Author Replaces Himself with AI Avatar on YouTube: Average View Lasts 1 Minute Instead of 5
The founder of a word-learning app built an AI copy of himself and launched a YouTube channel without filming anything. Claude Code wrote the scripts, ElevenLabs produced the voice, and HeyGen generated the face. A ten-minute video cost $2–3, but the average view duration for the avatar was about 1 minute, compared with about 5 minutes for the live author recording on a 300-ruble microphone. YouTube recommends videos that people watch to the end, so cheap production without audience retention does not work.
AI-processed from Habr AI; edited by Hamidun News
Imagine not having to set up lights, record ten takes, or fight with your diction: you write a script, and a digital copy of yourself reads it on camera. The author of a mobile app for learning foreign words tested this approach on YouTube. He built an almost fully automated pipeline in which a ten-minute video costs $2–3.
But YouTube's analytics showed the opposite. The average view duration, meaning how many minutes a viewer watches on average before closing the video, was about 1 minute for videos with the avatar. For regular videos, where the author sits in front of the camera himself, it was about 5 minutes.
And this is the metric that determines whether YouTube recommends a video to other people.
How the pipeline works
The author was looking for organic traffic for his app and decided to launch a channel about word-memorization techniques. The app could be mentioned during the videos.
First, he tried services that animate images. A video from Hedra turned out so weak that he had to delete it. King AI did even worse. Only HeyGen, on its most expensive model, Avatar 6, produced acceptable quality, at roughly $1 per minute of video.
To make the image resemble a multi-camera shoot, he took photos of himself from different angles. Using ChatGPT, he turned them into scenes with the same generated studio, and from these he assembled several avatar configurations in HeyGen.
Next, Claude Code joined in. It wrote the script and broke it into scenes: which text is spoken on camera, which angle to use, when to show a close-up, and when to show an infographic. ElevenLabs voiced each fragment using a clone of the author's voice. He created the clone by reading a short text into a microphone. All the steps were repetitive operations with clear inputs and outputs, so he wrote the scripts without difficulty. The hardest part was getting an acceptable face.
Economics: from $25 to $2–3
The first full video cost about $25. Most of that went to video generation in HeyGen. The author decided to cut the avatar's screen time. The face stayed at the beginning, at the end, and during transitions between blocks, about 20% of the runtime. The rest was taken up by infographics, illustrations, and inserts. This had two effects: viewers less often see unnatural facial expressions, and a ten-minute video came to cost about $2–3. The author believes the videos even benefited, since visual explanations replaced the talking head.
He also refocused the channel on an English-speaking audience learning Spanish. In his observation, HeyGen's facial expressions looked noticeably better in English than in Russian.
What the analytics showed
He posted the first video on X, and it was torn apart in the comments. People wrote about the uncanny valley, the term for the unpleasant feeling when an artificial human looks nearly real, but small inconsistencies in facial expressions cause discomfort. Blogger Vlad Ten shared the video on Telegram. The result: several thousand views, 10 likes, and 100 dislikes. The paradox is that YouTube began showing the video to other people because of the activity.
The author fairly pointed out that this was an unfair test: everyone knew they were looking at an avatar and was deliberately searching for flaws. So he published the English videos without an announcement and compared them with his regular channel. There, he sits in a dark room with a 300-ruble microphone, with diction defects and filler words.
According to YouTube analytics, the average view duration on regular videos was about 5 minutes, and on the avatar videos about one minute. If videos are not watched to the end, the platform does not promote them. The initial spike in views was driven by interest in the "cringe" factor, not the topic of languages.
Some acquaintances who were shown the videos without a warning did not notice anything off. But the videos still failed to hold viewers. The author does not know the exact cause and suspects that: a live person stumbles, changes intonation, smiles at the wrong moment, and searches for words. The avatar reads the text smoothly but looks like someone who does not understand what they are talking about. This is a hypothesis, not a measurement.
What practitioners can take from this
- Production can be automated, but that does not solve audience retention. Ten videos a day will not help if each one is closed in the first minute.
- The metric to test such an idea is average view duration, not views or initial reactions.
- It is better to test blind, without an announcement, otherwise viewers will search for artifacts.
- Neural networks remain useful for scripts, illustrations, and infographics. In this case, they did not become a replacement for a live presenter.
The limitations are obvious: this is one author's experience, one channel, one niche, and a sample of a few videos. Other formats or future models could change the result. But for those who are postponing a channel launch because of poor lighting or diction, the takeaway is practical: a real person on camera still holds viewers better than a perfect copy.
- Nov 21, 2026Microsoft Releases MAI Code 2, a Coding Model Less Than Half the Price of Flagships
- Nov 21, 2026OpenAI Lets Six Apollo Auditors Inside to Look for AI Sabotage During Model Training
- Nov 21, 2026At Holiday Extras, AI assistants send reports and animate in 10 minutes instead of days
- Nov 21, 2026Nvidia Doubles Quarterly Revenue to $115.84 Billion After Huang's Exit
Key Questions Answered
How the pipeline works?
The author was looking for organic traffic for his app and decided to launch a channel about word-memorization techniques. The app could be mentioned during the videos. First, he tried services that animate images. A video from Hedra turned out so weak that he had to delete it. King AI did even worse. Only HeyGen, on its most expensive model, Avatar 6, produced acceptable quality, at roughly $1 per minute of video. To make the image resemble a multi-camera shoot, he took photos of himself from different angles.
What are the key details on Economics: from $25 to $2–3?
The first full video cost about $25. Most of that went to video generation in HeyGen. The author decided to cut the avatar's screen time. The face stayed at the beginning, at the end, and during transitions between blocks, about 20% of the runtime. The rest was taken up by infographics, illustrations, and inserts. This had two effects: viewers less often see unnatural facial expressions, and a ten-minute video came to cost about $2–3.
What the analytics showed?
He posted the first video on X, and it was torn apart in the comments. People wrote about the uncanny valley, the term for the unpleasant feeling when an artificial human looks nearly real, but small inconsistencies in facial expressions cause discomfort. Blogger Vlad Ten shared the video on Telegram. The result: several thousand views, 10 likes, and 100 dislikes. The paradox is that YouTube began showing the video to other people because of the activity.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.