Three or four years ago, I tried YouTube.
I made stick-figure animation and some music content. I kept at it for more than six months and earned nothing.
Looking back, the problem was not only topic selection or persistence. The practical constraint was simpler: every video took too long.
Research, scripts, images, editing, thumbnails, and uploading all depended on me. Finishing one video was tiring enough. Testing several directions at once was out of reach.
Now I am trying again.
Not because earning money on YouTube has suddenly become easy, and not because I have found a guaranteed method. I returned because one of the constraints that stopped me has changed: AI and automation are changing the cost of sustained content production for a solo creator.
Two visual directions; each still needs human selection and review
The opportunity is not one-click video generation
The usual pitch for AI content is simple: enter a prompt, generate a video, upload it, and earn money.
I do not believe that story.
Generating a video is not the same as getting people to watch it. Views are not the same as sustained channel growth. Growth is not the same as revenue that covers production costs.
What interests me is whether AI can help one person build a complete content-production system.
Much of the work that wore me down was repetitive and necessary: finding candidate topics, analyzing videos, extracting transcripts, mapping structure, generating images, creating thumbnails, testing openings, and carrying feedback into the next iteration.
AI does not decide what is worth making. It does make continued production less expensive in money, time, and attention.
I did not choose history first; I screened for a production model
My new English history channel is Past Made Us. It starts with ordinary parts of modern life, then looks backward: how did people live without them, and how did those things evolve into what we know today?
I did not choose the niche only because I like history. I first found similar channels and asked whether this format could be produced consistently by one person using AI. I use three filters.
First, the channel should be relatively new. An established channel may benefit from years of accumulated audience. A channel that began publishing only a few months ago is more useful evidence that the format may still be worth testing now.
Second, I do not want to be persuaded by one viral video. I look at the broader performance of the channel.
Third, and most important, the production structure must be automatable. The format should be faceless, not dependent on the creator appearing on camera. Its visuals should be possible with generated images, light animation, or relatively simple shots. Its structure should be analyzable, and its source and copyright risks should be manageable.
I am not looking for a channel that merely looks profitable. I am looking for a content model I can realistically keep producing and testing.
Past Made Us has begun publishing; this is evidence of a working production loop, not proof of commercial success
Systems are built by producing, not by planning
Past Made Us has published its first batch of videos. I can currently produce roughly two or three in a day, a very different constraint from the last time I tried.
The tool I use most is Codex. It does more than help with a paragraph of copy; it participates in building and changing the workflow itself.
I use TranscriptAPI.com to retrieve transcripts from reference videos, then use Codex to analyze narrative structure, pacing, and information flow. Other skills help with thumbnails and visual assets. The primary video images come from OpenAI image generation.
The workflow was not designed perfectly before production began. The real loop looks more like this:
Production exposes bottlenecks that planning cannot predict.
Automation carries repetitive production; people retain decisions about direction, quality, rights, and continued investment
I am testing the opening as I go
My earliest videos opened with static images. I later used Grok to add simple motion to several images. Now I am experimenting with a third-party video API to produce a more forceful opening.
I do not assume that a more expensive moving opening will perform better.
The useful approach is variation: static, lightly animated, and more ambitious video openings can be published and compared. If the expensive version does not improve the result, there is no reason to keep paying for it. If it does, I can then work on lowering the cost.
That is what an AI workflow means to me. It does not make one decision on my behalf; it lowers the cost of testing decisions repeatedly.
The platform evaluates the finished work, not the tool alone
Automation is not a license to mass-produce low-value videos.
YouTube's current channel monetization policy says repetitive or mass-produced template content without original value is ineligible for monetization. It also gives examples where AI-assisted scripts, editing, and original visual storytelling can still qualify. The finished work must carry original narrative, perspective, educational value, or entertainment value.
YouTube's disclosure guidance for generative AI focuses on meaningfully altered or synthetic content that appears realistic. Even when generated historical imagery supports a narrative, it does not remove the need for factual review, rights review, or an appropriate disclosure decision.
I want to automate repetition, not responsibility.
The most valuable thing to automate may be boredom
This experiment is still early. The channel has published its first videos and received some views, but it is nowhere near proven, and certainly not a validated monetization system.
I am continuing because the hardest part of YouTube may not be making one video.
The hard part is publishing when the numbers are ordinary, doing it again without clear feedback, and continuing to research, produce, upload, and wait when you still do not know whether the direction is right.
It is boring.
Many people do not stop because they are incapable. They stop because repetitive work without feedback is difficult to sustain.
If a system can carry the most repetitive parts, a person can spend more attention on the decisions that matter: which topic deserves another attempt, whether the content is good, which opening works, whether the cost is rational, which steps can be automated, and what should not be published.
AI does not persist on my behalf. It makes the most abandonment-prone part of persistence easier to execute.
I am giving the experiment three months
I will not declare the idea dead because a few videos perform poorly. I will not declare victory because one video receives more views.
I am giving this experiment three months.
The test is larger than “did I get a viral video?” I want to know whether the system can keep running, whether quality holds, which topics deserve continued investment, what a video actually costs, which steps should be automated, which judgments must remain human, and whether the result can eventually produce sustainable revenue.
My first attempt proved that I could work for six months. It also showed that slow, manual production did not give me enough capacity to test properly.
This time I am testing a different question:
When AI and automation take over the boring, repetitive, time-consuming work, can one person build a YouTube content system that keeps running long enough to learn what actually works?
I do not know the answer yet.
But this time, we finally have the tools to test the question properly.