beyond policy lab 9
The Voice That Was Never Only Yours: A Live Lab for Singers and the Wisdom of What is Sung
Drawn together into Lab 9 by their shared interest of music, a small group of singers engaging with various genres of music drew from their knowledge, practice, and experience to imagine pathways for AI governance.
Amrita J sings Carnatic and Western classical music. She learned Western classical music from school, and learned Carnatic music online through YouTube. She shared that her caste positionality made it difficult to access formal music training, which is consistently gate-kept on caste lines. Her journey with music taught her that care and transmission must be humanised and equitably accessible to everyone regardless of their caste positionality.
Mangalam Sridhar grew up learning Carnatic music from her grandmother, and had her journey with light music, alongside explorations of various genres to identify what she most loved to sing. What her voice was trained to understand about care and transmission was different for her when she was a child, compared to what it is now. As a child, she understood that care and transmission were all about singing a piece exactly as it was taught. Now, she reflects on care and transmission as pathways to connect to those she is singing for, as music is an experience for her to share with others around her.
Sushma Soma trained in Carnatic music as a child, and was exposed to Ilayaraja’s music and folk music. Her journey with music has taught her that she owes it to herself to make music that is honest to her, that is an expression of her true self, and how it lands or what is taken from that is not on her.
Naming and framing the problem
The group reflected on the tension of the opposites around AI - as a technology that is enabling access on the one hand, and as a technology built on the edifices of power and elitism on the other hand. They explored how the same technology that is helping level the playing field in some ways in the fields of education, healthcare, financial literacy, and accessibility, is also a pivotal tool in warfare, enabling and furthering surveillance, and environmental harm.
Within the world of music, AI is voice cloning systems that can replicate any singer’s voice with minutes of audio. Music generation models are being trained on recordings without artist consent, and systems that produce particular styles or inspired compositions tend to strip the artform from its grammar, devotion from its context, and regional identity from its rootedness. The sound of a singer’s lifetime of practice can now be reproduced, sold, and deployed without their knowledge, consent, or any of the meaning, stories, and lived realities that made their work what it is.
The group also noted that the voice is not a sound file, but a body’s history of breath trained over decades, of histories of oppression and exclusion being embedded into whose voice can be trained and whose cannot, of emotions that are absorbed and metabolised and practiced until they become cell memory. When AI clones a voice, it takes the surface and leaves behind everything that made that surface possible. The room named a key question here, asking what it might mean to govern a technology that believes it has captured a voice when it has, at best, only captured its frequency.
The group also named the fact that microtonal intervals that give a raga its soul, namely shruti, lives between the notes and props up the entire composition. It is transmitted in presence, listening, and in the journey of learning. AI trained on recordings can hear the notes and weave them together based on patterns, but it cannot hear, experience, or embody shruti. AI governance is also missing this nuance, because it measures what is quantifiable and governs by what it can count, while missing the qualitative intelligence that also exists alongside.
The group reflected on how AI is, ultimately, a tool that can support something an artist may not be able to immediately access: A thought partner, marketing material while having to be your own marketer, and even ideas. The group noted that this also points to something more structural, where differential access to resources, support, and the ecosystem it takes to make music is always defined by power and privilege. It also acknowledged how AI offers availability at great speeds and low cost, leading more people to turn to it easily in the push for greater productivity. There was a reflection on whether we need to be as productive at all, in the first place.
Decisions around all that AI can do and cannot do, and the extent of excusability for some of its harms are defined by power, much like how caste has determined who accesses music and in what way. The question of “whose voice counts” is not only a symbolic parallel but a very real question that we must sit with and learn from.
Creating a repository of wisdom
Each member in the group was invited to share one concept, practice, relationship, or understanding from their singing life that carries a governance insight. The themes that emerged were as follows:
Ragas: Ragas define the parameters for melodic movements. Each raga defines the range of what is allowed and not allowed within a song through the Aarohanam and Avarohanam. It sets the boundaries for how the song will be composed.
Talam and gathi set the rhythmic structure for music, and rely on a precise count and cycles of counting. It offers transparency in pacing and is a cyclical check and balance on the emerging music itself so that you don't trip yourself up by speeding up too much and you don't drag the composition by slowing down too much.
Thaniyavarthanam, where the singer gives way to a percussionist to lead a whole solo on stage, acknowledging the full ensemble it takes to build a performance into what it is.
The stories and lived realities of each singer and what they bring to the music, which also shape the way a song or a raga should be understood.
The role of silence in music. It isn’t just about filling the space with lots to say, sing, or play.
The role of breath in music. How a singer arrives to a composition or brings a raga alive is fully a function of how they have trained their breath - which takes time, practice, effort, and a whole ecosystem to make possible.
Building a container for governance
AI has been able to replicate the tone and texture of artists’ voices, but what emerges is a sanitized, clean version that is bereft of unplanned pauses, intakes of breath, swallowing, flaws, or just the act of clearing the throat. These aspects emerge in live performances and are sanitized in recordings, which form the data that are used to train the AI model. Naturally, as a result, AI is learning from these as datapoints and knows only to replicate the sanitized version. This “x factor” in live performances can never be replicated, and any attempt to produce music through AI is inevitably flattening complexity, diversity, and even human imperfection. Governance should strive to avoid endorsing this flattening as acceptable, because it may soon become the norm for what is acceptable as “good music.”
Relatedly, context always matters. Just as in music, there is no one-size-fits-all for what a particular composition, song, raga, or set of lyrics could mean, there should be plenty of room for nuance and complexity in governance. Everything from the preparation and thought process to the actual delivery of a song or music performance goes into making a song what it is. Governance should not stop at what is and is not permissible, but should make space for nuance. A law or policy cannot capture the entire set of vocabulary and grammar of a system. The ways in which we learn from each other and evolve in our own journeys as human beings include pathways where we are constantly breaking and rewriting rules, which a system holds better than a law can. Governance should speak to the system rather than restrict itself to a law.
In the receipt of music by an audience, the way they arrive into the moment counts in making the song what it is: thus, governance must account for and acknowledge relationality. A single song performed live before an audience holds myriads of meanings for people in the room. The same song recorded can meet an audience of one in different ways each time they listen to it. That very song sung together in a group at a gathering emerges out of the meaning that the song holds for that gathering. Nothing about the song is “owned” by either side, because that entire moment and all that constitutes it is collectively held. All of these are driven by relationality - the two sides on either side of a song and what that song means for both sides. Governance should pay attention to all relationships that sustain human life, protect them from erasure by AI, and centre relationality as a core tenet in its functioning.
Power is a key factor in determining access and in designing the very architecture of music as an institution and AI as a tool, phenomenon, resource, and even institution in its own right. Interrogating power is vital: Who gets to access something? On what terms? In what ways are their uses constrained, enabled, or limited by the power of others? How does their agency play a role in determining access to something? Caste, race, class, health status, ability, age, and plenty of other factors have kept music inaccessible. And yet, people from oppressed caste backgrounds have made whole careers in music for caste-privileged people by making their instruments and through their labour - almost entirely with no compensation and high levels of exploitation. These realities are mirrored in the AI world: caste-oppressed communities perform labour to prepare data to train AI, tribal and indigenous communities are exploited to extract minerals to sustain the hardware that makes AI possible, and racially oppressed communities are displaced to make land and water accessible to cool AI systems. Governance cannot side-step this as it is a question of justice and accountability.
A related principle that emerged was respect for communal music forms, lyrics, instruments, and sounds. There is a dominance of particular forms of music that are treated as “core” while most others are relegated to the “periphery” or “margins.” In this, there is an inherent epistemic injustice that suggests that some streams of music are more important than others. This epistemic injustice also paves the way for extraction: Where artists have been known to callously “take” (but was there really an act of giving?) from indigenous, folk, tribal, and ethnic music forms, “mix” them with elements informed by their own “creative decision-making” and produce something they call their own. The world of AI is suffering from a challenge: On the one hand, the data used to train most models comes from an ethos that does not represent the entire world, and the inclusion of the data that are missing comes with a tradeoff of enforced visibility. The option to opt-out on respectful terms is rarely available, because the argument of inevitability is used to endorse the “be there or be square” metric. Governance should respect agency and treat with care, any data that is provided to the AI lifecycle with trust.
Finally, recognising diverse histories, stories, and lived realities was named as a key governance insight. Every song that is sung enters the terrain of subjectivity: There is the history, journey, and meaning of the song on the one hand, as well as the stories, experiences, realities, and complexities of the artist presenting it. Diverse lived realities endure, and make the impact of a composition all the richer. For instance, twenty years from now, a raga should not just be understood as the raga in an objective form. It should be seen in its rich, complete history that includes the stories embedded around it. AI’s rich history is often seen through the limited lens of what Alan Turing and World War II brought into fruition: But there is a richer history, with Al-Khwarizmi (whose name forms the base for the word “algorithm,” and whose work on algebra is foundational given that modern machine learning is essentially mechanized algebra operating across high-dimensional matrices and optimization functions), the Banu Musa brothers (who were pioneers of early automation - with a musical connection no less, given that they built the automatic flute player operated by interchangeable pinned cylinders, that is widely considered the earliest known programmable machine), Pingala, (whose work carried the two-state logic that maps to the binary system and the combinatorial pyramid which maps to Pascal's Triangle), and Panini (who came up with a generative formal language machine comprising 3996 rule-based programs that functioned like a rule-based computer compiler), among several others. The erasure of multiple contributions to this journey to support the dominance of the Western narrative serves to ensure colonial dynamics continue to endure. Governance must recognise and treat the question of whose pen writes the story as a question of power, and treat multiple contributions equally.