Last year, Suno and Udio released products that generate songs based on user prompts. Both services are based on large language models (LLMs), neural networks that are good at synthesising all kinds of data when trained on large enough datasets. To build these models, Suno and Udio likely would have needed millions of high-quality examples: audio, lyrics and metadata for real music composed by real artists. The launch of these products thus raised red flags for those in the music industry: what data did they use? If it was copyrighted, was this use of the data permissible? And if the models were somehow using bits of copyrighted music to build new pieces of music, was this plagiarism?

The smoking gun

After Suno’s splashy debut, boosted by a Rolling Stone profile in March 2024, many wondered about the source of their data, which was not disclosed. Ed Newton-Rex, a music tech entrepreneur (and, full disclosure, briefly my boss at TikTok), investigated the product and observed that Suno output seemed to plagiarise from copyrighted works. A few months later, the RIAA filed a lawsuit against Suno and Udio for copyright infringement. In response, the CEO of Suno (Mikey Shulman) said that “Our technology is transformative; it is designed to generate completely new outputs, not to memorize and regurgitate pre-existing content.” Suno answered the RIAA’s complaint a few weeks later, settling the factual question of whether they had used copyrighted works — they had! — but defending the practice as “fair use”.

The RIAA complaint cited examples of prompts that can lead Suno to plagiarise well-known songs. If Suno is prompted with the lyrics to Mariah Carey’s “All I Want For Christmas Is You”, along with ‘style’ prompt “m a r i a h c a r e y, contemporary r&b, holiday, […]”, the output sounds just like the original song. Mind you, it is not an exact copy: the timbre differs from the original, and there are vocal ornaments that don’t sound quite right. But it is clearly a copy of some kind. Ted Chiang has described ChatGPT as a “blurry jpeg of the internet” to explain how it reproduces text it has been trained on, but in a lossy way. Butchering his phrase, we might call Suno’s version of “All I Want For Christmas Is You” a ‘blurry mp3’ of the original.

Suno puts up minimal defenses against plagiarism: their terms of service forbid you from entering copyrighted lyrics, and if you enter artists’ names as prompts, the model tends not to sound exactly like them. However, the system does not complain if you enter copyrighted lyrics, and the part of the model that interprets the prompt seems easily tricked into producing soundalike vocals. The odd spacing of “m a r i a h c a r e y” is one strategy; Newton-Rex’s investigation found that prompts like “beminem” and “blank-184” also worked.

Retrieval vs. generation

Suppose you’re asked to fill in the blank — “The capital of France is _____” — and you write down the word “Paris”. You might imagine that somewhere in your brain, the ‘answer’ to the question was ‘stored’, and you ‘retrieved’ it. Some people imagine that if GPT4 (one of the LLMs used in ChatGPT) were prompted to fill in the same blank, it would do the same: find a datasheet about world nations, look up “France”, and ‘retrieve’ the word in the “capital” column. This is not how it works. GPT4 is a generative model: given an input sequence of words, the model predicts which word is most likely to come next, and continues the process iteratively. For the capital of France prompt, it might generate the token “Paris” with the highest probability (say, 97%), even though other continuations are possible. For example, there may be a 0.1% probability that the next token will be “located”, and that we are writing the sentence “The capital of France is located on the Seine”).

Many generative models for music work the same way, by generating a sequence of abstract tokens, one at a time, just like GPT4 generates words. If Suno works like this, then when their model happens to generate the same melody as “All I Want For Christmas Is You”, Suno could argue that the model did not memorise and then retrieve the melody, but generated it. The model started at the beginning, picked a note, and tried to determine at each step what the best next note should be. In this way, it did not copy the Mariah Carey song; it independently invented it.

Menard’s Quixote

Did Suno re-compose “All I Want For Christmas Is You” without plagiarising it? The notion is absurd, but it has (fictional) precedents. Jorge Luis Borges, in his 1939 short story “Pierre Menard, Author of the Quixote”, imagined an author who aims to write Don Quixote, three centuries after it has already been written. Menard does not want to adapt the plot of the book to a new setting, like “one of those parasitic books that set Christ on a boulevard” (or Romeo and Juliet among a village of garden gnomes, say). He does not want to rewrite Cervantes’ Quixote mechanically, either by copying it word for word, or by memorising and then recalling the book. He wants to go through all the labour of writing a new book — devising characters, plots, word choices all on his own — and yet, through deduction and the weighing of alternatives, to make inescapable choices such that he ends up producing, in his own hand and by his own decisions, a book that just happens to be, word for word, identical to Quixote.

In the story, set in the early 1900s, Menard considers different strategies for rewriting Quixote, which was written in the early 1600s. Among them is to renounce his modern possessions, move to Spain, emulate the original author Cervantes, and proceed from there. But Menard views this as an uninteresting solution: “Being, somehow, Cervantes, and arriving thereby at the Quixote — that looked to Menard less challenging (and therefore less interesting) than continuing to be Pierre Menard and coming to the Quixote through the experiences of Pierre Menard.” In other words, copying Cervantes could lead to recreating the Quixote, but it would still be an act of copying.

Complicating matters, Menard has already read Don Quixote as a teenager. He has not memorised the text, but it has visited his brain, and a trace of it remains — call it a compressed copy in his mental model, a blurry jpeg of the original; or, in Menard’s words, a “general recollection of the Quixote, simplified by forgetfulness and indifference”, “the equivalent of the vague foreshadowing of a yet unwritten book.” Menard complains that Cervantes had the easier task: he could be “swept along by the inertia of the language and the imagination”, letting “chance” guide him, whereas Menard needs to somehow arrive at the same original text, but with a different starting point.

Ardmen’s All I Want

Imagine a modern-day Menard — let’s call her Perrie Ardmen — who is trying to write a song identical to “All I Want For Christmas Is You”, but without plagiarising Mariah Carey. Ardmen considers putting herself in Carey’s shoes — moving to New York, attending beauty school, training her voice, and executing a random walk through life that just happens to retrace Carey’s. She rejects this as a simplistic solution. Instead, Ardmen engages in a greater labour: she begins to write a song, considering all possible starting points — “Shall I write a song about my mother? About an ex-lover? About what happened to me this morning?” — and must somehow make creative decisions to reject the alternatives and justify choosing the desired path: “I’ll write about Christmas, and how I don’t need all these material things, just my lover.”

Having picked a topic, Ardmen now needs to write some lyrics. She considers different starting words — “Christmas”? “Santa”? “You?” “If”? — and has to persuade herself to reject them all until only the necessary word, “I”, is possible. Ardmen then considers options for the second word — “I am”? “I will?” “I Christmas?” — and so forth. Menard described this process as a “game of solitaire” with two rules: “the first allows me to try out formal or psychological variants; the second forces me to sacrifice them to the ‘original’ text and to come, by irrefutable arguments, to those eradications”. An LLM engineer might describe it as an auto-generative model.

Complicating matters, Ardmen has already heard the 1994 holiday classic “All I Want For Christmas Is You” several times. She avoided hearing it last year and couldn’t recite all the words, but she knows it well. While writing her new song (which happens to be an anti-materialist love song set at Christmas), she relies on a ‘general’ mental model of music, but she must avoid accessing any part of that model that has retained information about the 1994 song. She laments that while Mariah Carey could be guided by inspiration, she needs to be rigorous and diligent, avoiding any thought of the original but nevertheless returning to it. Carey could write by chance; Ardmen must write by probabilities.

Labour and richness

Ardmen might not finish her song; it is a laborious task. Based on the excerpts I’ve seen, Suno only generated about 30 seconds of All I Want For Christmas. Menard made some progress but never finished his Quixote: the narrator of the story, reflecting on Menard’s writings, notes that they include “the ninth and thirty-eighth chapters of Part I of Don Quixote”. Even so, the narrator calls these chapters “the most significant writing of our time”; he appreciates the unfathomable labour that produced them; and, he writes, “The Cervantes text and the Menard text are verbally identical, but the second is almost infinitely richer.”

Writing music also requires unfathomable labour. That’s why, according to the CEO of Suno, “It’s not really enjoyable to make music now”. To address this, Suno has compressed the labour of millions of musicians, preserving it as a vast library of blurry mp3s that can be used to make lossy copies. The musicians’ recordings that trained Suno’s models and the lossy outputs generated by Suno’s users may sometimes be musically identical, but compared to the musicians, Suno is almost infinitely richer.