Moises separates what it will not train on from what it does

The system

Moises is a mobile and web application built around stem separation, pitch and key detection, practice tools, and a generative “AI Studio” that can produce new instrumental or vocal parts from a text prompt. As retrieved on 16 September 2026, Moises' help centre publishes specific answers to two separate questions a working musician is likely to ask: who owns the audio a user uploads or produces, and what data trained the generative models that produce new material. Moises Systems, Inc. (d/b/a Music AI) also publishes a governing Terms of Service that binds both questions together.

What the documents establish

The help-centre article “Who owns the output and music I create with Moises?”, last updated 27 November 2025, states plainly: “You do,” and that Moises does “not claim ownership” of uploaded audio or generated outputs, including commercial use, “as long as you have the necessary rights in all of the materials you incorporate.” The Terms of Service, last updated 27 August 2026, adds a narrower and separate commitment in Section 11.4: “WE DO NOT USE YOUR USER CONTENT OR OUTPUT TO TRAIN AI MODELS WITHOUT YOUR PERMISSION,” a restriction that applies unless a user opts in in writing or electronically. A third article, “What are the AI Studio models trained on?”, last updated 22 August 2025, states the generative models are trained on “a diverse foundation of fully licensed, high-quality data” obtained by working with “leading content providers, along with independent labels, artists, and our own internal production team.”

Craft and rights

These are two different promises and a producer should not collapse them. The no-training clause covers a specific user's own uploads and outputs; it says nothing about how Moises built the underlying AI Studio models, which the separate training article addresses by pointing to licensed sources rather than scraped catalogues. Both statements shift the remaining rights question onto the user: Moises disclaims ownership of output but conditions commercial use on the user actually holding rights to every incorporated element, and says this is not legal advice.

Outcomes and open questions

None of these pages name the specific labels or agreements behind “fully licensed” training data, so the claim cannot be checked against a public rights ledger. As living documents, all three carry update timestamps rather than fixed publication dates and can change. Watch for whether Moises publishes more detail on its licensed-data partners, and how the opt-in for training on a user's own content is actually presented in the product.

• Has this specific Moises policy page been updated since 16 September 2026, and does it still say the same thing?
• Does a given project incorporate any uploaded material the user does not hold rights to, despite Moises disclaiming output ownership?
• What would an explicit, written opt-in to model training actually look like inside the app?

Moises' own pages draw a reasonably clear line between not training on your material and training its generative features on material it says it licensed elsewhere, and that line, not a single blanket promise, is what a rights-conscious user should carry into the app.