How to Measure One Song's Contribution to an AI Model

Executive Summary

AI licensing deals across the creative industries have already reached nearly 300. In music alone, Suno, Udio, Sony Music, and Universal Music have signed contracts. Yet there is still no yardstick for the question that matters most to a rights holder: "How much did my song contribute to this model?" The deals are piling up, but the basis for splitting the money is missing.

Today's payouts mostly divide by prompt count or number of generations. That approach allocates shares regardless of how much a song actually contributed to the model's performance. The core problem is not copyright but measurement. Can we convert into a value how much a single training sample contributed to an output? That is the question of data attribution — a problem Pebblous readers already know well.

This piece lays out what each of three measurement technologies — influence functions, embeddings, and watermarking — actually measures and where each hits a wall, and how real services and recent research route around those limits. Prompt count measures "how much was used." A fair payout needs to measure "how much changed."

~300

Creative-industry AI deals

BPI · WPI Economics 2026 tally

26

Key deals in music

Suno · Udio · Sony · Universal, etc.

5%

Small labels participating

Deals skew toward the majors

0

Standard contribution metrics

No recognized basis for fair payout

1

300 Deals, No Yardstick

According to a 2026 report from the British Phonographic Industry (BPI) and WPI Economics, 274 commercial AI agreements have been struck across the creative industries, a figure the press rounds to "nearly 300." Narrow it to music alone and there are 26 key deals involving names like Suno, Udio, Klay Vision, Stability AI, ElevenLabs, Spotify, and Vermillio. Sony Music has partnered with Vermillio; Universal Music with Prorata.ai.

The numbers are climbing fast, but participation is uneven. In the same report, only 25% of mid-sized labels and 5% of small labels had completed licensing agreements for AI products. With deals concentrated among the large rights holders, the creators with the least bargaining power are precisely the ones who can least verify on what terms, and for how much, their songs were sold.

The deeper gap is not the number of contracts but the basis for them. An open letter led by the European Music Managers Alliance (EMMA) and joined by 31 music organizations took issue with clauses buried in existing contracts that fold works into AI use automatically "unless you opt out." The American Federation of Musicians (AFM) has sued to claim a share of the major labels' Suno and Udio licensing revenue. Before anyone has settled what to divide, or on what basis, the contracts keep stacking up.

The industry has given this gap a name: Creative Weight Attribution. Proposed by researchers at Fraunhofer alongside independent rights-management bodies, the idea is to measure not surface-level usage such as prompt count, but the "creative weight" a work carries within a model's training and generation. The direction is clear; what remains is the technology. Music-recognition tools still cannot reliably tell what came from which song. So the next question naturally splits in two: what exactly do today's payouts measure, and what technology would it take to measure contribution properly?

2

What Prompt-Counting Payouts Miss

Current practice comes in two broad forms: pro-rata splits proportional to the number of generations or prompts, and flat fees paid regardless of usage. Both are simple to compute. But a simple calculation is not a fair one.

First, prompt count is unrelated to how much a song actually contributed to the model's performance. Some songs are decisive in teaching a genre's harmonic feel; others are buried among tens of thousands. Prompt-based payouts treat the two identically. They count how many times a user made a request, never asking whose work the result came from.

Second, composition and master (recording) rights get tangled together. Most systems treat the recording as a single asset and check only for recording-level similarity. When an AI generates the same melody in a different arrangement, the composer goes unrecognized while only the master rights holder is paid. That is a structural flaw. Because there is a shortage of data that separates melody from production, the problem does not resolve easily.

The real problem with prompt-counting is not imprecision but direction. It measures consumption (how much was used). What creators need is a measure of contribution (how much was changed). Because the target of measurement is wrong, no amount of precise counting will converge on fairness.

▲ Original Pebblous diagram — usage-based (prompt-count) payout vs. contribution-based payout

3

Three Ways to Measure Contribution, Three Limits

The technologies that actually try to measure "how much one song contributed to a model" fall into three families. Each is clear about what it measures, and equally clear about the wall it hits in practice. Following the framing of the music-industry analysis outlet Water & Music, the three approaches compare as follows.

Technology	What it measures	Where it breaks down
Influence functions	How much the output changes when a specific training example is removed	Retraining cost is impractical at a scale of millions of songs; neural approximations are unreliable
Embedding similarity	Turns songs into vectors and computes proximity to the output	Proximity suggests similarity but is not causation
Watermarking	Embeds hidden signals in training data and traces them in the output	Signals can be lost during generation or deliberately removed

Influence functions have the most direct logic. Remove this song, retrain, and see how much the result changes — that shift is the song's share. The problem is scale. Repeating retraining for every single song is unaffordable for a large music model, and the neural methods that approximate it are not yet accurate enough to trust.

Embeddings are fast and scalable, but while they show what resembles what, they do not tell you why. Watermarking requires implanting a signal in advance, and that signal can be erased during generation — so it does not apply to unlicensed data in the first place. Water & Music's conclusion is sober: perfect attribution does not yet exist. It also notes that deciding which attributes to reward is itself a value judgment. Overweight melody and you bias toward composers, and that choice ends up defining the structure of the music economy for a long time.

4

The Market Didn't Wait for Perfect

The absence of perfect measurement has not frozen the market. Several companies have already shipped their own workarounds as services. Sureel uses simulation-based technology that tracks composition rights and master rights separately. Musical AI produces a royalty sheet after generation, computing which IP influenced the result and in what proportion. Soundverse tracks not output similarity but the path tokens travel through training and generation itself — though only inside its own model. LANDR and Lemonaide choose practicality over precision, distributing in proportion to the amount of data contributed.

In academia, one line of research has lifted the problem up to the level of contract design. "What's a Credit Worth?" (Zhang et al.), posted to arXiv in July 2026, binds data-contribution estimation and payment design into one. Its core device is the top-k reward: for each validation data point, a fixed reward goes only to the top k most influential training examples.

Why only the top k? Because of the randomness inherent in training, contribution estimates cannot be trusted equally for every data point — and estimates for small contributors are especially noisy. Top-k trims that unstable tail out of the reward calculation, preventing measurement error from working against small creators. The paper also presents, as a theorem, that a reward must be a monotonically increasing function of contribution to preserve incentive compatibility, giving theoretical backing to the critique that simple counting fails to distinguish the quality of a contribution.

▲ Original Pebblous diagram — for one validation point, only the top-k most influential examples get a fixed reward; the rest are withheld

Real services and recent research share one attitude: a design that admits what it does not know. Rather than forcibly assigning a value to unmeasurable contributions, they reward only the reliably top contributions, or withhold the uncertain shares. It is not perfect, but its direction sits closer to fairness than prompt-counting does.

5

Same Math, Different Stage

The problem of converting into a value how much one piece of data contributed to a result is not new to Pebblous readers. In the technology of putting a price tag on data, we covered the wide gap between the data-broker market and pure marketplaces, and the value-proof techniques and Data Shapley that bridge it. In how to price synthetic data, we examined a pipeline that converts quality scores into contribution and automates all the way to compensation.

Measuring contribution in music uses the same math as these problems. Shapley value's marginal contribution, influence functions, and incentive-compatible reward design are the same skeleton whether the input is an industrial dataset or a creative work. What differs is the stage.

Industrial datasets mostly deal with anonymous samples. In music, that sample is one person's labor. It has a name, and its rights split between composition and performance. Here, "we can't measure the share" is not merely a payout error. It extends into a question about the dignity of creative labor: someone's work is absorbed into a model without being recognized as having value. That is why the same measurement problem grows harder.

▲ Original Pebblous diagram — anonymous samples and named creative works feed the same contribution-measurement engine

6

What to Measure for Fairness

To sum up: prompt count measures usage. What a fair payout needs is contribution — how much the model's output changed depending on whether a given song was there or not. With no perfect causal tracing available yet, designs that admit what they do not know, such as top-k and uncertainty-aware rewards, come closest to fairness, at least for now.

The minimum conditions owed to creators follow from this. First, what is being measured must be disclosed, because which attributes carry value is exactly what decides who gets paid more. Second, contributions that cannot be measured should be withheld rather than scored low. The moment you shave a small contributor with a noisy estimate, the calculation may look precise but the result is unfair.

What 300 deals tell us is that the market has already opened. What has not been settled is the basis on which that market will split the shares. That basis will be decided in measurement before it is decided in court. The technology of putting a price tag on data is now being tested on the coordinates of the dignity of creative labor.

Editor's Note

Pebblous has explored the problem of measuring how much one piece of data contributes to a result — and linking that to compensation — in data value proof and synthetic-data contribution scoring. Music licensing shows what that contribution-measurement problem looks like when its subject is human creation.

Pebblous Data Communication Team
July 17, 2026

R

References