The advent of generative AI has unsurprisingly evoked diverse responses. As with any major historical shift, diverse perspectives emerge, with individuals interpreting the current situation often in notably contradictory ways—consider the Armageddon-like doomerism juxtaposed with the naïve optimism permeating discussions.
One way that I find useful to understand this moment is through a semiotic approach. Without being too technical, semiotics is the study of signs, symbols, and their interpretation, examining how meaning is created and communicated through various elements in language, communication, and culture.
In the realm of generative AI, signs abound; Large Language Models (LLM) tokenizes linguistic signs whereas text2audio models, such as Meta’s MusicGen, tokenizes sonic signs (all converted into data and then, almost miraculously, back to their original domain). Yet, within this technological revolution, a specific ensemble of signs takes center stage—style. To be more specific, the transformation of styles into abstracted signs.
Style, a multifaceted term, can be understood in semiotic terms as a distinctive and meaningful arrangement of signs, symbols, and expressive elements within a communication system, contributing to the creation of specific aesthetic, cultural, or individual identities.
As you see, under this definition, style is something that is the result of a constellation of signs, so what does it mean to treat style not as a constellation of other signs but as a sign in itself?
Styles of Styling Styles: A Semiotic Perspective
Anthropologist Eitan Wilf has argued that contemporary forms of algorithmic-based creativity can be aptly described as styles of styling styles. In Wilf’s words:
“this historical moment should not be analyzed in terms of one or a number of fixed styles, but rather through the prism of styles of styling styles. In other words, the age of computer- mediated, algorithmic forms of sociality might be better analyzed through the restraints that govern the practices of styling styles with the aid of these algorithms and similar technologies” (Wilf, 2013, p.719)
If in the pre-algorithmic era, style emerged out of a play with lower-level symbols (icons, words, musical notes, etc.), in the algorithmic era, we are going up a level to work directly with the styles themselves in a way that was never possible before.
Moving forward, once style is abstracted (and this is what Wilf calls “styling styles,”), it becomes a sign that can be mixed, manipulated, and transformed. Now think about prompting MidJourney, Dall-E, etc. with the prompt “Sunflowers in an impressionist and Baroque style”; this is the act of styling styles.
But there's more to it. Once we work with style as an object, a higher level of style emerges, aligning with our human inclination to classify and categorize our world. Now we are speaking about styles of styling styles. Wilf alludes to the style characteristic of the technology itself—think of the distinctive yet consistent quality in common MidJourney outputs or the flickering aesthetics of first-generation Diffusion-based videos. This distinct fingerprint arises from the technology's limitations and affordances.
Flickering effect
The Slippery Slope of Style Abstraction
The abstraction of style is remarkable. Who would have imagined just a couple of years ago that we could effortlessly blend entirely different styles on the fly? Let's set aside any judgment about the quality of these fusions for now.
However, there's a potential pitfall here. The abstraction of style, like any abstraction, poses the risk of obscuring the foundational components that originally shaped the style. Consider a parallel in programming languages. Initially, there was Machine code, characterized by binary instructions and demanding an in-depth understanding of a computer's hardware architecture. In contrast, modern programming languages prioritize abstraction, offering high-level representations that shield developers from intricate hardware details. One consequence is that many developers are no longer aware of the underlying processes operating under the hood.
Returning to generative AI, it's evident how the abstraction of style could obscure the various layers contributing to this style, encoded in data from the outset. But this time, these contributing layers aren't computing processes; they are the endeavors of human creatives – artists, authors, musicians.
Let's zoom in on music. Picture a scenario where I can create music without relying on specific artists, using prompts like "Rock beat with West African percussion". This prompt won't reveal the human musicians behind the algorithm's extrapolation of the style of a "Rock" beat or "West African percussion".
To be fair, this process of erasing the creative contributors is nothing new. This oversight is prevalent in various domains, including music technology, where anonymous sample packs or VSTs inspired by ethnic music often omit proper acknowledgment of the musicians, and if mentioned, it's buried in a manual that almost no one reads.
Yet, generative AI takes this tendency to an extreme. The training approach adopted by leading AI companies, utilizing massive datasets without ensuring proper attribution, compensation, and copyright clearance, is not just a legal and ethical concern gaining more visibility. It fundamentally alters our perception of the creative process, dehumanizing music by presenting it as faceless, abstracted genres, styles, moods, and more.
The Ethical Imperative: Putting Creatives at the Forefront
If you've stuck with me this far, you might be expecting a flat-out objection to generative AI, but that's not my stance. You might even assume I'm an anti-technological Luddite, not the founder of a startup developing AI-powered music creation tools.
The truth is, generative AI has the potential to tremendously benefit humans, becoming a valuable partner in facilitating and expressing creativity in unprecedented ways. It opens doors to new forms of expression and, importantly, enables a growing number of people to experience the act of artistic creation.
Yet, in the current landscape, we believe there's an ethical imperative for generative AI companies to place creatives at the forefront. While many companies today talk about fair compensation and data transparency, we go further, advocating for creatives to be fully acknowledged and respected. Let’s not forget that training materials are not mere data points; they are inscriptions of creative labor.
But what does this mean in practice? In the face of the looming risk of effacing human creativity behind stylistic abstraction, we must humanize technology in more radical ways. Let’s inject personality and humanity into the process! This shift is the only valid foundation for true innovation that genuinely benefits creatives and allows culture at large to flourish, avoiding the pitfalls of stagnation.
It's a collective responsibility—both for entrepreneurs and users—to ensure that the technologies we use and build don't delude us into thinking we can replace creatives. Even with the most sophisticated algorithms, we can't afford to lose sight of the irreplaceable human touch in the creative process.
Comments