How music AI attribution actually works
The state of music AI today can be measured in the millions — millions of dollars raised, millions of users engaged, millions of AI-generated tracks uploaded to streaming services.
As the market continues to grow at breathtaking speed, an urgent business question has emerged: Who deserves credit and compensation when AI generates music? The answer could determine how billions in potential revenue flow through the music economy in the coming years.
Introducing attribution
Attribution — or the linking of AI outputs back to the specific training inputs that influenced them — underpins the ideal of a more granular rights holder compensation framework for AI.
Unlike copyright or deepfake detection systems that simply flag similarity or infringement, attribution aims to establish causal relationships between inputs and outputs. The concept overlaps with the fields of interpretability and explainability (XAI), which seek to understand how a model produces output for a given input.
For the music industry, attribution tech can help answer questions like:
- Which songs in the training data most influenced this output?
- How strong was each influence, and how should we quantify it?
- What specific musical elements were borrowed or transformed?
Several startups are tackling this complex problem with their own distinct approaches. Musical AI, Sureel, and Soundverse promise more granular, model-level tracking of influence, while others like LANDR, Source Audio, and Lemonaide offer simpler, pro-rata splits based on total dataset contributions. These ventures are also actively brokering a new generation of AI licensing deals between developers and rights holders, helping shape industry standards ahead of regulation.
In the sections ahead, we'll examine some of the technical underpinnings of attribution, the commercial frameworks being built around them, and the challenges in deploying these tools at scale. More than a mere technical curiosity, attribution will likely determine how — and for whom — the music industry thrives in an AI-centric future.
Watch our webinar
For a video walkthrough of attribution with visual references, we recommend watching the following recording of a webinar we presented to Water & Music members. Links to nearly all the papers we reference in the webinar are included throughout this written report.
But wait… what are we attributing?
While many compare music AI generation to sampling, the two processes operate fundamentally differently.
Traditional sampling takes identifiable fragments from existing recordings; think, for instance, of the iconic drum break from "Funky Drummer" appearing in hundreds of hip-hop tracks. In contrast, AI models absorb musical concepts at a more abstract level, learning patterns and concepts rather than merely copying sections. These learned patterns are captured across its multilayered architecture, forming an intricate weave of dependencies — similar to a modular synth, where each module represents its own layers with many parameters to tweak, each change cascading a complex network of changes. A single AI-generated track will draw from dozens of abstract layers of influence, which are challenging to map computationally and with 100% confidence.
Sonic characteristics
When examining any AI-generated song, multiple dimensions of musical influence come into play simultaneously:
- Composition. AI models learn fundamental musical structures like chord progressions, melodic patterns, and longer song forms. A model might learn the distinctive I-V-vi-IV progression that powered countless pop hits, without attributing it to a specific creator. This raises a critical question: When an AI recreates a common chord sequence, who (if anyone) deserves attribution?
- Recording. AI models also absorb production techniques, sonic textures, and mixing approaches from the recordings they train on. For instance, models exposed to lo-fi hip-hop beats often internalize techniques like vinyl crackle textures, relaxed drum compression, and intentionally reduced high-frequency content.
- Time. Unlike static media like images, music unfolds over time in ways that compound complexity. The introduction of a song influences later sections, creating causal chains that attribution systems must somehow trace. Moreover, different moments in a single AI-generated song might draw from entirely different sources in the training data.
Beyond audio: The full attribution picture
The attribution challenge extends far beyond audio files alone. A fully comprehensive view would consider:
- Data augmentation. AI developers routinely expand their training datasets by creating modified versions of existing data — e.g. altering pitch, tempo, or adding background noise for a single song, which can generate thousands of derivatives. This process exponentially complicates attribution: If a model trains on 50 variations of one song, how should that impact credit allocation?
- Prompt engineering. Text-to-music generators like Suno and Udio rely on written descriptions from users that shape musical output. While the U.S. Copyright Office has ruled that prompts alone typically don't warrant copyright protection, they undeniably influence the creative result. Should exceptionally detailed prompts receive attribution alongside training data?
- Multimodal influences. Newer AI systems like Tencent’s XMusic can generate music to accompany video, learning cross-modal relationships between visual emotion and sound. When visual data directly influences musical decisions about timing, dynamics, and emotional tone, attribution frameworks should arguably account for this additional layer of complexity.
- Model design. Different AI architectures process musical data in different ways, meaning choices about model design can significantly influence which training inputs have the strongest impact on final outputs.
The stakes: Attribution as value distribution
Every attribution system ultimately makes value judgments about which musical elements deserve compensation.
These judgments have substantial financial implications:
- A system that overvalues melodic similarity might disproportionately compensate songwriters over producers.
- Attribution focused on recording characteristics might favor master rights holders over composers.
- Models that track thousands of minor influences could fragment payments into economically meaningless amounts.
These decisions are both technical and ethical in nature, and will set precedents that could redefine music's economic structures for decades to come.
How do we attribute?
The gap between the music industry's ideal of attribution — a detailed royalty sheet for every AI-generated song — and technical reality remains substantial. Let's examine how some common attribution technologies actually work, and why each approach comes with significant tradeoffs that shape the commercial solutions emerging in the market.
Influence functions: The statistical ideal
A fundamental concept in statistics, influence functions measure how much a single data point affects a model's output, by comparing results with and without that point included (“leave-one-out”). The principle is straightforward: The more a prediction changes when a specific data point is removed, the more influential that point is considered to be.
As a practical example: Picture a music AI that generates a trap beat. Influence functions would theoretically measure how different that beat would be if Drake's "God's Plan" were removed from the training data, versus removing Kendrick Lamar's "HUMBLE." Whichever removal causes a greater change would be considered more influential to the output.
Despite their elegant simplicity, influence functions face several significant limitations. First, they tend to overemphasize outliers in the data, as unique examples can disproportionately affect the output. More critically, a "leave-one-out" approach becomes impractical with modern AI models: For a model trained on millions of songs, computing true influence would require retraining the model millions of times — a task that would take years even on advanced hardware.
To address these computational challenges, the explainable AI field has developed approximation techniques that attempt to capture the spirit of influence functions without the full computational burden. However, these approximations often break down when applied to neural networks — i.e. the foundation of modern generative AI. The result is that these approximated influence measurements no longer reliably indicate how the model would actually behave if specific training data were removed.
Embeddings: Similarity as a proxy
Whether we’re talking about Google’s search results, a photos app grouping images of your face, or Spotify’s song recommendations, a critical component of modern AI is the ability to mathematically quantify similarity. These similarity mapping systems rely on embeddings: Translations of raw, unstructured data (text, images, audio) into compact numerical coordinates in an abstract space, where similarity can be measured precisely.
For example, a song’s raw audio waveform, typically chaotic and computationally unwieldy, might be compressed via embedding into a vector like [0.24, -0.73, 1.2, ..., 0.45], a string of hundreds or even thousands of numbers. These vectors act as GPS coordinates in an abstract space — but instead of two dimensions (latitude/longitude), they use hundreds to capture nuanced relationships. In this space, proximity reflects perceived similarity: Two songs near each other might share tempo, mood, or genre traits a human would recognize. Music projects like Everynoise are exemplary in how they visualize similar genres as clusters in embedding space.
Crucially, however, proximity implies resemblance, not causation. While an AI-generated track near Taylor Swift’s songs might directly sample her work, it might also be echoing her style unintentionally, or coincidentally converging on similar patterns. A clone of a song’s "coordinates" might be theft — or pure algorithmic pareidolia.
For rights holders, this distinction is critical: Embeddings can flag potential influences (what sounds alike), but cannot interrogate the creative process (not why it sounds alike).
Watermarking: Digital receipts for data usage
Watermarking takes a different approach entirely, embedding hidden signals (e.g. inaudible tones, spectral tags, metadata) in training data that survive the generation process and appear in outputs. Unlike the other methods, watermarking doesn't try to measure degree of influence; it simply confirms whether specific training data contributed to an output at all.
This significantly limits watermarking’s scope in the context of music rights. For example, an AI trained on a watermarked pop song might learn its "catchy hook" structure but generate an original melody, leaving the watermark intact yet revealing nothing about this creative borrowing.
As we’ve covered previously, watermarking is also incredibly vulnerable to inaccuracy and manipulation. While detecting watermarks in AI outputs can confirm the tagged data was part of the training set, they often degrade during training or generation, leading to false negatives. In response, researchers have adopted convoluted workarounds, like applying the same signal to massive portions of a dataset. Imagine tagging every item in a warehouse with the same barcode; this brute-force approach might prove the dataset was used, but it’s useless for identifying specific tracks that influenced the output.
Adversaries can also exploit loopholes such as stripping watermarks via audio editing (e.g., EQ filtering), or training models on "clean," unmarked copies of the same content. In short, watermarking answers a limited question — and even then, not reliably.
Challenges in the details
Even the most sophisticated attribution technologies face inherent challenges rooted in how modern AI systems work:
- Architectural complexity. As detailed in the previous section, today's music AI systems transform raw inputs into songs through nonlinear conceptual stages, blending influences from thousands of training examples into fundamentally new outputs. Current attribution tools cannot reliably map how specific artists’ works contribute to these intermediate transformations, particularly when examining subsystems that operate at different abstraction levels. This practically means there is a long tail of music data, all responsible for the model learning the fundamental concepts for “what is music,” that might not be credited.
- Hidden dependencies. AI systems often integrate specialized subsystems trained for distinct tasks, including components designed to process music in highly specific ways. For instance, some models employ preprocessing modules that compress musical data into streamlined representations (e.g. a latent space), optimizing computational efficiency during generation. These subsystems require their own training data; a compression module, for example, might be trained on vast datasets of raw audio to learn how to distill music into a compact, analyzable format. Though invisible to end users, this foundational preprocessing step shapes every output the system produces. This raises a critical question of whether data used to train such subsystems should be assigned a different value than data used in later stages of the pipeline.
- Data debt and negative influence. AI models don’t discern from the data they are fed; they learn as much from negative (i.e. low-quality or conflicting) examples as positive ones. How do we credit artists if their work inadvertently undermines another’s creative intent? Further complicating accountability, recent research shows isolating these negative influences may actually help improve models. For instance, a poorly mixed track might teach a system what to avoid, paradoxically making it more influential on the final result by showing what not to do.
- Synthetic data. As companies increasingly use AI-generated data to train other AI models, attribution becomes a chain-of-custody problem. Each generation further obscures the original human sources, creating attribution chains that become increasingly difficult to trace.
These inherent constraints underscore why major players now face a dual mandate: Refining imperfect tools while inventing entirely new commercial frameworks to meet industry demands. The next section will examine how these players are building businesses despite — or perhaps because of — the technical constraints in front of them.
Who are the key players?
While the technical challenges of attribution remain substantial, several pioneering companies are developing commercial solutions that balance future ideals with current market realities. Their diverse approaches reveal not just different technical strategies, but fundamentally different visions for how the music AI economy should function.
Sureel
Founded by AI and robotics PhD Tamay Aykut, Sureel has emerged as one of the most technically sophisticated players in music AI attribution. They've secured multiple patents for their tech, and have inked partnerships with OpenPlay and Rostrum Pacific, (parent company of Rostrum Records), nearing 10 million registered assets under their belt.
Their key innovation is addressing what many consider the central problem in music rights: Distinguishing between compositional and recording contributions.
Most attribution systems treat music as a single asset, and therefore only examine recording-level similarities between outputs and training data, without capturing compositional relationships. The discrepancy occurs because training data often takes the form of recordings, as opposed to separated recording and respective compositional (e.g. MIDI) representations.
In practice, this can lead to a severe gap in financial outcomes between both sides. For instance, when an AI generates a track with the same melody as a training example, but with completely different instrumentation and production, many attribution systems would miss this connection entirely — potentially leaving songwriters unpaid while only compensating the master rights holders.
Sureel's technology attempts to separate influences on melodies and lyrics (publishing rights) from influences on production and performance (master rights), with the ambition to go even more granular in the future to the level of crediting specific performers of an original song. Their tech is designed to be integrated during model training, with one of their patents outlining a process to track how individual pieces of data influence the neural network throughout the training process.
Musical AI
Co-founded by Sean Power, Nicolas Gonzalez Thomas, and Matt Adell (former CEO of Beatport), Musical AI is taking a similar infrastructure-focused approach to Sureel, building attribution tech that can be integrated into other companies' AI platforms.
The company is also actively brokering the rightsholder and developer deals needed for their solution to scale. They have partnerships with the music AI generator Beatoven and distributors Symphonic and Kanjian, and announced a CA$2.1M (~US$1.5M) funding round in February 2025.
While specific technical details remain limited, a 2024 demo reviewed by the W&M team showed their technology producing "royalty sheets" that identify specific pieces of IP and their percentage influence on generated outputs. Like other leading solutions, their system must be present from the start of model training to be fully effective.
Due to the complexity of music rights, Musical AI’s current strategy is to limit their IP partners to “one-stop rights holders” — i.e. those who own both their masters and publishing. This presumably narrows down their serviceable market to production libraries like Epidemic Sound and Artlist that own all the rights to their catalogs, or to DIY artists with no label or publisher signings.
Under the Musical AI x Beatoven deal, 30% of revenue from Beatoven’s upcoming generative models will be distributed back to the artists and rights holders who provided training data (aligning with collecting society GEMA’s recommendations).
Soundverse
Soundverse — founded by Sourabh Pateriya (who previously co-invented Spotify’s audio-to-MIDI converter Basic Pitch) — offers a suite of AI tools geared towards content creators and hobbyist musicians, including stem separation, lyric generation, text-to-audio synthesis, and full track creation. Linnea Sundberg, Head of Operations at UnitedMasters and former business exec at Splice and Spotify, joined Soundverse as an advisor in February 2024.
In May 2024, Soundverse launched its Content Partner Program, which allows artists and rights holders to submit their content for inclusion in training the company’s own AI models, in exchange for receiving royalty payouts based on how their data is used.
Under this program, Soundverse has evolved its attribution approach significantly through trial and error. They initially analyzed outputs for similarity across multiple dimensions (voice similarity, song structure, lyrics, instrumentation, key), but found this approach difficult to implement reliably.
Their breakthrough came with integrating attribution directly into their own foundational model. "Attribution accuracy is very much dependent on the architecture, and it's highest when it's infused with the model itself,” Pateriya told us.
Now, rather than examining outputs after generation, Soundverse now tracks how tokens flow through the model during training and generation. This allows them to monitor which parts of the neural network are activated by specific training examples, and therefore create a more accurate connection between training inputs and generated outputs. The company is still exploring exactly where in the model pipeline to implement attribution — e.g. tracking token flow end-to-end versus focusing only on final token generation.
On the other side, rights holders can access a partner portal to monitor how their data influences generated content, with attribution mapped to specific royalty payments.
Notably, Soundverse is one of the few companies today combining attribution tech with their own consumer-facing AI music platform. This vertical integration gives them a significant advantage: They can tune their attribution system specifically for their own models, rather than feel pressured to build a one-size-fits-all solution.
Simplified attribution alternatives
Not all companies are pursuing technical attribution at the individual song level. Several are opting for less tech-dependent deals with rights holders, through simpler revenue-sharing and pro-rata payout models.
Lemonaide
Founded by Michael “MJ” Jacobs and Anirudh Mani, Lemonaide launched in 2021 as a browser-based MIDI generator, and has since expanded its product line to include Spawn (DAW plugin for MIDI generation) and Collab Club (producer-branded AI melody generator, partnering with the likes of Kato on the Track and KXVI).
With Lemonaide’s Collab Club models, attribution is straightforward: Each AI model is trained exclusively on a single producer's data, with 100% of attribution flowing back to that producer. Users agree to share 5% of revenue with the producer for tracks exceeding one million streams.
Recognizing this single-source approach won't scale to larger datasets, Lemonaide is developing what they call "cohort-based attribution," which categorizes outputs by genres and sub-genres (ranging from hip-hop to Nepalese folk music), determines which cohort most influenced a given output, and distributes revenue to all rightsholders who contributed to that cohort. Importantly, this attribution still relies on detecting output similarity — probabilistically attributing an output back to an input cohort, rather than gaining insight into the internal mechanisms of the models themselves.
“The idea isn’t to enforce traditional genre definitions, but rather to create meaningful ‘groupings’ that help us trace influence in a way that aligns with musical patterns, musical intention, and other production related factors,” Mani told us over email. “Genres and subgenres are also a great starting point because artists and rights-holders still think in genre terms when using tools like ours. I anticipate we’ll move towards more meaningful groupings organically as we keep working and improving.”
While Lemonaide is investing in more granular "fractional attribution" for the future, they are prioritizing building trust with rights holders over achieving near-term technical perfection. As Mani told us: "Attribution is as much a business, ethics, and trust problem as it is a technical problem.” In other words, a technically sound attribution system is worthless if industry rightsholders won't adopt it.
LANDR
One of the earliest AI-native music tech companies, LANDR has evolved from an automated mastering tool to a full-service platform with distribution, sample libraries, and plugins. With over 6 million registered users and tens of millions of mastered tracks, they have substantial built-in market reach for their current and future AI tools.
Their Fair Trade AI program allows artists who distribute through LANDR to opt into providing their music as training data. Rather than attempting granular attribution, LANDR will distribute 20% of revenue from AI products back to contributing artists on a pro-rata basis, determined by the volume of songs contributed to the dataset.
This approach sidesteps the technical challenges of precise influence tracking, providing rights holders with a much more predictable revenue model. The simplicity of this system may appeal to many independent artists who value upfront transparency and predictability over potentially higher but less certain returns from more complex attribution systems.
Source Audio
Music catalog management and sync licensing platform Source Audio is expanding into AI tools with their new platform SongLab, which officially launched in February 2025. Their substantial dataset includes 33 million tracks (14 million full mixes plus 19 million alternate versions and stems), positioning them as one of the largest rights-cleared music repositories available for AI development.
SongLab offers a comprehensive toolkit including lyric translation, catalog augmentation (creating derivatives like remixes), stem separation, and audio similarity search. Underlying these tools is a distinctive approach to rights management that focuses on specificity rather than flexibility of use.
Drew Silverstein, Head of AI Strategy at Source Audio, explained their licensing framework over email: "Our agreements are limited to specific, pre-defined use cases with strict guardrails preventing copying, sub-licensing, or replacing human-created content with synthetic data. The multi-year, fixed-term structure provides clarity and stability to all parties rather than open-ended commercial exploitation."
Unlike systems that attribute ongoing royalties based on usage, Source Audio's agreements build compensation directly into the upfront terms. "We license access to data for training and improvement purposes within carefully defined boundaries, which exclude the right to commercially reproduce or distribute the original works," Silverstein noted. "The value exchange is built into the upfront terms of the license, and additional royalty mechanisms are not aligned with how the data is actually used."
Human Native
Founded by Jack Galilee and James Smith (formerly of Google's DeepMind), Human Native AI is positioning itself as a broker of licensing deals between AI developers and rights holders rather than a technology provider. The company has won the Music Ally S:IX startup competition, and is a member of the BPI’s latest Grow Music innovation program.
Notably, Human Native has deliberately chosen not to pursue attribution tech at all. Instead, they focus on facilitating dataset licensing deals with fixed terms and upfront fees. Their marketplace model aims to standardize these transactions, giving rights holders access to multiple AI partners without negotiating individual deals with each. While they're experimenting with revenue-sharing models, none depend on attribution to individual pieces of content.
What does this mean for the music industry?
After examining the sheer diversity of attribution and licensing approaches in the music AI market, several critical insights emerge for industry stakeholders:
Technical limitations shape deal structures
The attribution deals being struck today directly reflect each company's technical capabilities, rather than ideal market structures. Those without granular attribution simply cannot offer usage-based royalty models similar to traditional music licensing. Instead, they're creating alternative compensation frameworks built around flat-fee licensing or pro-rata data contribution volume.
These simpler approaches may appear less "fair" theoretically, but they offer crucial advantages:
- Predictability — Rights holders can forecast revenue with greater confidence.
- Transparency — The terms are more easily understood by all parties.
- Scalability — These models don't require complex technical infrastructure.
The most critical infrastructure: Trust
Several attribution companies are acting as deal brokers, suggesting that technical solutions alone don't solve the market coordination problem. This intermediary role is proving essential because:
- Rights holders need assurance about how their IP will be used.
- AI developers need confidence in the legitimacy of their training data.
- Both sides benefit from standardized deal terms and compliance verification.
Many of the attribution companies gaining early traction are those that prioritize trust-building alongside technological development.
Strategic advantages for unified rights
Entities controlling both master and publishing rights have substantial advantages in the AI training market. This explains why early AI licensing deals are concentrated among:
- Production music libraries (e.g. Source Audio)
- Digital distributors with comprehensive rights (e.g., Symphonic)
- Artist-owned catalogs (e.g. Soundverse)
These "one-stop shops" can move quickly without navigating complex approval chains. For major labels and publishers with fragmented rights, the path to AI monetization will likely require more sophisticated attribution that can track and allocate revenue across multiple rights holders.
Reframing the purpose of attribution
The perhaps uncomfortable truth underpinning all of these findings is that perfect attribution for music AI doesn't currently exist.
All the attribution approaches we studied provide approximations at best, rather than definitive proof of influence. The options for rights holders and AI developers today essentially boil down to:
- Pursuing highly technical attribution methods that remain cumbersome to implement and audit, or
- Adopting simpler solutions that compromise on detail, but offer more practical viability.
Ultimately, real progress hinges on confronting these limitations. Rights holders must ask what "fair compensation" means in an environment where every track contains countless micro-influences. Meanwhile, AI companies must design systems transparent enough to build trust, while being candid about where the tech still falls short.
At the same time, rather than viewing attribution as merely a technical problem waiting to be solved, the industry might also benefit from reconsidering its fundamental purpose.
Attribution itself isn’t even necessarily the end goal; rather, it's one possible mechanism to ensure rightsholders are fairly compensated when their work contributes to AI systems.
Understanding this motivation helps explain why the companies featured in this piece define attribution in such different ways. Some are building fine-grained, forensic systems that aim to trace exactly what happened inside a model. Others, limited by current tech, rely on approaches like output analysis or similarity scoring — methods that echo traditional musicology in infringement cases, but do not give primary source sight into the models’ workings.
Interestingly, the best technology does not ensure "winning" in this space; we already see companies able to gain trust and coordinate deals by balancing commercial viability with practical constraints.
Hence, while there is a gap between what the market expects from attribution and what the technology can actually deliver, if the goal is compensating rights holders, then attribution is just one tool in the box, and the right mix of incentives and innovation will eventually close that gap.
The companies most likely to succeed won't necessarily be those with the most technically sophisticated attribution systems, but rather those who can solve the immense political challenge of fostering cross-industry trust, and creating compensation frameworks that all parties view as both practical and fair.
Supplemental resources
Training data influence analysis and estimation: a survey
Tracing Model Outputs to the Training Data
Studying Large Language Model Generalization with Influence Functions
Is There A Coherent Theory of Attributing AI Training Data?
MusicLIME: Explainable Multimodal Music Understanding
AudioGenX: Explainability on Text-to-Audio Generative Models
Towards Assessing Data Replication in Music Generation with Music Similarity Metrics on Raw Audio
God Help Us, Let's Try To Understand AI Monosemanticity
Watermarking Training Data of Music Generation Models
XAttnMark: Learning Robust Audio Watermarking with Cross-Attention