Good artists borrow, great ones steal…what does generative AI do?
When it comes to plagiarism, intention trumps the technical details
Left: Heart on My Sleeve, Right: Ed Sheeran
The AI-generated song ‘Heart on My Sleeve’ was splashed across the headlines recently, showing how it is now possible to automate the vocal style of musicians. The track has unmistakable shades of Drake, easily conquering what has been described as his ‘nasal monotone’ lacking in emotional heft. And while it may not quite have nailed The Weeknd’s distinctive tenor, Universal Music Group was compelled to issue strongly worded letters to streaming services demanding that they take down AI-generated music.
The same story is playing out across the creative arts, with Getty Images filing a lawsuit against Stability AI, the company behind text-to-image technology Stable Diffusion. The age-old question of what constitutes an original piece of work will be scrutinised through the lens of generative AI.
Is there anything new under the sun?
The creator of Heart of My Sleeve, who goes by the apt moniker of Ghostwriter, produced the track by training an AI model on the musicians’ voices. Generative AI models rely on being fed prior knowledge. Suggestions by Google that its Bard chatbot had learned Bengali all by itself (perpetuated by CEO Sunder Pichai), without any training data in the native language, has been soundly debunked by former Googler Margaret Mitchell. It is no more credible than US Senator Chris Murphy’s claim that ChatGPT had taught itself chemistry, which he was disabused of in short thrift.
Charges of plagiarism can only be answered by looking closely at how models are trained. One study shows that diffusion models, which purportedly generate whole new images, often depend on lifting entire images or isolated objects from their training data. Ghostwriter’s track does not fall victim to this type of clear-cut plagiarism because there is no part of the song - no lyrics, no vocal sample - that is ripped directly from other works. Even so, AI-generated music that ‘sounds like’ a given artist lies somewhere in between unauthorised recording of an artist’s voice (an obvious violation of their rights of publicity) and human impressionists who go around imitating others (which goes unchallenged).
As Jarod Lanier explains, generative AI is ultimately a tool for mashing up human thought, and is best understood as ‘illuminating previously hidden concordances between human creations, rather than as the invention of a new mind.’ An AI-generated track, or drawing, may not have been at the forefront of the artist’s mind, but through an interpretive mechanism these tools can bring to surface plausible and often striking new possibilities that remain faithful to the artist’s style.
Generative AI advocates may dismiss the charge of plagiarism on the grounds that it merely emulates and amplifies our own human creative thought processes. However novel we consider a piece of work, it never simply springs into being. It is an argument as old as the Bible itself:
The thing that hath been, it is that which shall be; and that which is done is that which shall be done: and there is no new thing under the sun.
(Ecclesiastes 1:9)
What is an artist to do? There may be technical solutions, such as tools like GLAZE, which alter an artist’s work in ways that are imperceptible to humans but that make it harder for algorithms to mimic their style. Thus ensues a cat-and-mouse game between artist and imitator, the victor of which remains far from obvious.
Why intentionality matters
We cannot simply engineer our way out of plagiarism conflicts. There is wisdom to be found in legal precedent. Let’s take another example from the music industry, featuring two lawsuits issued by the Marvin Gaye estate, whose vastly different outcomes offer a different insight into what constitutes artistic theft.
Ed Sheeran recently fought a copyright case which claimed that his 2014 hit ‘Thinking Out Loud’ copied Marvin Gaye’s classic ‘Let’s Get it On’. Sheeran had already won a court battle the previous year over whether his hit song ‘Shape of You’ was plagiarised. The definition of plagiarised music has shifted over time: for years, the litmus test for whether one song ‘ripped off’ another was similarity in lyrics and melody; in the 1980s and 1990s, as sampling came into the mix, new limits were imposed over how freely samples could be used before crediting.
The latest lawsuit claimed that Sheeran and Gaye’s tracks share the same chord pattern. In a Manhattan courtroom Sheeran rebutted, guitar in hand, that while the chords were similar, they were ‘basic musical building blocks’ shared by dozens of songs. The court was swayed by Sheeran’s argument that chord progressions lack the requisite ‘original expression’ to warrant copyright claims. According to Sheeran’s testimony, modern pop music has a very limited ‘harmonic palette’ of chord progressions - ‘most pop songs can fit over most pop songs’.
This sounds like good news for Generative AI - maybe it, too, amounts to no more than combinatorial play of fundamental building blocks of artistic expression that belong to no single artist.
But now consider the Gaye estate’s successful 2018 claim against Pharrell and Robin Thicke’s ‘Blurred Lines’, where the song shared nothing as tangible as a melody and lyrics, but only a ‘groove’ with ‘Got to Give It Up’ by Marvin Gaye. This case led to a $5 million payout to the family of Marvin Gaye. The deceive judgement had no technical basis; it rested significantly on the stated intentionality of the artists. The giveaway was Robin Thicke’s 2013 GQ interview, in which he was asked about the ‘origin story’ of Blurred Lines. He recalled:
Pharrell and I were in the studio and I told him that one of my favorite songs of all time was Marvin Gaye’s ‘Got to Give It Up.’ I was like, ‘Damn, we should make something like that, something with that groove.’ Then he started playing a little something and we literally wrote the song in about a half hour and recorded it.
In sworn depositions, Thicke later backtracked and said he was ‘high and drunk every time I did an interview last year’; both he and Pharell also testified that they exaggerated Gaye’s influence to help sell more records. But the damage was done.
Copyright infringement claims are no simple matter. Historically, in the music industry, who gets sued has tended to be highly arbitrary. Settlements have been agreed for songs with passing resemblances, while others with glaringly evident similarities have seen no challenge. Recent lawsuits arguably show that it has more to do with the popularity of the song, and how much one can profit from it, rather than any clearly defined musical concept.
As the debate inevitably collides with generative AI, the question of intent may have to trump all technical considerations, given the growing ease with which this technology can emulate various aspects of human artwork. If a model churns out content that is uncannily similar to your favourite artist, there’s a good bet they have either sourced the artist’s content to train their model, or that they have deployed a deliberate technique to codify and replicate the artist’s work without drawing on their actual content. The intent is the same regardless - to profit from the artist’s reputation and known style.
Some artists may be fine with AI replication (subject to a royalty split), others loath to it. Proponents of generative AI, for their part, regularly lean on Picasso’s adage that good artists borrow, great ones steal (it was in fact Stravinsky who coined the phrase, in relation to composers). Appealing to such tropes conveniently bypasses the nuance of how we undertake our work, how we source inspiration from others, where we choose to credit them.
Intention is everything, and we should be wary of any technology that indiscriminately scrapes data off the web without acknowledging its sources of ‘ground truth’.
 | A guest post by
|