Dario Amodei's 2024 essay Machines of Loving Grace is one of the most serious, expansive and influential visions of AI's positive potential yet written by anyone in a position to know. As CEO of Anthropic, Amodei has thought carefully about both the risks and the promise of powerful AI systems, and the essay reflects that seriousness. His vision deserves to be engaged with on its own terms — not dismissed, not merely celebrated, but extended.
What follows is an attempt at extension. Amodei's framework is materially comprehensive: he envisions AI accelerating progress in biology and medicine, mental health, economic development, and the stability of human institutions. These are certainly important domains to care about. But there is a variable his framework does not include — one whose absence reshapes every conclusion he draws, and whose inclusion would change the picture in important ways.
That variable is consciousness itself.
What Amodei's Vision Gets Right
The core of Amodei's argument is that sufficiently capable AI systems could compress decades of scientific and medical progress into years — not by replacing human researchers but by acting as tireless, highly capable collaborators who can hold and analyze vast bodies of knowledge; simultaneously generate and test hypotheses at scale; and navigate the space of possible solutions with a speed no human team could approach.
The specific ambitions he names are worth taking seriously. He envisions the effective defeat of most infectious and degenerative diseases within a decade of capable AI deployment. He imagines AI-assisted breakthroughs in mental health treatment that could reduce the global burden of depression, PTSD, and addiction — conditions that currently affect hundreds of millions and resist conventional intervention. He sees AI as a potential equalizer in economic development, extending to the global poor access to expertise — medical, legal, educational — that has historically been available only to the wealthy.
None of this is fantasy. The underlying logic is sound: if AI systems can genuinely function as expert collaborators across domains, the leverage they provide is enormous, and the domains Amodei identifies are exactly the right places to apply it. His optimism is grounded, experienced, not naive.
And yet. Throughout this vision, one question is never asked: what kind of thing is the human being who will benefit from all of this? The people whose diseases are cured, whose mental health improves, whose economic circumstances transform — what are they, considered not as biological systems or economic agents, but as knowing subjects? What is the structure of the experience that all of this material progress is meant to improve?
The question seems abstract. It is not.
The Materialist Assumption
Amodei's framework, like most contemporary thinking about human welfare, rests on an implicit materialist/behaviorist assumption: that the relevant facts about a human being are physical, measurable, and external. Health is the absence of disease. Mental health is the regulation of neurochemistry. Flourishing is the availability of resources and the reduction of suffering defined in biological terms.
This framework is not wrong. It identifies real goods and real harms. But it is incomplete in a specific way: it treats the knowing subject — the conscious being who experiences health or illness, wealth or poverty, connection or isolation — as a dependent variable. Consciousness, in this picture, is what benefits from material improvement; it is not itself a subject of investigation or a source of insight into what those improvements should aim for.
The consequence is that certain questions cannot arise within this framework. What does it mean to experience wellbeing, as opposed to merely being measurably well? If a person's neurochemistry is optimally regulated but their inner life is characterized by distraction, fragmentation, and a persistent sense of meaninglessness, are they flourishing? If AI systems extend human lifespan but simultaneously erode human cognitive independence and the capacity for sustained attention, is that progress — or is it merely unbalanced, one-sided growth?
These are not rhetorical questions. They point to a genuine gap in the most sophisticated current thinking about what AI should be optimizing for — and more fundamentally, in how we understand the human being that AI is supposed to serve.
The Missing Variable
The Advaita Vedānta tradition of Śaṅkarācārya offers something that is rarely available in Western philosophical frameworks: a rigorous, systematic account of consciousness as the primary datum of inquiry rather than a byproduct or epiphenomenon of material processes.
The central insight — developed through careful phenomenological analysis rather than metaphysical speculation — is that the knowing subject, the Ātman or witness-consciousness (sākṣin), is not produced by the brain or body but is the field in which all experience — including the experience of having a brain and body — appears. This is not a claim that the physical world is unreal; it is a claim about the structure of experience, and it is a claim that can be examined directly by anyone willing to look. After all, without consciousness, does anything else even exist?
What changes when you take this view seriously?
In medicine, the shift is subtle but consequential. A science of health that includes the knowing subject is not limited to the regulation of biological processes; it also attends to the quality of the subject's relationship to those processes. The difference between a patient who understands, accepts, and participates in their own healing and one who experiences their body as an alien system being managed by experts is not merely psychological — it affects outcomes, compliance, recovery, and the durability of health gains. AI systems designed with a model of the knowing subject can support this relationship rather than bypass it.
In mental health, the implications are deeper. Amodei envisions AI accelerating progress in treating depression, PTSD, and addiction — primarily through improved diagnosis, pharmacological intervention, and the scaling of therapeutic access. These are genuine goods. But the Advaita framework points to a dimension of mental health that pharmacology and talk therapy both struggle to reach: the structure of the relationship between the knowing subject and the contents of experience. Much of what presents as depression, anxiety, and compulsive behavior involves not just dysregulated neurochemistry but a specific confusion — the misidentification of the witness with what it witnesses, the sākṣin with the objects of its attention. Traditions that work directly with this confusion — vipassanā meditation, Advaita inquiry, certain forms of contemplative psychotherapy — have demonstrated clinical effects now well-documented in peer-reviewed research. A 2016 Lancet randomised controlled trial found mindfulness-based cognitive therapy as effective as maintenance antidepressants for preventing depressive relapse (Kuyken et al., The Lancet, 2016); a broad meta-analysis across 61 studies confirmed significant benefits for depression, anxiety, addiction, and stress-related conditions (Wahbeh et al., PMC, 2021). An AI-accelerated science of mental health that ignores this dimension is leaving some of the most powerful tools on the table.
In human-AI interaction itself — the domain most directly relevant to everything Amodei envisions — the absence of a model of consciousness produces the failure modes already being documented: cognitive dependency, learning atrophy, the erosion of independent judgment, and what is beginning to be called 'AI psychosis'. A 2025 arXiv study found that AI-assisted programming actually decreases the productivity of experienced developers by increasing technical debt and maintenance burden, with core developers showing a 19% drop in original code output (Jain et al., arXiv:2510.10165); a separate study found that AI tooling slowed experienced open-source developers by 19% on average (arXiv:2507.09089). These are not merely alignment bugs to be fixed with better guardrails; they are the predictable consequences of designing interaction systems on a limited materialist/behaviorist model of the human being.
It is worth being precise about this, because the dominant responses to these failure modes — stronger content filters, improved alignment techniques, more sophisticated behavioral guardrails — all operate at the wrong level. They attempt to constrain outputs without addressing the model of the human being that produced the problem in the first place. No guardrail can compensate for a missing foundation. If the system does not know what the knowing subject is, it cannot be redesigned to serve it by adding restrictions to what it says. The cause is structural, and only a structural solution — a different account of what a human being is, built into the architecture from the beginning — can address it. Currently, no such solution exists within the mainstream AI development paradigm. The tools, the frameworks, the evaluation metrics, and the research priorities are all organized around the behaviorist premise. This is not a gap that more compute or better training data will close.
An AI system that does not know what the knowing subject is cannot be designed to serve it.
What a Conscious Foundation Would Change
If Amodei's vision were extended to include consciousness as a primary variable — not a dependent outcome but an active dimension of inquiry — several things would change.
The definition of flourishing would expand. Material welfare is certainly necessary, but not sufficient for a completely satisfactory human life. A complete account of what AI should optimize for includes the quality of inner life: the capacity for sustained attention, clarity of perception, the ability to distinguish one's own understanding from borrowed information, the experience of meaning rather than merely the conditions that correlate with it. These are not soft or unmeasurable goods; they are increasingly the subject of serious empirical research in contemplative neuroscience, phenomenology, and cognitive science — a field whose emergence and findings are surveyed in Lutz, Dunne, and Davidson's foundational overview (Meditation and the Neuroscience of Consciousness) and more recent advances documented in the journal Frontiers in Psychology (Contemplative Sciences, 2025). AI systems designed to support them would look different from AI systems designed only to maximize material outcomes.
The science of mental health would deepen. The most intractable problems in psychiatry and psychology — treatment-resistant depression, the chronic relapsing nature of addiction, the persistence of social, political and family trauma across decades — share a common feature: they involve the relationship between the knowing subject and its own contents in ways that purely pharmacological or behavioral interventions cannot fully address. A science of mental health that includes a rigorous model of consciousness — not as mysticism but as precise inquiry into the structure of experience — has access to interventions and understandings that the current framework cannot reach.
The design of AI systems would be governed by a different principle. If the knowing subject is the primary beneficiary of everything AI does, then every AI system should be evaluated not just by what it produces but by what it does to the clarity, independence, and cognitive integrity of the humans who use it. This is the principle at the heart of witness-centered design: the sākṣin is the constant against which all system outputs are measured. An AI system that produces better outputs while degrading the user's own capacity to evaluate those outputs is not, on this account, a better system but a worse one.
The Contribution of Formal Verification
One reason the Advaita framework can be taken seriously as a foundation for AI development — rather than as an interesting cultural artifact — is that it has now been formally verified as a logically consistent system. Matthew Scherf's formalization of Advaita Vedānta in Lean 4, the first machine-verified proof of a non-western philosophical system, establishes that the conceptual architecture Śaṅkara built holds together without internal contradiction. Its 69 axioms, ten modules, and derived theorems have been checked by a proof assistant that does not traffic in cultural sympathy or philosophical fashion. The full formalization is publicly available at github.com/matthew-scherf/Advaita.
This matters because it closes the gap between ancient wisdom and contemporary rigor. The Advaita-derived witness-based framework is not proposed as a supplement to AI development on the basis of tradition or spiritual authority; it is proposed on the basis of conceptual precision — now independently verified — and on the basis of what it can see that the dominant materialist/behaviorist framework cannot.
Amodei is right that AI could compress decades of progress into years. But progress toward what? In domains defined by material metrics, the answer is relatively clear. In domains that involve the quality of conscious experience — which is to say, ultimately, every domain that matters to a living human being — the answer requires a model of consciousness that the current framework does not provide.
That model exists. It has been carefully developed, carefully verified, and carefully applied. What remains is to take it seriously as a foundation — not alongside the materialist framework, but beneath it, as the account of what the human being actually is that all the rest of the vision is meant to serve.
A Complementary Direction
None of this diminishes what Amodei envisions. The defeat of disease, the treatment of mental illness at scale, the extension of expert knowledge to people who currently have no access to it — these are genuine goods, and AI's potential to accelerate them is real.
The argument here is not that Amodei is wrong. It is that his vision is incomplete in a way that matters — and that completion is available. A parallel direction for AI development, grounded in a rigorous model of consciousness rather than a purely materialist account of human welfare, would not replace his vision but would give it what it currently lacks: a foundation adequate to the full depth of what a human being is.
The work of building on that foundation — in interaction design, in mental health research, in the definition of what flourishing actually means — is not waiting for better AI. It is waiting for better questions, better answers. And the tradition that has been asking those questions most precisely, for the longest time, has now been verified to be asking them coherently.
That is a starting point worth taking seriously.
This essay is a companion to Witness-centered Design — A Conscious Foundation for AI. Matthew Scherf's Lean 4 formalization of Advaita Vedānta is available at github.com/matthew-scherf/Advaita. SpecStudio is at specstudio.net.