← Journal

Should AI Music Be Labeled? Where I Land on Disclosure

I turn the AI toggle on every upload and it costs me reach. Why I still do it, and why a label on the file is not the same thing as a confession in the caption.

There is a switch on the TikTok upload screen marked AI-generated content. I turn it on every time. It costs me.

Labeled posts get handled differently. The platform knows, the ranking knows, and the video goes out into a smaller room than it would have otherwise. I have watched it happen across a whole album rollout. So the honest version of this question is not should artists disclose. It is whether you will disclose when disclosure is the expensive option.

I do. Here is the reasoning, and it is not moral.

A label is metadata. It belongs to the file, the same way tempo and key and songwriter credits belong to the file. Spotify moved on this, backing an industry credits standard so a record can carry how it was made down the pipe. Deezer started tagging fully generated tracks in its own catalog. The infrastructure is arriving whether artists cooperate or not, and the artists who wait will have the tag applied to them instead of by them.

Then there is the other reason, which is Velvet Sundown. A band appears, pulls a million monthly listeners, denies being AI when asked, and later confirms it was a generated project the whole time. The music was not the scandal. The denial was. Every listener who had defended them found out they had been managed. That is the actual risk of hiding it, and it is not a fine or a takedown. It is that the audience learns your first instinct was to let them assume.

So I label the upload. I do not label the caption.

That distinction is the whole thing for me. The toggle is a fact about the file. A caption that opens with a disclaimer about how the song was made is something else: it hands the listener a frame before they have heard a note, and it turns the tool into the subject. I wrote the words. I built the story bible, the nine cards, the sequence, the thousand rejections nobody sees. The model made sound. Disclosure tells you that honestly, in the place built to hold it, and then gets out of the way.

Where I am still unsettled is the threshold. Nobody has drawn the line between a record generated end to end and a record where a producer ran a vocal through a model for two bars. Autotune was never disclosed. Sampling was, because money was attached. My work sits far enough over the line that the answer is easy for me, and I know plenty of people whose work sits right on it and who are being asked to sort themselves into a box that does not exist yet.

Label the file. Tell the truth in the place the truth goes.

Then make something good enough that the label is the least interesting fact about it.