beyond policy lab 7
The Question that was Never Just a Question: A Live Lab for Philosophy Enthusiasts on AI and the Wisdom of Thinking Carefully About Everything
An interest in philosophy was a common denominator for this group, comprising Chandru, Durga, Sandeep Chandran, and Vijay Somanath. Each participant had a different point of entry into the world of philosophy, with its wisdom giving them an anchor to reflect on a range of different realities they were each grappling with.
Sandeep, an engineer, was drawn into philosophy after reading Fyodor Dostoyevsky’s The Brothers Karamazov. The book left him with a lot of philosophical questions, particularly the chapter The Grand Inquisitor, which has been published as a separate book in itself. Following this, he picked up the thread by reading Marcus Aurelius’ Meditations, Seneca’s Stoicism, and spanned a wide spectrum of branches of philosophy. He grapples with these different streams rather than to identify any one as an absolute fount of wisdom.
Durga works in Governance, Risk, and Compliance in cybersecurity, and her foray into philosophy began with her upbringing. Raised in a Hindu household, she came across different spiritual philosophical views rooted to the religion. Reading and engaging deeper helped her become familiar with Buddhism, Stoicism, and other schools of thought. A personal tragedy pushed her to reflect on the paradoxes of life and morality. Durga’s reflection also included a memory from watching a movie from the Avengers franchise, Avengers: The Age of Ultron, where Iron Man creates Ultron from his consciousness. Leaving it aside for a few minutes, Ultron learns everything about everything and decides to kill everyone in the world. That moment was scary - the very potentiality of it happening. Bridging both, AI and philosophy, Durga finds this to be an intriguing moment in human history and notes that governance is one of the best things we can do at the moment to address this technology and its full impact.
Chandru’s beginnings with philosophy were also similar, having grown up in a Hindu household and encountering the spiritual side of things. However, the “core of philosophy,” as he put it, came through physics. While doing his BSc Physics, he had picked up Sanskrit as his second language, and came across the Ṣaṭ Darśana (pronounced: shat darshana) or the Six Indian Systems. He learned that science and physics offer an understanding of the what and how, but never the why, and philosophy helped him reflect on that. Engaging with individual texts helped him see that there were a lot of different kinds of interpretations, with some dominating over the other. He realized that the knowledge he had gained by the time was insufficient to really get to the deeper point of uncovering the why. A deeper study of Sanskrit led him to Indian philosophy, and comparative analysis led him to philosophical streams from the West, particularly Greek and Roman.
Vijay’s beginnings with philosophy came through an epiphany about existential thoughts. Waking up one day and feeling everything was pointless started the whole inquiry. His exploration of philosophy helped him realize that the question of “What is the point” was in itself meaningless. This helped him grapple with the full depth of philosophy as a point of inquiry, and a framework to reflect.
Naming and Framing the Problem
The group noted that AI comprises systems that process language, generate text, make predictions, classify images, recommend content, make decisions about people's lives in domains from hiring to criminal sentencing to medical diagnosis to credit at a scale and speed that human civilisation has never had to govern before.
From a philosophical perspective, AI is a set of systems that raise, in acute and practical form, nearly every major question that philosophy has ever asked. Specifically, these questions include: What is intelligence? What is consciousness? What is moral responsibility? What do we owe to entities whose inner life we cannot verify? How should decisions be made under radical uncertainty? Whose values should a system optimise for when values conflict? What is justice when the algorithm cannot explain its reasoning? None of these are new questions, though. If anything, they are the oldest questions humankind has grappled with. They are just now attached to systems that are making consequential decisions about billions of people's lives.
AI systems generate outputs that carry the surface appearance of knowledge while having a fundamentally different relationship to truth than does human knowledge. A large language model produces the most statistically probable continuation of a text. It is unlike humans in that it does not know whether what it says is true, it does not have beliefs of its own, and it does not have the capacity to be wrong in the way that a person who holds a belief can be wrong, because holding a belief requires a relationship to evidence, to revision, to the possibility of learning that you were mistaken.
Responsibility for harm was a major question. Who is responsible when an AI system causes harm, when an algorithm denies someone a loan they deserved, when a predictive policing system targets a community unjustly, and when a content recommendation system amplifies content that damages a teenager's mental health? There is no clarity on who is responsible in these situations. Existing governance approaches do not answer these questions sufficiently because they are built on an idea of moral responsibility that is defined for individual humans and their rights, rather than for distributed agency, collective responsibility, and the moral status of entire systems.
AI systems make decisions that allocate benefits and harms. These decisions are made using training data that reflects historical patterns of injustice, optimising for metrics that embody particular values, in ways that their developers often cannot fully explain. These historical patterns of injustice have never been addressed through the lens of justice. Harms have been perpetuated in a certain way, and now, the same harms are continuing with different sites of manifestation.
As a group interested in philosophy, the focus on consciousness was of particular interest. There is no concrete answer on whether AI systems have any form of inner experience - that is, is there anything that defines or explains what it is like to be a large language model processing a query? The philosophical frameworks for answering this question are also contested. If AI systems have morally relevant inner states, the obligations we have toward these systems may be different from what we are currently assuming to be true. If they do not have morally relevant inner states, the simulation of inner states can create deception risks that governance frameworks have barely begun to address.
AI governance decisions that we make in the present will shape the future of the world. Our current approaches to governance are restricted to calling for accountability in the present. They lack any reflection or theoretical reflections on our obligations to future beings who cannot consent to decisions being made on their behalf. We are looking out toward a future we cannot predict and the decisions we make now will be critical for the future because a countless number of yet nameless and faceless lives will bear these impacts.
With this in place, the group reflected on three questions:
Every major AI governance question is a philosophical question that has been asked before. What is the impact of the people making AI-related decisions proceeding without engaging with those answers?
Philosophy's most important contribution is not to provide answers but to prevent the wrong questions from being treated as if they have obvious answers. What aspects of AI development / use are currently being treated as absolute / inevitable? What does a philosophical inquiry of these aspects reveal?
The diversity of philosophical traditions represents a diversity of answers to the most fundamental questions about knowledge, ethics, and what it means to live well together. What does it mean to build AI as if there were one answer to those questions? What would it look like if we take philosophical pluralism seriously?
Creating a repository of wisdom
Each member in the group was invited to share one concept, argument, thought experiment, or philosophical practice from the tradition or thinker that speaks to governance. The themes that emerged were as follows:
Avoiding reductionism, where most systems tend to reduce something to a statistical median. We are building AI systems trained on reductivism to produce reductive outputs. In a global system with an ever-evolving value system and infinite diversity, it is critical to find out how to minimize and avoid reductivist approaches.
Yin and Yang or balance of the universe between cosmos and consciousness: The things people in power do (be it good or evil), will inevitably come back to them. The way AI is developed and implemented will eventually affect the creator.
Butlerian Jihad, from Dune, which tells us that machines that think would be used to enslave others.
Paropakarartham Idam Shareeram, or the body exists to serve others, tells us that AI should exist to serve and do good. What “good” is, should be defined by context and governance.
Democracy, where AI should be owned by communities from start to finish.
Ubuntu: The Southern African philosophical concept that a person is a person through other persons, that identity and flourishing are fundamentally relational. Ubuntu offers a direct challenge to AI systems built on a model of the individual user as the unit of analysis, and to governance frameworks that focus on individual rights without attending to the relational fabric those rights exist within.
Anekantavada: The Jain philosophical principle of many-sidedness, the view that reality is complex and can be validly described from multiple perspectives, none of which is complete. Anekantavada is a powerful governance model for holding the genuine pluralism of values and knowledge traditions that AI governance must navigate, and as a challenge to any framework that claims to have found the single correct answer
Rawls’ veil of ignorance, a thought experiment that asks us to choose principles of justice without knowing our position in the society those principles will govern. Rawls’ veil of ignorance works as a tool for stress-testing AI governance frameworks: what principles would we choose if we did not know whether we would be the developer, the deployer, the user, or the person most harmed by the system?
Wu wei from Daoist philosophy, or the concept of non-coercive action. Wu wei calls for governance that works with the nature of things rather than against it, of the wisdom of knowing when not to act — as a challenge to AI governance frameworks that assume more regulation and more intervention is always better, and as a model for thinking about what governance should leave alone.
Kant’s categorical imperative, or the requirement to act only according to principles you could will to be universal laws. Kant’s categorical imperative is a direct test for AI governance decisions: could you will that every organisation, in every context, make this decision in this way? What does the answer reveal?
Pratītyasamutpāda, or the Buddhist concept of dependent origination. Pratītyasamutpāda offers us the understanding that all phenomena arise in dependence on causes and conditions and that nothing exists independently — as a governance model that requires AI systems to be understood not as isolated tools but as nodes in vast networks of causes and consequences, and that any governance framework must account for the full ecology of effects rather than only the immediate outputs.
Mill’s harm principle, or the view that the only legitimate basis for restricting liberty is to prevent harm to others. Mill’s harm principle is a framework for AI governance that is both powerful and limited: powerful because it gives a clear criterion for intervention, limited because it requires agreement on what counts as harm, who counts as an other, and what the relevant comparison is.
Epistemologies of ignorance from contemporary feminist and critical race philosophy, or the philosophical analysis of how systems of power produce systematic not-knowing, how dominant groups develop elaborate practices for not-knowing what it would be inconvenient to know. Epistemologies of ignorance from contemporary feminist and critical race philosophy offers us a framework for understanding why AI systems reproduce injustice not through malice but through the structured ignorances that their developers inherit from the broader social context.
Socrates’ view of the examined life, or the insistence that the unexamined life is not worth living, that genuine wisdom begins with the honest acknowledgement of what one does not know. Socrates’ view of the examined life is a governance requirement: what would it mean to require AI developers to practice the Socratic examination of their own assumptions before deploying systems that affect billions of people?
Ahimsa, or the principle of non-harm from Hindu, Buddhist, and Jain traditions. Ahimsa is not as a simple prohibition but as a demanding positive practice of attending carefully to the ways one's actions ripple outward and cause suffering that was not intended and was not foreseen, and taking responsibility for that suffering anyway.
The capabilities approach from Amartya Sen and Martha Nussbaum, or the view that justice requires attending not to what people have but to what they are able to do and be. The capabilities approach is a framework for evaluating AI systems not by their outputs but by what they enable and what they foreclose for the people they affect.
Oyèrónkẹ́ Oyěwùmí's critique of the ocularcentrism of Western philosophy, or the analysis of how Western thought's privileging of vision as the primary sense has produced a particular and limited understanding of knowledge, identity, and social order. This is a challenge to AI systems that are overwhelmingly trained on visual and textual data, and to governance frameworks that privilege what can be seen and measured over what can only be felt, heard, or known through relationship
Building a Container for Governance
AI-related decisions are often made without engaging deeply with the wisdom that has emerged from reflecting on philosophical questions in the past. Drawing a parallel with the question of using violence, the group looked at how some philosophies entirely abhor violence, while some say violence is a means and can be justified in some instances. Nation-states and the general public have seemed to consider violence an inevitable necessity at different points, and this percolates into how and for what AI is being used. This is in itself a pattern. The use of violence, and now AI, can also be detrimental to human existence itself, and the most vulnerable are the ones who have already been historically disadvantaged. Certain kinds of violence are acceptable in the name of national security, but a group trying to take back something that was stolen from them is expected to respond peacefully. This is where nuance matters. Standardizing the rules around violence without answering the nuanced distinctions between different regions, people, and communities of the world will only produce more harm. AI governance should pay meaningful attention to context, nuance, and lived realities to be effective.
There is an assumption that AI will replace humans. However, this is a belief held by and among certain people. This is a choice made by those in positions of power, holding a profit mentality. AI does not have to replace humans. Philosophy helps us explore a variety of questions before we make choices, including recognising potential outcomes rather than to assume inevitability. It doesn’t have to be the case that humans will lose their jobs.
There is also an assumption of competition. States are competing with each other around AI development, because that is how the narrative on AI has been set. It is true that AI is used in war, defense and security, and scientific innovation. All of these become sites for competition and how political governments are looking at it, and this is all being presented as inevitable. Why is collaboration or cooperation not treated as the inevitable norm, but competition is considered inevitable? Philosophy reminds us that the nature of every narrative we frame, including inevitability, itself needs interrogation.
Diverse philosophical traditions offer a diversity of answers to the most fundamental questions about knowledge, ethics, and what it means to live well together. Not acknowledging these pluralities in developing and building AI may mean that any tool emerging from a singular starting point would be impractical for use. Philosophical pluralism also tells us that you may not have single, one-size-fits-all, set-in-stone answers to certain questions at all, and that that in itself offers a rich approach. Varying perspectives and objectives must be surfaced to make decisions, and there could be varied decisions for varied contexts. Diversity should be centred, and space must be made for plural outcomes rather than a flattened, singular outcome for everyone.
Philosophy gives us a way to look at the world from different perspectives. This in itself is a powerful offering for AI governance. The entire purpose of AI governance is to ensure that there should be no harm, which leads us to a philosophical question on society and social good. If we don’t engage with this aspect of governance with a philosophical bent, we may risk normalising harm. For instance, in data collection, if we don’t include data on particular groups, such as women, historically oppressed groups, people with disabilities, and so on, the emerging tool will carry that bias and its benefits will not accrue to those groups. Rather, harm will, because a system built on the use of such a tool will exclude and discriminate against these groups. Acknowledging diversity of thought, vulnerability, power, and privilege are vital to AI governance.
An important learning from philosophy for AI governance is to be attentive to unintended consequences, because they are scaled heavily when left unaddressed. An AI model is a file sitting on a server somewhere and cannot do much by itself. But they are scaled and integrated into everyday life at a staggering pace, and their use can result in unintended consequences of an undesirable kind. This calls on us to recognise displacement, extraction of mineral resources, environmental degradation, occupation of land, and the destruction of knowledge as issues to address. We don’t know everything that can happen with the emergence of AI, but being attentive to all that we can imagine as possible or likely to happen will enrich governance.
While reflecting on philosophy, it is important to acknowledge that certain streams of philosophical thought are guiding the development of AI to the exclusion of other worldviews. As people in positions of power make these decisions, they work to corral that power rather than redistribute it. The dominant idea is to put technology first, and to evangelise whole communities into buying the narrative that technology of a particular category is inevitable. The assumption underpinning this is the economic idea that for anything to come into existence, something should be destroyed. These dominant frameworks exclude other ways of existing and approaching this technology. Governance must recognise different perspectives rather than assume a single path or dominant approach.