Musical AI raises $4.5M to trace which songs shaped an AI output. What can attribution prove?

Musical AI announced a $4.5 million round on 13 Jan 2026, led by Heavybit with BDC and Build Ventures. The company, formerly Somms.ai, sells attribution: software that says which pieces of licensed music a generated track drew on, and in what proportion. If AI music is ever going to pay artists per output rather than through a one-off licence fee, something like this has to work. The question is what "work" means.

What did Musical AI raise, and from whom?

The round is $4.5 million. The release says the money goes on hiring, developing the attribution technology and supporting industry deals. Music Business Worldwide adds that the company plans to grow its music partnerships before moving into other creative sectors, and that it launched its rights platform in 2024.

The named rights partners are catalogue and distribution businesses rather than labels: Pro Sound Effects, SourceAudio and Symphonic Distribution. The release also names SoundBreak AI, an AI music company that it says trained its models on licensed works through Musical AI. SoundBreak's chief executive, Kevin Griffin, is also the singer of Better Than Ezra. In the release he says SoundBreak's business is built on licensed data with attribution. Musical AI is also listed by Fairly Trained, the non-profit that certifies AI companies that license their training data.

The most concrete earlier deal is with Beatoven.ai. In December 2024 the two companies announced a white-label song generator trained on more than 3 million songs, loops, samples and sounds. Musical AI was to handle licensing, attribution and payouts, with rightsholders paid a revenue share based on how outputs are used.

How does the attribution actually work?

Musical AI's own pages are more specific than the funding release. Its attribution page says the system works "at the output boundary". The AI company sends a generated output, or a representation of it, after generation, and Musical AI returns a record. It doesn't inspect model weights or gradients, doesn't train models, and doesn't decide infringement. Each record splits the output across the licensed sources that influenced it, master and publishing are attributed separately, and, in the company's words, the total "always sums to 100%".

The home page says pricing is per attribution event, so cost rises with the number of generations rather than with catalogue size. Neither page publishes accuracy figures, benchmarks or a description of the method. The company calls its records "not a probabilistic guess" but doesn't show how it backs that up.

What does the research say attribution can prove?

Less than the marketing implies, though more than nothing. In an October 2025 paper, Sony AI researchers split existing methods into two families. Similarity methods assume the training items closest to an output were the ones that shaped it, and the authors point out that similarity doesn't imply causation. Methods that look inside the model only give probabilistic estimates. Their conclusion is that there can be no certainty a specific training item was used for a specific generation, and they call exact training-time attribution intractable. They also concede that uncertain estimates may still be a fairer basis for paying people than a flat rate.

A separate royalty-model paper by Deng and colleagues, first posted in December 2023 and revised in December 2025, finds that perceived similarity, the usual legal proxy, picks up the most influential training samples but misses broader contributions. On that evidence it may be a poor basis for splitting royalties. Other researchers are testing heavier methods: a NeurIPS 2025 Creative AI track paper applies "unlearning" to a text-to-music diffusion model to estimate which training data contributed most to a given output. Deng and colleagues note why such shortcuts exist: exact causal attribution would mean costly retraining.

Put next to that work, an output-side system without access to the model looks like it is measuring resemblance between the output and a licensed catalogue. That is our reading, not something the company states. Its pages don't say which method it uses, so we can't confirm it.

What it means for artists

For an artist licensing a catalogue to an AI company, attribution decides whether you get a share of each output that sounds like your work or a cut of a flat pool. A system like this can give you per-output records, separate master and publishing lines, and an audit trail to dispute. Those are real improvements on a lump-sum training licence that never reports usage.

Before you sign, ask four things: what is being measured (similarity or influence), what error rate the vendor will put in writing, who can audit the records, and what happens to the share when the output resembles nothing in the licensed pool.

What we think

The detail that matters is "always sums to 100%". A split that must total 100% across the licensed material answers one question: how to divide the money among the people already in the deal. It can't, by design, report that an output owes most of its character to music outside the pool, such as an unlicensed catalogue the model also learned from, or the wider style of a scene. For a closed, fully licensed model like the Beatoven arrangement, that may be fine. Attached to a model trained on data no one has disclosed, the same report would make it look as if every output was fully accounted for. AI companies would have every reason to like that, and artists should notice it.

The pricing points the same way. Charging per attribution event means Musical AI earns more as AI output grows, which is a reasonable business but ties the referee's income to the volume of the thing it measures.

The likely winners are mid-sized production libraries and distributors, the kind already on the partner list, which can turn catalogues into recurring AI income without a major label's legal team. Session players and writers whose influence is diffuse rather than recognisable are the likely losers under any similarity-based split. The Deng paper suggests that is exactly the contribution similarity misses.

Attribution is worth building, and $4.5 million is a small, sensible bet on it. But until a vendor publishes its method and error rates, treat an attribution percentage as a negotiated estimate, not a measurement, and write that into the contract.