'Human-First' Is a Promise Nobody Can Check Yet.

By Cynthia Zhu, Co-Founder of HeartMuLa

The most interesting thing in Jason English's research on AI and streaming isn't the blind spot. It's a sentence near the end that he only casually lingers on:

"The "Human-First" label that's showing up among independent artists and labels is a market response to the blind spot working in reverse. The audience is paying attention. The artists who say plainly that the songs were written, performed, and sung by real people are gaining ground their AI-heavy peers think is being given away."

So, a "Human-First" label is showing up among independent artists and labels, and it's gaining ground. That's a market pricing "authorship." After three years of arguing about whether anyone cares, someone's putting a sticker on it. That's the most encouraging development in this whole mess.

Yet, right now, it's worth nothing as a signal. Anyone can type this label out, slap it on an image or webpage and use it to their own ends. But what does it really mean?

A label nobody can check is just a "vibe."

Every label that ever meant anything had machinery behind its proliferation and reputation. "Organic" certification has audits and a paid inspectorate. "Fair Trade" has chain of custody. "Made in Italy" has customs law and a fraud squad. The claim is what you see; the enforcement is what makes the claim worth seeing.

"Human-First" is giving you an artist's word. But an artist's word is precisely what stops being reliable at the moment the label starts converting — because that's when the incentive to claim it falsely shows up. The label gets more valuable and less trustworthy on the same schedule. That's not cynicism about musicians, it's just how unaudited claims behave, in every market where they've ever been tried.

So the question isn't whether Human-First is good, it's who's going to build the machinery, and out of what exactly?

The blueprint is next door.

Software went through this exact fight and the receipts of those transactions are public. For twenty-odd years, "open" was a marketing word in software the way "human" is becoming one in music. Anyone could say it, use it, manipulate it.

The Open Source Initiative's answer wasn't to audit every project — nobody could afford that. It was narrower and much cheaper: maintain a definition, hold it outside the companies that benefit from it, and let anyone check a claim against it in ten seconds.

That still works, and you can watch it work. Stability AI ships Stable Audio Open. The name says open. The license is the Stability AI Community License: it does permit commercial use — that's an actual section — but the grant terminates automatically once your organization passes $1M in annual revenue, and commercial use requires registering with Stability.

Useful model, real license, entirely legitimate product. It's just not open source by the definition, and because a definition exists that isn't Stability's, anyone can say so without it being a slur.

Contrast ACE-Step. ACE Studio and StepFun put ACE-Step 1.5 out under MIT — no revenue cliff, no registration, no takebacks on the copy you've got. Both models are "open" in ordinary English. One word, two completely different sets of rights, and the only reason you can tell them apart in a sentence is that somebody neutral maintains the word.

That's the transferable part. Not audits. A definition by an owner who isn't selling anything.

*Screenshot courtesy of* *HeartMuLa's Explore page*.

Where will music's version have to live?

Here's where I'd push back on my own analogy, because a definition alone won't take us all the way there. OSI works partly because software claims are checkable by strangers — the license is in the repo, the code is right there. "A human sang this" isn't checkable by inspection. There's no detector worth trusting and there won't be one soon.

So the enforcement has to come from somewhere else, and there's an obvious place: the upload form.

Every release already passes through a distributor. DistroKid, CD Baby, TuneCore, Bandcamp — they already make you attest to things, under contract, with account termination behind it. That's not a hypothetical inspectorate; it's a checkbox and a Terms of Service that already exists. If the attestation happens there, it's contractual rather than aspirational. Lie on it and you're not being unfashionable, you're in breach.

And the attestation has to be granular, because the current version is useless. "AI was involved" covers a mastering assistant and a synthetic lead vocal identically, which tells a fan nothing and correctly gets ignored by artists. Five lines fixes it:

Composition: human / AI-assisted / AI-generated
Performance: human / AI-assisted / AI-generated
Lead vocal: human / AI-assisted / AI-generated
Production chain (mastering, stem separation, cleanup): AI used / not
Marketing assets (copy, artwork, video): AI used / not

I'd bet English's 81% is answering about the first three. But because disclosure today is one undifferentiated blob, the artist who used a mastering plugin and the artist who generated a vocal make the same admission and eat the same penalty. That punishes the honest and instructs everyone else to say nothing, which is roughly what's happening.

Fans get five fields instead of a vibe. Artists get to disclose the boring parts without being lumped in with the synthetic-vocals crowd. And the platform — which already has the contract, the account, and the kill switch — becomes the enforcement layer, at approximately zero marginal cost.

The obligatory caveat, and then the one about me...

English's sample size deserves its asterisks and he's upfront about them: 573 respondents, self-selected, distributed through the Curious Goldfish audience — a podcast about Americana, folk, roots, bluegrass and country. That's close to the most human-authorship-friendly listener base in music. The 81% who called AI music inauthentic aren't a cross-section of the market. Read it as a signal from one room, not a verdict on the industry, which is exactly how English framed it.

But the room matters. Those are the genres where people still buy tickets and merch and turn up twice.

And me: I work in GenAI music. English's data says people in my position read this worst — creators who don't use AI put fan concern about human authorship at 67%, frequent AI users put it at 14%, a 4.7x gap. I have no rebuttal. I'm inside the sample that's wrong.

Which is the whole argument, really. Open source never assumed anyone would be honest, including itself. It wrote the definitions down, put them somewhere its own funders couldn't edit them, and made claims checkable by people with no stake. That isn't idealism — it's a system built by people who assumed everyone involved, themselves included, would eventually be tempted.

Human-First is currently running on honor, in an industry where most independent artists are under real financial pressure. That's a lot of weight on a lot of honor.

Give the label some teeth before somebody hollows it out. The fans are already asking. We just don't have a way to answer that means anything yet.

Cynthia Zhu is a Co-Founder of HeartMuLa, an open-source AI music project that releases its models under Apache-2.0 on Hugging Face. She writes about open source, transparency, and where the music industry goes next.