top of page
  • Writer's pictureLeslie Predy

Cognitive Dissonance: Re-evaluating AI through Noise and Improvisation




My goal in this residency is to use my background as a noise musician to explore the creative potentials of noise and improvisation in human and AI systems. However, it is first necessary to clarify that I don't wish to explore current “artificial intelligence” technologies like ChatGPT, Dall-E, Copilot, and others. These systems are not “intelligent” or even really “artificial,” the hype and panic around these is big tech corporate propaganda. To develop anything resembling artificially intelligent technology, the dominant narratives of the tech industry need to be explicated and challenged.


Matt Nish-Lapidus summarizes today’s state of technology well: “so-called 'AI' does not think. It doesn't know things. It doesn't model the world or thought processes or ideas or anything like that. It doesn't know math, or english, or physics, or what a hand is.

What it does do is create immensely complex and inscrutable statistical models of the data that is processed when creating the model. It contains a set of deeply nested statistical weights that determine which tokens are probable given a set of previous tokens (tokens being words, letters, sentence fragments, clusters of pixel colours tagged with words, etc).”


Their outputs are what Hito Steyerl calls “statistical renderings''; word or pixel groupings that are the stochastic combinations of huge quantities of data we have created and placed on the internet. The statistical nature and heavily sanitized data sets comprising these systems produces an averaging effect: the outputs are the most generic and mediocre relationships of conceptual proximity. They're reminiscent of stock images, representative but in an oddly uncanny way. But something deeper is being reflected back.




Interfaces // Ideologies


The simple conversational interface seamlessly renders images or text complete with their inherent meaning. Yet, this interface conceals the inner working of the algorithm. At its core–statistics—with its origins in eugenics, is visualizing the given prompt, but also dominant social patterns, values, and inequalities. The result is a flattened version of the world, but a highly ideological one (Steyerl). The eerie emotional reaction is a reaction to the ideology on display.



The outputs of these systems do not represent reality, but a carefully designed version its creators want you to see. Behind the curtain is big tech’s capitalist obsessions with data extraction and commodification as a means to consolidate wealth and control. Many excellent critiques on these topics exist, and we won’t go into detail.


In relation to agency and interaction, the concern is the effect of these systems to control the way we think about ideas and see the world. Leif Weatherby called ChatGPT and other AI systems “an ideology machine,” using the Marxist definition of ideology as the "production of ideas, of conceptions, of consciousness." These systems both instantiate and contribute to the conventions and culture of the dominant ideas of society. Weatherby says they are automating ideology and this is dangerous.


This is especially problematic as the current set of “AI” systems don’t do anything (I personally) find innovative or even useful, they aren’t even tools. As we use these systems we are being trained and exploited in service of corporate agendas of power and control and get very little in return. If we wish to imagine truly complex systems of interaction or intelligence, computational tools must be conceived of and built differently. To move forward on a new path it is necessary to set these in a historical context, and we need to take a small diversion to the history of computing.


Capital // Control


Post WWII, the US government sponsored research projects that partnered with large corporations, giving us many technologies/infrastructures we use today (the internet, GPS, etc) and in the process coining the term cybernetics. In his 1948 book Cybernetics, Norbert Wiener  stated “...the present time is the age of communication and control,” succinctly summing up the original foundation of the field. Values of cybernetic capitalism are deeply embedded in our current landscape of big tech. As Timothy Erik Strom outlines in capital and cybernetics, this relationship of tech companies to the American military industrial complex persists to this day. Computing technologies have always been inherently about capital, power and control. 


Even if unconsciously, these ideas of command and control are entrenched as best practices in the development of algorithms and software. As Wendy Chun noted, one facet of software is hegemonic ideology, embodying Antonio Gramsci’s notion of hegemony as ideological common sense, and this status quo must be challenged.


An illustrative example is the concept of abstraction. Abstraction, a fundamental principle in “proper” programming, is not inherently bad. All programming languages are fundamentally abstractions, effectively "hiding (and thereby securing) the machine" (Chun). However, for technologies stemming from capitalist frameworks, this concept permeates all levels of the stack in problematic ways. The interfaces of today’s AI systems are designed to abstract the functioning of the algorithms, thereby diminishing comprehension and agency of its users. A contradiction of agency lies at the heart of software, where agency denotes both one's ability to act and the ability for others to act on one's behalf (Chun). In these systems, the ability to interact with the algorithm is governed by programmers who define its parameters and conduct training, aspects that remain opaque to users. Abstraction manifests at other levels as well, hiding and securing many things, including the typically marginalized labor of cleaning the data sets, and the looming environmental devastation of cloud infrastructure.


Adorno and Horkheimer’s phrase the “leveling domination of abstraction” is fitting here: the leveling is a flattening of the perception of what is possible, reducing capacity for creativity and agency. Through abstraction, these systems perpetuate the dominant ideologies that benefit their creators. They allow the access of some things while concealing others, shaping our modes of participation. 


Today’s software has created assumptions that we rarely examine. By exposing software’s inherent ideology, we can move beyond a purely technical narrative, toward exploring more nuanced possibilities for the future. OpenAI defines artificial general intelligence as “highly autonomous systems that outperform humans at most economically valuable work.” This is a sad vision for the future of computing rooted in elitism and capitalism. Technology holds so much more potential if we can break free of this thinking. Rather than strive toward some ill-defined goal of "artificial intelligence" (because how do we even define intelligence?), could we imagine a complex, algorithmic systems that enable human creativity?


 

"One produces what technology makes possible, instead of creating the technology for what one wishes to produce." Jacques Attali in Noise, the Political Economy of Music

To reconsider our approach to these systems, we can draw insights from sonic perception, music, and our relationship with sound. Fundamentally, letters and numbers are representations of sounds, organizing sound for communication. Sound and frequency are the original basis for human to human communication. Noise emerges as an essential element in this paradigm, conceptually emerging in all forms of communication (noise as sound frequency, noise as visual information, noisy datasets in computation, etc.). Noise is a contrast to signal and meaning and a means of de-centering subjectivity, allowing a restructuring of meaning beyond traditional narratives of value.


Signal // Noise


Composer Pauline Oliveros embraced noise in her practice of Deep Listening, which involves a conscious acceptance of all sound without judgment. Sonically deepening, as opposed to a flattening (in Adorno’s notion of abstraction), Deep Listening embraces chance, "error," and distortion of signal or noise as means to expand perception. This is crucial in prioritizing difference over standardization and the absurd over the practical to widen the sonic experience. 


In relation to computation, this means reconsidering the training data provided to algorithms. Today these are heavily curated, culturally biased, and exploitative. Mistakes, outliers, and chaos are essential for evolution and creativity, technology today actively attempts to eliminate these. Whenever one optimizes for something, much must be left out, and we are currently in a shallow pool that must be deepened with a more ethical framework.


The inscrutable statistical models behind these systems need to be demystified. Algorithms start with noise (either as combinations of letters or Gaussian noise), to generate their outputs, this must be embraced and opened up to the users of these systems to interact with. Inserting processes with a robust ability to handle chaos could potentially have these systems truly learn, rather than mimic back to us what we have already done. To change this paradigm, the subject/object relationship must be re-considered.


Denoising Diffusion Probabilistic Models (Ho et al, 2020)

Furthermore, the contradiction of agency between programmers and software users is paralleled in the musical realm with composers, performers, and listeners. Historically, the musical score provided by a composer reduces the agency of players in creating compositions, and listeners traditionally have no agency except to consume. Many composers and musicians challenge this hegemonic structure through participatory and improvisational practice, often paired with technological innovation. This can be seen in musical practices from Sun Ra to Stockhausen (Stapleton and Davis). 


We can look again to an example in Oliveros’ work with Sonic Meditations, which are “an attempt to return the control of sound to the individual alone…attempting to erase the subject/object or performer/audience relationship.” These are prompts that anyone, regardless of musical training, can engage with. They are designed for group participation, and create emergence when performed in large numbers. Oliveros subverts traditional music “scores” in favor of participatory meditations. Likewise, computing needs to enable the agency of a system’s users in building and reconfiguring that system. Rather than a structure dictated by hegemony, a participatory or communal model is needed. This participatory model of computing would extend throughout the stack: from building the datasets, the algorithms, and creating an interface that empowers rather than hides.




Improvisation // Agency


David Borgo and Jeff Kaiser characterize improvisation as a mutually constitutive process through which users, technologies, and environments are dynamically engaged in refashioning one another in a feedback loop. An improvisational model of computation could create such nonlinear systems with distributed agency (similar to Sonic Meditations), blurring the subject/object relationship.


One could imagine interacting with artificially intelligent systems similarly to musical instruments–as a tool aiding expression. Musicians can be in dialog with other musicians, their instrument, and the environment. The fundamental frequencies of communications are open for all to contribute to. Applying this to computation would mean machines, algorithms, programmers and users interacting with technology all contribute to the outcomes.The development of technology can be influenced by individuals and social groups who in turn can be influenced through their interactions with technology. Rather than consuming ideology, participants can help shape it.


Creating a new paradigm for computing is not something I can do in a residency, but with this research I’ll be experimenting with algorithms and sound, foregoing pre-existing datasets and models to work with local sound datasets I’ve created, and looking to practices like Sonic Meditations.







References

Matt Nish-Lapidus (@emenel). 'So-called "AI" does not think...', March 6, 2024 https://post.lurk.org/@emenel/112049395846617922

Hito Steyerl. ‘Mean Images’, New Left Review 140/141, 2023

T. Gebru and Émile P. Torres. 'The TESCREAL bundle: Eugenics and the promise of utopia through artificial general intelligence', First Monday, 29(4), 2024

Leif Weatherby. ‘ChatGPT Is an Ideology Machine’, Jacobin, 2023

Timothy Erik Strom. ‘Capital and Cybernetics’, New Left Review 135, 2022

Wendy Hui Kyong Chun. 'On Software, or the Persistence of Visual Knowledge', Grey Room 18(4):26-51, 2005

OpenAI charter, https://openai.com/charter, accessed 17 April 2024

Paul Stapleton and Tom Davis. 'Ambiguous Devices: Improvisation, agency, touch and feedthrough in distributed music performance', Organised Sound 26(1):52-64, 2008

David Borgo and Jeff Kaiser. 'Configurin(g) KaiBorg: Interactivity, Ideology, and Agency in Electro-acoustic Improvised Music'. Beyond the Centres: Musical Avant- Gardes Since 1950 – Conference Proceedings, Thessaloniki, Greece, 1–3 July, 2010






89 views

Recent Posts

See All

Comments


bottom of page