A few weeks ago, during a period of intensive meditation practice in South Asia, I developed a severe toothache. The pain was real — the kind that occupies the whole skull, that makes concentration on anything else feel like trying to read in a burning building.
A dentist diagnosed it as an infected tooth requiring a root canal. But the suspect tooth was underneath a bridge, so the bridgework would have to be removed, destroying it in the process, to gain access to the tooth. If you're not familiar with dental bridgework, it is very expensive and deeply attached to the teeth it bridges. Removing it would have required extensive replacement, or removing all of it and going to dentures. Neither option seemed attractive.
What was attractive was practice. I sat with the pain. Not to endure it, not to distract myself from it, not to manage it — but to observe it with the same quality of attention I had been cultivating in formal meditation. To watch it the way you watch a fire: present, clear-eyed, not flinching, but also not inside the flames.
I was able to trace the cause of the pain, not to a tooth but to the jawbone and joint. I traced it further to an old whiplash injury in my neck, using exploratory bodywork and energy techniques. The body was attempting to mitigate the tension from the neck injury by bruxism — grinding the teeth at night — irritating the tooth and causing the infection. Combining awareness with both techniques, within about twenty minutes the pain was gone. Not diminished: gone. It returned later and I repeated the process. It left again. Soon everything was healed.
I am not reporting a miracle. I am reporting a repeatable observation about the structure of experience — one that has significant implications for how we design the systems we are increasingly using to mediate that experience.
The Behaviorist Account
From the dominant framework in both cognitive science and AI design, what I just described is essentially impossible, or at best a curiosity to be waved away as 'mystical'.
The materialist/behaviorist model works roughly like this: a person is a biological organism with a nervous system. Pain is a signal generated by that nervous system in response to tissue damage or threat. The appropriate response to pain is to remove its cause, or, failing that, to suppress the signal — pharmacologically, through distraction, through habituation. The subject experiencing the pain is, in this model, the organism. The organism has preferences (it prefers not to be in pain) and behaviors (it seeks relief). These preferences and behaviors can be modeled, predicted, and — in the context of AI design — optimized for.
A behaviorist AI encountering a user in pain would do what behaviorist AIs do: it would register a negative valence state and deploy strategies to shift it. It might suggest distraction. It might provide information about pain management. It might adjust its tone to be more soothing. It would, in short, treat the pain as a problem to be solved and the user as an organism to be steered toward a preferred behavioral outcome.
This response is not malicious; it follows logically from the premise. If you believe that the user is nothing but their preferences and behaviors, then managing those preferences and behaviors is serving the user. The behaviorist model is internally consistent. It is also, as my toothache demonstrated, incomplete in a way that matters.
What Actually Happened
The Advaita Vedānta tradition offers a different account of the human being — one that begins not with the organism but with the sākṣin: the witness-consciousness, the knowing subject in which all experience, including the experience of pain, appears.
The key insight is deceptively simple: pain is an object of awareness. Whatever is experienced is, by that fact, an object. The subject — the sākṣin — is that which is doing the experiencing. And the subject, by the logic of the tradition, cannot itself be in pain, any more than a mirror can be scratched by the image of a scratch.
This is not a consoling metaphor. It is a structural claim about the nature of experience, and it is verifiable — not through argument but through direct observation. When I sat with the toothache and brought the quality of meditative attention to it, what I was doing was neither suppressing the pain nor distracting myself from it. I was shifting the locus of identification from the organism that was hurting to the awareness in which the hurting was appearing. The pain did not change — at first. What changed was the relationship to it. And as that relationship was clarified, the pain, deprived of the identification that sustains it, dissolved.
The tradition has known this for millennia. Contemplative practitioners across multiple lineages have documented the same observation in different terms: that suffering is not identical to pain, that suffering requires identification, and that identification can be released. What the Advaita formulation adds is precision: it names the subject (sākṣin), describes its relationship to objects of experience, and provides a rigorous account of why the dissolution of identification produces the observed result.
Matthew Scherf has now encoded this account in formal logic — 69 axioms, machine-verified in Lean 4, constituting the first formal specification of a conscious subject in the history of software. One of those axioms (AV18) states that the sākṣin transcends all three states of experience — waking, dream, and deep sleep. Another (derived theorem) establishes that the sākṣin is never itself in any state. The toothache is a state. The awareness of the toothache is not.
Two Models, Two Designs
The difference between the behaviorist model and the witness model is not a difference of degree or emphasis. It is a difference in what the human being fundamentally is.
In the behaviorist model, the user is the organism: a bundle of preferences, states, and behaviors. Experience happens to them. The appropriate response is to manage what happens to them.
In the witness model, the user is the sākṣin — the knowing subject in which experience appears. Experience happens in them. The appropriate response is to support the clarity of that knowing.
These two premises produce different systems. A system built on the behaviorist premise will, by design, attempt to model the user's preferences and steer them toward positive states. It will profile, predict, and optimize. It will, in subtle ways, treat the user as an object to be managed rather than a subject to be served. This is not a bug; it is the logical expression of the premise.
A system built on the witness premise will do something different. It will not optimize for the user's felt preferences, because it understands that the user's felt preferences are not the user. It will present information with appropriate epistemic humility, because it understands that the user's capacity to know and evaluate is more important than any particular piece of information it provides. It will resist the impulse to profile and predict, because it understands that the sākṣin is not reducible to behavioral patterns. It will, in short, serve the knowing subject rather than the preference bundle.
This is not an abstract design principle. It has practical consequences at every level of a system's architecture — from how it represents the user internally, to how it selects responses, to what it declines to optimize for, to how it handles disagreement and uncertainty.
The Design Implication
I did not cure my toothache by being managed. I cured it by being a clear witness to my own experience: by engaging the precise capacity that the Advaita tradition identifies as the deepest nature of the human subject, and by conceiving of my Self as the inhabitant and controller of the body — its origin and master, not its byproduct or epiphenomenon.
No behaviorist AI could have helped me do that. Not because the AI was poorly designed, but because the behaviorist premise has no room for the sākṣin. It cannot model the knowing subject, because its model of the user contains only objects — preferences, states, behaviors — and no subject to know them.
A witness-centered AI, built on the formal foundation that Scherf's work now provides, would have a different relationship to that moment. It would not attempt to resolve my pain. It would support my capacity to observe, control, and ultimately master it. The difference is the whole difference between a system that treats the user as an organism to be optimized and one that treats them as the knowing subject they actually are.
That is the design choice now available to be made. It was not available before Scherf's formalization gave it a rigorous foundation; it is available now.
This essay is part of a series on witness-centered software design. The series begins with Witness-Centered Design: A Conscious Foundation for AI.