What does an attribution engine actually do?

MBW Views is a series of op-eds from eminent music industry people… with something to say. The following MBW op/ed comes from Monica Corton, music publishing veteran and founder of Go To Eleven Entertainment. Below, she returns to AI — and to the industry’s one great hope for deciding who actually gets paid when machines make music…

An attribution engine for music takes an AI-generated output and traces it back through the model’s training data, identifying the compositions, source recordings, and styles that shaped it. It then assigns a relative weight to each song and track that influenced the output.

The goal is to turn that map of influence into something the music industry has never had before: a mechanism for compensation that follows the creative debt back to its source. That makes the choice of attribution engine one of the most controversial – and most important – questions in any licensing model for gen AI music.

Billions of dollars are on the line with gen AI music platforms, but unless we can find a fair and equitable attribution engine, that money will not reach the correct creators. Every music licensor wants accurate attribution, because it will form the basis of who gets compensated for each output. One of the biggest tensions for rights holders is the gap between the desire for precision in attribution and the complexity of measuring how any given song or track has actually influenced an output.

Why Metadata Isn’t Enough

Simplistic methods like genre tagging, metadata matching, or embedding similarity don’t just fail; they actively distort how creative influence is measured. Small uses of a famous guitar riff, or repetitive rhythmic or melodic use, often go undetected when they appear in a different genre from the original.

Compositional influence isn’t confined by genre; it can be deeply interpolated into a track, making it hard to detect. Basic similarity methods reward quantity over quality, erasing the true impact of standout works that, even when they appear infrequently in an output, still carry an outsized influence on the new work.

The influence of one work on another is not a black-and-white issue. The relationship between a song’s emotional weight, its cultural impact at release or re-release, and the connections that people make with songs does not fit neatly into algorithms. Creativity is often a wild card with unequal influences, whereas AI systems treat every song as raw data, constantly seeking the homogeneous middle.

Attribution sounds intuitive, but once examined in detail, it becomes technically and philosophically difficult. Is compensation based on similarity? And if so, similarity of what — harmony, melody, lyric, rhythm? Causality matters too: a song can be statistically present in training data without meaningfully shaping an output, and a song can deeply shape an output without appearing frequently in the data at all, if it contains a highly recognized element (like the cowbell in Blurred Lines). Then there is perceptual resemblance, which operates on an entirely different axis and determines what a listener recognizes, regardless of what the math says. All three measurements are required, but how should we weight them against one another? The balance between human perception and machine-level causality is delicate, and we have never had to reconcile them at scale before.

What The Math Misses

Similarity is mathematical. We take representations of both the generated audio and the catalog track and measure the distance between them, extracting audio features such as rhythm, pitch, timbre, and chord progressions. It’s precise, repeatable, and objective — but it doesn’t always reflect what a person hears. Perceptual resemblance, by contrast, is about the listener’s experience. It is driven primarily by timbre, rhythm, tempo, and genre, rather than technical or harmonic structure alone.

“The balance between human perception and machine-level causality is delicate, and we have never had to reconcile them at scale before.”

Two tracks can be mathematically close but sound nothing alike. They can share the same chord progression and tempo, yet one might be a jazz trio and the other a pop song — and in those different arrangements, they can sound like entirely different songs.

The reverse can happen as well. An AI output can capture the feel of an artist’s vocal delivery or production style in a way that’s immediately recognizable to a listener, even when the surface features of the music differ. Because the vocal style is so distinctive, listeners will associate the output with that artist — even if the underlying song bears no resemblance to anything that artist has previously recorded.

Determining attribution, then, is neither easy nor static. For that reason, whatever attribution engines are ultimately used by AI platforms, the entities responsible for running them should have no connection to any given AI platform and no connection to any given rights holder. Attribution cannot be a political decision, and transparency is critical if all parties are to trust the system.

The Independence Problem

Self-regulated attribution — say, by the AI platform itself — feels like assigning the most important task to the parties that never prioritized attribution in the first place. None of the large AI platforms have addressed attribution engine development in any detail in their public communications. These same platforms have offered little transparency around their training data practices, including how metadata was handled during the ingestion of songs and tracks from the internet — practices that make it difficult to determine what was used in the original training data.

And if there is no harmonization around how we weigh the different factors in attribution — similarity, causality, perceptual resemblance — how will music publishers, labels, and self-released artists ever feel confident that the songs and tracks credited to any given output are correctly identified, both by influence and by nature of use?

It is essential that we identify and elevate the people and organizations working on sophisticated, cross-referenced attribution systems. Harmonizing standards, aligning incentives, and writing rules that give these systems actual force are still up for debate.

For music industry professionals, lawyers, and technologists, that is the work to focus on now. Understanding what these systems can and cannot do, being part of the discussions that will set the standards, and demanding frameworks with real legal teeth are critical if we are going to have gen AI music platforms that equitably compensate human creators.