I Spent an Afternoon Proving I'm Human. The Machine Got the Benefit of the Doubt.
The Human Frequency — Notes on making music by hand in the age of AI
By Roger Ricks
Let me ruin my own post in the first paragraph.
I set out this week to get falsely accused. There are free tools on the internet that listen to a song and estimate whether a machine made it, and I had read enough about their error rates to expect that mine would come back flagged. That was going to be the piece: the guy who builds his own guitars, marked as a robot. Good headline. Easy to write.
I passed. Four songs, four clean verdicts. So that post doesn't exist, and what I found instead is stranger and, I think, worse.
What came back
I used SubmitHub's AI Song Checker, which is free, needs no account, and runs two models side by side — one reading the frequency spectrum, one reading how the music unfolds over time. Its creator says on the page itself that it gives you one data point and not proof, which is more honesty than any paid competitor manages.
"Clarity" came back 91% human on the spectral side, 96% on the temporal. "That Day" came back 89% and 99%. "Ready To Stay" came back 91% and 82%. All three: made by humans, strong human characteristics detected.
Then, as a control, I sent it "Billie Jean". Recorded in 1982, played by session musicians in Los Angeles, produced by Quincy Jones. It came back 91% and 100%. The same spectral number as two of my songs, to the percentage point. I'll be honest about the small pleasure in that, and then move past it, because it isn't the story.
The song with nobody in it
In August 2025, Rick Beato sat down for a CBS News segment with the reporter Dave Malkoff and built a pop singer out of nothing. He used Suno. The persona and the lyrics took seconds, the melody and the voice took about two minutes. He called her Sadie Winters and the song "Walking Away", and he did it on camera specifically to show people how far this had come.
It escaped. Strangers cut music videos for it. More Sadie songs appeared. One channel gave her a biography that reads like a real one.
There is no human performance anywhere in that track. Not a note, not a breath. It is the cleanest test case anybody could ask for, and it is publicly documented by the man who made it.
SubmitHub's verdict: Hybrid — AI and human.
The spectral model was not fooled. It put the odds of a human being involved at 5%, twice, on two separate runs from two different uploads. It saw exactly what was there. But the verdict printed at the top of the page still credited a person, and the temporal model went further: on the first upload it thought there was a 44% chance a human had played this, and on the second it raised that to 84%. Same song. Different YouTube channel, different compression. Forty points.
Five runs across two tools, and not once did any of them produce the sentence that happens to be true: nobody was in the room.
That's the part I keep turning over. These tools can confirm me. They cannot deny the machine. They will tell you what something probably is not, and they go quiet on what it is — which is precisely the gap I keep running into everywhere else in this argument. Spotify's new badge marks profiles whose identity is generated, and says nothing about how the music was made. The detector marks audio that smells generated, and then hedges the conclusion into a shared credit. Both stop at the same wall.
Why I wanted a hundred
Here is the thing I actually want to tell you, and it is smaller than any of the above.
When the first result came in at 82%, my immediate reaction was: but it should be a hundred. I sat in my flat, in front of a song I wrote, played, sang and paid three people in Nashville to help me finish, and felt a flicker of wanting to argue with a webpage.
Eighty-two is a fine score. There is measurement error in any system like this and I know that perfectly well. That isn't what the feeling was about. The feeling was about a machine having an opinion on whether I exist, and me caring what it said. And that's the small version of something I don't have a good answer for.
I come from a completely analogue childhood, and I like technology — I test things, I enjoy them, I record through modelled amplifiers and I'm not the purist in this story. What has changed isn't my equipment. It's the setting I approach everything in. The default used to be that a thing was real until something felt off. Now it's the other way round, and the cost of that reversal gets paid in small amounts, all day, by everybody.
The place I notice it most isn't music at all. It's podcasts. I'll be watching something on YouTube where a confident person makes a serious claim and cites a source, and I cannot tell whether any of it holds. Not because I'm gullible. Because checking it properly would take the evening, and I have a day job and a family and three songs a year, and there are forty more of these tomorrow. So I close the tab with the question still open, and it never gets answered, and that happens again the next day.
Multiply that by everyone. Something gets worn away that I don't think we've named yet. Not truth exactly but the ordinary willingness to take what someone shows you at face value, which is most of what holding a society together actually consists of.
Now the part where I argue against myself
My instinct is to go one step further and say that algorithms are hardening everyone's opinions and shrinking their horizons, that we're each being fed our own confirmation until the other side looks insane. I looked into it, and I have to walk that back.
The research is genuinely divided. The Reuters Institute's literature review on echo chambers and filter bubbles found that people who use social media tend to encounter more news sources, including ones they disagree with, than people who don't. Around 10% of UK respondents said they almost never see political content they disagree with — which means 90% do. A Swedish study found that heavy consumers of news from one side of the spectrum were also heavy consumers from the other. And a systematic review of 129 studies found that the answer you get depends largely on how you measure: track people's actual clicks and echo chambers look small, model their friendship networks and they look large.
So the confident version of my complaint isn't supported, and I'd rather say so than have you find out from someone else. What survives is narrower and I still believe it. The problem isn't that I only see one side. It's that I can't fully trust any side I see, including the one I agree with, and suspicion is a poor material for building anything shared out of.
Where I land
The detectors work better than I expected and they help less than I hoped.
They gave me a verdict nobody asked for. No curator has ever accused me. No listener has written to ask if a machine made "That Day". I went looking for the accusation myself, which tells you something about the weather rather than about my songs, and when the machine cleared me it changed nothing, because I can't carry it around and hold it up in front of the next thing somebody sends me.
What I'd want instead is the boring thing I keep coming back to: a line under a track saying who played what, declared by the artist, enforced the way false credits get enforced everywhere else in this business. Not a probability. A claim, made by a person, that can be checked and that costs them something if it's a lie. That's the difference between 91% and a name.
I could be wrong about how much of this matters to anyone but me. It's possible that in five years nobody asks the question at all, and the whole worry looks like a man from an analogue childhood being sentimental about provenance.
So tell me one thing, honestly: when was the last time you couldn't tell whether something you were reading or watching was real — and what did you do about it? Did you check, or did you close the tab?
I'd like to know, because I mostly close the tab. And if you want to hear what one stubborn human still does by hand, my music lives here.
Sources & further reading
SubmitHub AI Song Checker and the tool's own help page — free, no account, two runs per hour as a guest, a spectral and a temporal model, and the creator's own line that it provides one data point rather than definitive proof. All results in this post were run on 6 September 2026.
LetsSubmit AI Music Checker — the bAbI v2 model, MERT embeddings and a logistic regression classifier, and the only detector I found that publishes its real holdout accuracy rather than a marketing figure.
Deezer's public AI music detector — the 99.8% accuracy claim and the stated false-positive rate of fewer than 1 in 10,000, on the page that would not run for me.
CBS News on the making of Sadie Winters — Rick Beato building the persona and the song with Suno on camera for Dave Malkoff, and the fan-made videos that followed within days.
Walking Away, as reposted on YouTube — one of the uploads I tested. The two runs came from two different channels.
Reuters Institute: Echo chambers, filter bubbles and polarisation — a literature review — the UK and Swedish findings, and the general point that most people encounter more cross-cutting content online than off.
A systematic review of echo chamber research, Journal of Computational Social Science, 2025 — 129 studies, and the finding that the measurement method largely determines whether echo chambers appear at all.
Spotify: Introducing a new label for AI-generated artist identities — the AI Persona badge, rolling out mid-September 2026, covering the artist's public identity rather than how the music was made.
Kommentare