen's own policy forbids avatars of real people without explicit consent; VOA said HeyGen did not respond to its questions. After complaints from Loiek and her followers, about a dozen imitating accounts were removed. The lesson for creators is uncomfortable: if you post long videos of yourself talking, you have already published the raw material.

An everyday scenario: an author's welcome video in two languages

Picture Marta, who writes cosy mysteries and sells well in the US and Mexico. She films herself for a few minutes in good light, reads HeyGen's consent statement on camera, and builds her own avatar. She writes a 60-second welcome video for her website ("Hi, I'm Marta, here's what my books are like"), generates it, and watches it closely: one hand gesture looks odd and her fictional village's name is mispronounced, so she respells the name and regenerates. She then translates the video into Spanish, and asks a Spanish-speaking reader to check the translation before posting. Because the video promotes her books, which is commercial use, she pays for a month of a paid plan rather than using the free one, and she adds a line under the video: "Made with an AI version of me; the words are mine."

Try it yourself

Plans, limits and menu names change often; check HeyGen's pricing and terms pages for the current details before you start.

  1. Sign up for the free plan at heygen.com. At the time of writing, the free plan needs an email sign-up but no credit card, and allows three videos a month of up to one minute, with a HeyGen watermark on every video.
  2. Start with a stock avatar. Choose one of the ready-made presenters rather than yourself. It is the quickest way to learn how the tool behaves.
  3. Paste a short script of about 60 to 80 words, which is roughly 30 seconds of speech. For example, you could paste: "Hello, and thanks for stopping by. In the next thirty seconds I'll tell you what this channel is about, who it's for, and how often new videos come out. If that sounds useful, stay a while."
  4. Choose a voice and preview it. Listen for mispronounced names, numbers and abbreviations. Fix them by editing the text, for example writing "twenty twenty-six" instead of "2026", or spelling a name the way it sounds.
  5. Generate the video and watch it three times: once normally, once with the sound off to watch only the mouth and hands, and once looking only at the edges of the face and hair. Note anything that looks wrong. This is also practice for spotting AI video in the wild.
  6. Try the translation feature on a short clip of yourself, if the free plan allows it at the time you try. Use a quiet recording with only you on screen. Then ask someone who speaks the target language whether the translation is right; don't trust it blind.
  7. Only then consider your own avatar. HeyGen asks for a separate consent video of under 30 seconds, in which the person being copied reads a script and a code on camera. If you are making an avatar of someone else, they must record that consent themselves. Never try to work around this.
  8. Read the terms before you publish. At the time of writing, HeyGen's terms say that videos made on the free plan are for personal, non-commercial use only and may not be sold or used commercially; paid plans (the cheapest was about 29 dollars a month when this was written, with a monthly allowance of credits) give you rights to use your output commercially. The terms also say you must disclose that your video was made with AI. Check the current version yourself.

Common mistakes and misunderstandings

Thinking the free plan covers business use. A welcome video for your book, an advert or a paid course lesson is commercial use. At the time of writing, HeyGen's terms don't allow free-plan videos to be used that way, and the watermark would be there anyway. Budget for a paid month if you want to publish.

Making an avatar of someone without their permission. HeyGen's policy forbids it, many places now have laws about using someone's likeness, and it is simply not yours to do. That includes a friend, a relative who has died, or a celebrity "as a joke". The consent recording exists for a reason.

Not telling viewers. HeyGen's own terms, at the time of writing, require you to disclose that a video was made with AI. It is also good sense: viewers who later discover an avatar was presenting without being told tend to feel tricked, and trust is hard to rebuild. A single line on screen or in the description is enough.

Trusting the translation because the lip-sync looks perfect. The visual polish says nothing about accuracy. Idioms, jokes, names and numbers are where machine translation slips, and a confident face makes an error look deliberate. Have a fluent speaker check anything you publish.

Expecting a finished video in one click. Avatars still produce odd hand movements, stiff pauses, over-bright faces and mispronunciations. Longer videos give more chances for something to look off. Plan to watch everything and regenerate the parts that fail, which also uses up credits on paid plans.

Assuming an avatar will feel like you. An avatar reproduces your face and voice, not your judgement or warmth. For a quick explainer that may be fine; for a heartfelt message to readers, many people find a simple phone video of the real you works better.

Believing labels and logos will always show what is fake. HeyGen says it is a member of the Content Authenticity Initiative, a group promoting "content credentials", the invisible origin labels explained on the OpenAI Images page. But labels can be stripped, and the misuse cases above show videos circulating far from where they were made. Judge a surprising video by where it came from, not by how real it looks.

Forgetting who owns what. HeyGen's terms, at the time of writing, give paid users ownership of their output, but the wider question of whether AI-generated material can be copyrighted is unsettled; the US position so far is covered on the Midjourney page.

Related elements