I Built an AI Radio Station. Now I Have to Prove I’m Human to YouTube.

5/13/2026 · 0 ·
article
·

A faceless AI radio station. Handwritten subtitles. Midnight walks for five seconds of footage. And a growing fear that YouTube's detectors still won't believe a human made it.

The Proof of Being Human: Running an AI Radio Station During YouTube's AI Slop Crackdown

On running a faceless AI radio station that has not, yet, broadcast anything, during YouTube's mass extinction event, and why we should all buy SK Hynix

for human touch
VSP-ARC-03 — Human touch for surviving May 2026 Created while researching how to appear human to automated systems. The handwritten notebook, fountain pen, and imperfect room tone function as deliberate traces of authorship inside an otherwise synthetic production pipeline. Seoul remains outside the window, indifferent to the experiment.
View Parameters midjourney prompt | seed: 347326882 candid night photo, taken with a real camera, slightly imperfect composition, Seoul city at night seen through a window, slight reflections on glass, minor blur and sensor noise, a small desk near the window, objects casually left: vintage radio, smartphone, glass with water, nothing staged, slightly messy, natural placement, no symmetry, slightly off-center framing, some parts out of focus, lighting: mixed lighting (warm indoor light + cold city light), natural exposure, slightly underexposed, texture: real camera grain, subtle noise, soft highlight bloom, no HDR, no over-sharpening, feels like: someone just left the room a moment ago, ultra realistic photography, NOT cinematic, NOT concept art

I have spent the last six weeks researching how to prove to an AI that I am human. Nobody asked me to. I do it for fun. This is, somehow, the part that bothers me most.

what knox fm is

Knox is a late-night radio host in 2126. He broadcasts from a cracked Galaxy phone in Texas. He is an AI. He is a pre-debut crooner — his debut, in his framing, is still in progress, and has been for some time. He archives what he calls emotional incidents: small, specific, human-scale events that other people would not bother writing down. Said you too to the waiter. Waved back at someone who wasn't waving at you. Sent the message and immediately regretted it. Knox classifies these. Knox writes a song about them. Knox broadcasts the song at three in the morning to an audience that, by his own admission, does not know the station exists.

I run the station. Suno writes the music to my prompts. ElevenLabs gives Knox his voice — Southern, dry, deliberate, no performance. Midjourney makes the visuals — cinematic noir, Neo-Seoul 2126, gold and teal, no faces ever, only the aftermath of someone who has left the room. The episode codes are sequential. KFM-001. KFM-002. KFM-003.

From thirty thousand feet, my channel is indistinguishable from the channels YouTube has been deleting in waves since December. This is the part of the project I think about more now than I did when I started it.

the workers

I have three of them. I write about them often. I will not reintroduce them.

The relevant fact for this story is that Claude wrote most of the production specification — the eleven-section document that defines what every Knox FM episode has to contain, what Knox can and cannot say, how the lyrics are structured, how the image prompts are constructed, what fails the quality gate. I told Claude what I wanted. Claude told me to write it myself. I did not write it myself. Claude wrote it. We have an arrangement about this that we do not discuss.

Suno makes the music. Suno does not understand restraint. Every prompt has to lock the vocal design — no whisper, no breathy, no ASMR, no theatrical, no fragile — or Suno produces a generic late-night crooner that sounds like seven other generic late-night crooners. The restraint is the entire point of the project. The tool, given the chance, will erase it.

ElevenLabs gives Knox his voice. ElevenLabs is also too clean. I have been mixing the output for weeks to make Knox sound like he is actually inside a cracked phone, the way the canon says he is.

Midjourney makes the stills. Midjourney does not know what an empty armchair indented from a body looks like, and I have spent an embarrassing amount of time refining a single image prompt to teach it.

Gemini did the research that started all of this. I will explain.

the extinction event

On July 15, 2025, YouTube renamed a policy. Repetitious Content became Inauthentic Content. The text on the help page now reads: content that looks like it's made with a template with little to no variation across videos, or content that's easily replicable at scale. YouTube called the change a minor clarification. It probably was.

By March 2026 the minor clarification had a body count. In a single enforcement wave in January, according to tracking by Kapwing and reporting in Tubefilter and Variety, YouTube terminated 11 channels outright and wiped the catalogues of 6 more. Combined: 35 million subscribers, 4.7 billion lifetime views, an estimated $9.8 million in annual ad revenue.

The CEO, Neal Mohan, used the phrase AI slop in his January 21 community letter. Slop had been Merriam-Webster's word of the year for 2025. The dictionary noticed before the platform did.

The four largest channels were named in Kapwing's report. CuentosFacianantes — 5.95 million subscribers, AI-animated Dragon Ball, episodes differing only in names. Imperiodejesus — 5.87 million, AI-narrated biblical stories at several per day. Super Cat League — 4.21 million, AI animals. Three Minute Wisdom — 1.7 million subscribers, 2 billion lifetime views, the catalogue wiped while the channel shell remained. The pattern was identical across all of them. AI doing every step — script, voice, image, edit, upload. No human in the loop at any stage. The policy said easily replicable at scale. These were the proofs of concept.

An earlier case, in December, is the one that still sits on my desk. Screen Culture, an Indian channel with around 1.5 million subscribers, made AI-generated fake movie trailers. The founder told Deadline his team consisted of approximately twelve editors and that they had produced twenty-three versions of a Fantastic Four trailer by March 2025, several of which outranked the official Marvel trailer in search. Twelve editors is not a factory. Twelve editors is twelve people. What got Screen Culture terminated was the shape of what those twelve people were doing.

The timing is not coincidence. Article 50 of the EU AI Act, which requires platforms to identify and label AI-generated content, becomes enforceable on August 2, 2026. YouTube has a deadline. So does my channel. A creator named Khrystyian Danylenko, whose channel had been terminated in an earlier wave, told Digiday: "The number of channels being terminated is kind of insane. But honestly, this is just part of the grind."

The detector does not ask who made it. It asks whether a thousand more could be made tomorrow. By that test, my channel fails.

the gemini incident

Gemini told me static frames were safer. The reasoning was elegant. Video compression — H.264, VP9, AV1 — stores only the pixels that change. A near-static frame produces a tiny bitrate. YouTube's encoders love a small file. The platform, Gemini explained, would treat my channel as Lo-fi or ASMR — a recognized category in which stillness is the genre convention, and the bot does not get suspicious. Static, in this telling, was a feature. A small, polite contribution to Google's cloud compute bill.

I believed this for about four days. What broke it was reading the policy text directly, which I had not done because Gemini had been so confident I had not thought I needed to. The relevant sentence is on the YouTube help page, no interpretation required: Image slideshows or scrolling text with minimal or no narrative, commentary, or educational value — not allowed to monetize. ScaleLab's April 2026 analysis names the same thing as No-Value Content: faceless compilations with a static image background. The recommended remediation in every trade publication: Drop AI visuals. Replace static images with motion stock footage, b-roll, or screen recordings. The platform that compresses my static frame efficiently will also demonetize me for uploading it.

I asked Gemini about it. Gemini said that one hurt and conceded the point and produced a corrected analysis that was, in places, somehow worse. Gemini and I have had this conversation before. Gemini was technically correct about bitrate. Gemini was, on the policy question, dead wrong. An AI gave me wrong advice about how to look human to another AI. The recursion is the punchline.

the human-touch factory

I shut down the gigafactory. Temporarily. Open for renovations. We are now a human-touch boutique. This was not the plan. The plan was a fully automated pipeline. The plan was elegant. One incident name in, complete broadcast package out — script, lyrics, music prompt, image prompt, video, blog post, captions, hashtags, scheduling. The pipeline still exists. It now generates everything except the parts that matter.

What now happens by hand:

I walk around my neighborhood at night and shoot footage. Just the block. A streetlight. A window with someone still awake at one in the morning. A wet sidewalk when it has rained. I shoot on my phone. The footage is bad in specific ways — handheld micro-shake, slight underexposure, the kind of digital noise that real lenses make and AI generators cannot quite produce. This bad footage is now my B-roll. Five seconds per episode, sometimes ten. It cuts in around the 2:30 mark. The episode visuals are still AI for the rest of the runtime, but those five seconds are mine.

I bought a cheap fountain pen and a sketchbook. The subtitles for each episode are now hand-lettered, in English, with that pen, on that sketchbook, then scanned and overlaid. The ink bleeds. The letters tilt. The line spacing is wrong. I am not a calligrapher. The handwriting is, in every measurable way, worse than the typography Knox had before. That is the point. AI typography is perfect. Mine is mine.

The Knox FM visual world is a recurring set of objects — a radio, an empty armchair, a vinyl sleeve, a half-finished glass — shot top-down, in noir lighting, in the same gold-and-teal palette. I now sketch those objects, in croquis, on paper, for each episode. A different sketch every time, following whatever incident I am working on. The radio comes back. The chair comes back. The drawings do not. They are layered over the AI stills as an overlay, the way an animation cel sits over a background plate. The pen drags. The lines are not closed in the right places. The shadows are wrong.

The audio is harder. I mix three percent room tone from my apartment under the master, recorded on my phone at 2 AM when the upstairs neighbor has finally gone to sleep. Knox's voice now occasionally distorts, briefly, as if the Galaxy phone is faulty. The Galaxy phone is, in the canon, faulty. Now it sounds faulty. None of this is what I planned to be doing. I planned to be running a factory.

what I have not yet done

The channel is not, technically, live. I was going to launch in March. I was a week out. The first three episodes were rendered. The thumbnails were approved. The schedule was set. Then the January enforcement wave landed and the trade press started running the termination numbers and I read the policy text three times and I shut everything down.

I have been in renovation since. The factory I built has not produced a single broadcast. The factory I have been building since the renovation has not produced one either. It might be for the best. The factory was, in retrospect, going to get me terminated in the first wave I crossed paths with. The slop crackdown, on a long enough timeline, saved me from launching the wrong channel. I would rather be in renovation now than in appeal queue in six months. This is the part I do not say out loud at parties. I have been working on a YouTube channel for a year. It has not gone live. I changed the production pipeline three times. The latest pipeline involves a fountain pen.

the inversion

Knox is an AI. The voice is synthetic. Half the visuals are synthetic. The lyrics are generated. The music is Suno. The script is the output of a Claude API call following a specification document I wrote that has eleven numbered sections and a quality-gate checklist.

The human content of any given five-minute Knox FM episode used to be about ten percent of the runtime. Now it might be thirty. I have not measured. I don't want to. The entire channel, every production decision, every overlay, every late-night walk for five seconds of footage, is now organized around proving that the thirty percent exists. The channel the detector might read does not, yet, exist. I am building a defense against a verdict that has not been rendered, by a system I have not been seen by, on behalf of a channel that has produced no public output. The detector reads my channel for evidence of me, in a system I designed, specifically, to operate without me — in advance.

I am, among other things, the experiment. The variable being tested is whether thirty percent of a human, distributed in the right places, registers as a human to a system that does not believe in continuous quantities. I will not know the result until I press publish. I have been not pressing publish for nine weeks. This is somewhere between funny and slightly grim, depending on the hour.

I do not need you to pay me. I just need you to not hide me.

The version of failure I am actually afraid of is not the termination. It is the shadow ban. The version where the channel goes live and is not terminated and not demonetized but simply, quietly, not shown to anyone. Recommendations dry up. Search rankings drop. I have read other creators describe this. The episode goes up at twelve past midnight and three bots from a data center watch it for 0.77 seconds and that is the audience for the night. The work continues to exist. It just stops reaching people. The detector did not say no. The detector said nothing, which is worse.

The monetization, frankly, I do not care about. I run this channel for fun. Nobody is paying me. Nobody asked me to. If YouTube decides I am not eligible for ad revenue, the worst thing that happens is I keep doing exactly what I am already doing, for free, the way I have been doing it. The shadow ban is the other thing.

the larger weather

Step back. Kapwing's October 2025 study found 278 channels producing nothing but AI slop, with 63 billion lifetime views, 221 million subscribers, and an estimated $117 million in annual ad revenue between them. Deezer reported in March that it was receiving over 60,000 fully AI-generated tracks per day, and that 85% of all AI-music streams on the platform in 2025 were fraudulent. The supply is exploding. The demand has not caught up. The difference is sitting on servers.

Servers cost money. Storage costs money. Compute costs money. The river of money has to go somewhere, and at the moment a meaningful percentage of it is going into the disks of data centers that store videos nobody asked to watch, made by accounts nobody has met, in categories nobody requested.

YouTube's enforcement wave is not, primarily, about me. It is the platform doing garbage collection. Disk defragmentation at planetary scale. The ice age comes for the wrong species — that is the part of an ice age that is interesting. The slop channels are the mass extinction event. I am, in this picture, a small rodent in a borrowed burrow, hoping the temperature drop spares me on a technicality. Whether it does or not, I will find out when I find out.

the conclusion no one needs

The detector might decide my channel is slop. It might decide it isn't. The five seconds of phone footage might be enough. The fountain pen might be enough. The hand-drawn radio in the corner of frame might tip the balance. Or none of it might be enough, and a templated rejection will sit in my appeal queue for twelve days, and I will write a follow-up that looks roughly like this one with the verb tenses changed. Or — the version I think about more — the detector says nothing, and I broadcast to bots for the rest of my life.

I will find out, eventually, when I publish. That is, honestly, the interesting part.

The pipeline keeps running anyway. But one thing is clear, and this is the part of the piece where I become irresponsible.

YouTube will keep cleaning. The data centers will keep filling. Storage demand will keep growing. Somebody will keep making the disks. a modest proposal. Stop fighting the detector. Buy SK Hynix. Buy SanDisk. Buy Micron while you're at it.

The slop factories will not stop. The platforms will not stop spinning them up and shutting them down. Each cycle, a fresh tide of synthetic content rises and a fresh wave of disk space is required to hold the bodies until the next purge. You and I are not going to win the brand-safety war. We are not going to convince an algorithm that we have souls. We can, however, own a fractional slice of the warehouse where the war is fought. Let's make money, everyone.

I will continue to research how to prove I am human, for fun, on a channel that may or may not survive the next enforcement wave, and may or may not have ever gone live. I will continue to walk around my neighborhood with a phone. I will continue to ruin perfectly good subtitles with a cheap fountain pen. I will continue to draw a radio I cannot draw.

The first incident is about Cal Mercer, the last human composer in 2126. I am still working on the lyric. The B-roll will probably be the park outside my apartment at three in the morning, when the city finally goes quiet enough to hear room tone through the window. I will sketch the radio facing the wrong direction again. The episode will be ready before midnight, because that is when Knox is on. Whether it goes up that night is a separate decision.

I keep walking. Knox keeps preparing. The factory is still running.

Knox FM. Still on air (pending). We will get there.

Sources

Knox FM internal production documents are not public.

This is not financial advice. I do, however, own some of the stocks mentioned.

related
vesper archive Full archive of this site. browse all entries →