Your music is never used to train an AI. Here's exactly why.

If you make music, you have every reason to be wary of AI. Over the last few years some of the largest AI models were built on artists' work that was scraped without permission, without credit, and without a penny going back to the people who made it.

So when a platform you trust with unreleased tracks mentions "AI," you deserve a straight answer to a simple question: what is it doing with my music?

Here is ours, in one line: Songbox never uses your music to train AI. Not ours, and not anyone else's.

That is not a promise we are asking you to take on faith. It is a consequence of how we built the system, and below we will show you how it works and name the models behind it.

Analysis is not training

When you turn on AI tagging, Songbox listens to a track to describe it: genre, mood, tempo, key, and the instruments it can hear. Think of a librarian reading the spine of a book to shelve it correctly. The librarian does not become smarter by memorising your book, and our model does not learn anything from your track. It reads, it labels, it moves on.

This is the distinction that usually gets blurred. Training is when a model studies millions of songs and permanently absorbs their patterns into itself. Analysis is when a finished model listens to one song and reports what it hears. Songbox only ever does the second. Your track is never added to a training set, never used to improve a model, and never kept by the AI once the tags are written.

We run our own AI

Songbox used to send audio to an outside AI service for this. We have brought the whole thing in-house. The AI that tags your music now runs on our own servers, the same infrastructure that already holds your catalogue. Your tracks are not handed to any third party or exposed to anyone outside Songbox.

What is actually inside our AI

Most companies will not tell you what powers their "AI." We will, because everything we use is open, permissively licensed, and free to inspect. The models that understand your music were built by universities and non-profit research groups and released for anyone to use.

Where we went one step beyond that, and trained those models internally for our own specific purpose, we learned from one thing only: music that artists deliberately released to the world under an open licence.

Our sources, in full

• LAION-CLAP handles genre, mood and instrument understanding. Released by LAION, a non-profit AI research organisation, under a public-domain licence (CC0).
• Beat This! measures tempo and beats. Published by the Institute of Computational Perception at Johannes Kepler University Linz, under the MIT licence.
• librosa detects musical key. A long-standing open audio library, ISC licence.
• OpenMIC-2018 is the dataset our instrument detector learned from: 20,000 Creative Commons clips, released for exactly this purpose (CC-BY 4.0).
• Free Music Archive is the dataset our genre detector learned from: an openly licensed collection of tracks that artists chose to share under Creative Commons.

Every one of these is free to use commercially. We turned down other models because their licences were restrictive or non-commercial, or couldn't be clear on how they were trained.

What this means for you

• Your uploads are never training data. Ever.
• The AI reads your track to tag it, then forgets it.
• Nothing is sent to an outside AI company.
• AI tagging is opt-in. Leave it off and no model ever touches your music.

• We can name every model we use, and we just did.

Why we bother

Songbox exists so musicians can share work they are not ready to release, privately and safely. An AI that quietly fed on that work would break the whole promise. So we built one that cannot.

Your music is yours. Our job is to help you organise it, share it, and get it heard. Never to mine it.