Musical AI is a Toronto startup that sits between music rights holders and generative AI companies. Rights holders license catalogues to it. It supplies that music to AI developers as training data. Then it uses attribution technology to estimate how much each licensed track influenced a model’s output, and pays rights holders according to that influence, not as a single flat fee. In November 2024 the round was reported at about $2.1 million. The company was not offering a blockchain ledger that “tracks every note,” which is how this deal was often described. It was offering a licensing intermediary with an influence-based way to split payments. Whether that model takes hold depends less on the software than on the lawsuits and licensing deals shaping AI music, which have changed a lot since 2024.
What does “attribution” mean for AI-generated music?
Attribution here means training-data attribution: estimating which items in a training set contributed to a particular output, and by how much. It is a statistical estimate, not a record of where each sound came from. Researchers have shown that it can be measured for image models (Wang et al., “Evaluating Data Attribution for Text-to-Image Models,” ICCV 2023), and the same idea is being applied to audio.
The weakness is checking the results. No independent public benchmark tests music attribution accuracy. Some coverage in 2024 quoted figures such as “95% attribution efficiency” or “12 major label partners,” but nothing published supports them. Treat any precise accuracy claim from a vendor as marketing until a third party has audited it.
Who owns music made with generative AI?
In the US, nobody owns copyright in material a machine generated without meaningful human creative control. The US Copyright Office’s Copyright and Artificial Intelligence report, Part 2 (January 2025) says prompts alone are not enough. Human-written lyrics, arrangement, editing and selection can still be protected. In March 2025 the D.C. Circuit upheld the human-authorship requirement in Thaler v. Perlmutter.
Generator terms of service decide who holds commercial rights to whatever protection does exist. For most of 2024–25, the free tiers of the major music generators kept ownership with the platform and allowed only non-commercial use. Paid tiers assigned output to the user. Check the current terms before every project, because they have changed several times.
Is training AI on copyrighted music legal?
The answer depends on where you are, and courts are still deciding it.
- United States: It depends on fair use. The Copyright Office’s Part 3 report on generative AI training (pre-publication, May 2025) concluded that some training uses that compete with the original works are unlikely to qualify. In June 2024 the major labels sued Suno and Udio. Universal and Warner have since settled with Udio, and Warner with Suno, in exchange for licensed models launching from 2026. Those settlements are the market moving toward the licensing approach Musical AI sells.
- European Union: Text-and-data mining is permitted unless the rights holder opts out (Directive 2019/790, Article 4). Since August 2025, providers of general-purpose AI models must respect those opt-outs and publish a summary of their training content (AI Act, Article 53). In November 2025 a Munich court ruled for the German collecting society GEMA against OpenAI over song lyrics reproduced from training data.
How is AI involvement disclosed and tracked?
There are two separate tools, and neither one handles payment:
| Mechanism | What it does | Limit |
|---|---|---|
| C2PA Content Credentials | Attaches signed provenance data to a file showing which tools created or edited it | Stripped by most re-encodes and many upload pipelines |
| DDEX metadata standards | The industry format for delivering releases to services. It is gaining fields for AI disclosure, and Spotify adopted these in 2025. | Self-reported by whoever delivers the release |
| Training-data attribution (e.g. Musical AI) | Estimates which licensed tracks influenced an output, as a basis for paying rights holders | Works only inside models trained on licensed, participating catalogues |
What this means if you commission or produce video
We license music for client video every week. AI music has not removed our licensing problems. It has moved them. Here is what actually goes wrong:
- Content ID claims. AI tracks that sound close to existing catalogue, or that someone else uploaded to a distributor first, get claimed on YouTube. A paid campaign can lose its monetisation or be blocked mid-flight.
- No indemnity. A $30–$300 stock licence or a custom score (about $1,500 to well over $15,000, depending on length, usage and territory) comes with clear rights, and often an indemnity. Most generator terms put the infringement risk on you.
- Client contracts. Brand and broadcast agreements increasingly require a warranty that no AI-generated audio was used, or that any AI use is disclosed. Agree on that before the edit starts, not at delivery.
- Weak protection. If the track is purely generated, you probably cannot stop a competitor using the same music. For a brand sonic identity, that alone rules it out.
Our working rule is to use AI music for temp tracks, pitches and internal cuts. For anything paid, broadcast or tied to a brand identity, we use licensed library music or a commissioned composer. When AI-assisted audio does ship, we keep the prompts, the generator tier, the terms in force on the date of generation, and every human edit, so the chain of rights can be documented if someone challenges it.
What changed since this deal was announced
In November 2024 there were no major-label licensing deals with AI music generators, no Copyright Office guidance on training, and no EU transparency obligations in force. All three now exist. The direction is toward licensed training data with paid attribution, which is the model Musical AI was funded to build. But that approach covers only participating catalogues, and its accuracy still has not been independently measured.
If you’re deciding on music for a project now, start with the intended use. For paid media, broadcast or brand identity, license human-authored music with an indemnity. Where you do use a generator, choose one that publishes its training-data licences, and keep generation records alongside the project files.
