Opinion
Made With AI is Not the Same as Generated by AI
A songwriter's view on what detection software actually detects, and what it misses.
I have been writing and producing songs for decades. The ones I'm releasing now were written and arranged decades ago, demoed in a home studio where I played the parts myself, or cut them in New York rooms with first-call session players and the kind of studio singers you hire when you want the song to land. I have used the best production technology of every era I worked in. AI is the next one. I don't find that controversial; I find it obvious.
Here is what I do with it. I take a finished demo of a song and process it through an AI music platform, then stem the output, add MIDI, edit across multiple takes, and finish it on a 2026 desktop rig. For $24 a month I get something I could never afford otherwise: unlimited session musicians and studios, available 24/7, who hold no rights in my song. Directing them is no different from telling a live rhythm section to take the second verse quieter, or asking a pianist to try different inversions. The players change. The song doesn't.
That distinction is the one the current conversation keeps missing, so it's worth being precise about it.
A song and a recording of that song are two different things, legally and creatively. The song is the melody, the harmony, and the words. It's the part you'd recognize if it were played as a quiet bossa nova or as Nu Disco. The recording is one performance of it. You can process a recording through reverb, autotune, a tape emulator, or an AI model, and none of that touches who wrote the song. And to be clear about scope: this isn't a conversation about text prompted AI-composed music. It's about professional and aspiring writers honing their creative craft with new performance technology the way every generation of musician has used the studio tools of their day. AI processing leaves a mark on the audio that the others don't, and given musical input like a complete song, it can just process a performance without authoring a new composition.
What it cannot do, working from the audio alone, is tell you who wrote the lyrics and melody, whether anything was copied, or whether a human spent a month shaping the result.
That mark on the audio is exactly what detection software finds. The newest tools look for a synthetic signature in the audio: a temporal perfection, spectral fingerprints, the too-clean quality of a generated waveform. That detection technology works. It does what it says. What it cannot do, working from the audio alone, is tell you who wrote the lyrics and melody, whether anything was copied, or whether a human spent a month shaping the result. It detects that an AI platform touched the sound. It infers nothing reliable about authorship. Those are different questions, and only one of them is what copyright is about.
The reason this matters is that the inference is being made anyway, at scale. Deezer reported in April that it now receives roughly 75,000 AI-generated tracks a day, about 44% of uploads. My decades-old songs are counted among them. But the same press release contains the more interesting numbers: by Deezer's own measurement, up to 85% of streams on those tracks were fraudulent, and in a blind test 97% of listeners couldn't tell AI-made music from human-made music. Read those together and the real problem comes into focus. The harm is fraud: bot farms diluting the royalty pool. Not the existence of music made with a new tool. "AI-generated" is being used as a stand-in for "fraudulent" and "worthless," and honest human work gets swept into the same bucket because it happens to share an audio signature with the fraud.
I'm not arguing the training-data question, which is real and unsettled, and I'm not arguing platforms should ignore fraud. They shouldn't. My point is narrower: the leap from "we detected a synthetic signature" to "this is not human creative work" is a marketing claim, not a finding, and it's now being licensed to the rest of the industry as an industry standard. A synthetic signature should open a question about provenance and authorship. It shouldn't close one with a ban.
The people who get this right will be the ones building the tools and the platforms to tell craft from fraud, instead of treating every producer who walked into the new studio as a counterfeiter.
· · ·