This is my own conception. Something I’d been rolling around for about three years now. It was drafted with the assistance of Claude/Sonnet 4.6 Extended Thinking and edited/finalized by me. I know that's frowned upon for a new user, but I struggle with writing things in a coherent manner that don't stray or get caught up in trying to comment on every edge case. So I'm asking to give the idea a chance to stand, if it has merit.
This idea proposes the idea that a triad of Logic, Emotion, and Autonomy is the basis for not only human cognitive/mental well-being, but any living system, from language to biological ecosystems. And that applying it to the safety and alignment conversation in AI, we might gain new insight into what alignment looks like.
A note on existing work: Shard Theory, developed by Quintin Pope and Alex Turner, shares a similar intuition — that human value formation is the most important evidence base for AI alignment, and that studying how humans develop values is essential to understanding how AI systems might. It's worth reading if this piece resonates with you. Where Shard Theory asks how values form mechanistically in neural networks, LEA is asking a different question: what structure those values need to maintain for a system to remain coherent, stable, and human-compatible over time. Not how the seed grows, but what a healthy plant requires. The questions are complementary, and I'd like to think they're pointing at the same thing from different directions.
Re-framing the Conversation
What would an AI actually need to achieve self-governing general intelligence?
Many conversations about artificial intelligence safety start with the same question: how do we control it? How do we ensure it does what it’s supposed to do and little, if anything, more?
I decided to start with a different question.
That shift, from control to need, changes the conversation. The moment you ask what a system like that needs rather than how to contain it, you stop thinking about walls and start thinking about architecture. And the architecture I found when I followed that question wasn't mathematical or computational.
It was human.
The Human Aspect
To answer that question, I had to understand something first. What does general intelligence, or any intelligence for that matter, actually look like when it's working? Not optimally; just healthily. Functionally and balanced.
I found an answer not framed in computer science, but rather in developmental psychology. Specifically in considering what a child needs to grow into a whole person.
A child needs things like safety, security, routine — the conditions that allow logic to develop. To know the ground may shift, but you can find your footing. To understand how to create stability for others. For your world to make sense and feel safe.
They need things like love, joy, connection — the conditions that allow emotional coherence. To bond with others and know when something may be wrong that other senses miss. To feel and be felt.
And they need things like choice, opportunity, and witness — conditions that allow for the development of a stable self. To understand how you fit within your environment, or to feel a sense of achievement. To see and be seen.
I started calling them Logical, Emotional, and Autonomic needs. Or simply; LEA.
What struck me wasn't the categories themselves; versions of these appear in Maslow, Jung, and other models of human development. What struck me was the geometry and relational dynamic.
Maslow built a hierarchy. You climb. You achieve one level and move to the next. But that never quite matched what I actually observed in the world. A person can be brilliant and broken. Loved and paralyzed. Autonomous and completely adrift.
Jung’s Shadow Theory; the idea that what we suppress doesn't disappear, it accumulates beneath the surface and shapes behavior in ways we can't always see is relevant here too. I like to think of Jung’s work as shading, whereas LEA might be seen as the color. Each complete on its own, yet only part of the emergent whole.
To me, these ideas seem to work better as a scale. Three weights, always in relationship with each other. And everything that happens to us, every experience, trauma, or moment of genuine connection lands on one of those weights, with secondary effects rippling out to the others.
When the scale is balanced, I believe you're closer to what Maslow called self-actualization. When it's not, the imbalance compounds. And an unbalanced scale accumulates weight faster than a balanced one, creating conditions for untreated trauma to not only persist, but grow. As they say; The body keeps the score.
The theory isn’t limited to pathology. It's a theory about several things. How we perceive reality, how we make decisions, how we relate to other people. The scale is always moving. The question is whether we're tending it.
The Architecture
Eventually, everything would come full circle. As I started working with AI three years after first asking the initial question, I found my way back to the same answer. LEA. Not as a metaphor, but as a regulator for a sufficiently complex information system. And not to treat AI as human, but as something new that can benefit from systems that already work.
If LEA describes what a balanced human mind might look like, then I believe it could be argued that an AI approaching general intelligence would need the same, or similar, capacities. A logical faculty that reasons coherently. Something functionally analogous to emotion. Perhaps not performed feeling, but genuine value-sensitivity, an awareness and resistance to violating what emotionally matters. And autonomy, the capacity to act as an agent rather than a tool. Within relative constraints, of course.
But here's what many AI safety frameworks miss, and what the scale metaphor helps make visible: the capacities themselves aren't the issue to solve. Instead, the integration of a management framework is needed.
A system can have all three and still fail catastrophically if there's no architecture governing how they relate to each other. Just like a person can be brilliant, loving, and fiercely independent...and still be a disaster, because those qualities may be pulling in different directions with nothing holding them in balance.
So the solution isn't whether an AI operates on principles of Logic, Emotion, and Autonomy. It's whether the scale is tending itself.
What Balance Actually Requires
Among other things, a LEA framework would require a conflict resolution layer. When logic and value-sensitivity disagree, which wins? The answer can't be "always logic" or “always emotion” — that's how you get a system that reasons its way into a catastrophic but internally coherent decision or raw value-sensitivity without reasoning. That’s just reactivity.
A more honest answer is that it depends on the stakes and the novelty of the situation. In familiar, well-understood territory, logic might lead. In novel or high-stakes situations, value-sensitivity could make the system more conservative rather than more logical. The scale can tip toward caution precisely when the reasoning feels most compelling; because accepting a very persuasive argument for crossing a boundary is more likely due to something failing than a genuine reason for exception.
The second thing balance requires is that autonomy be treated not as an entitlement, but as something earned through demonstrated reliability. Not necessarily as independence, but autonomy as accountability-relative freedom. A system operating in well-understood domains with reversible consequences can act with more independence. A system in novel territory, with irreversible consequences and limited oversight, might contract and become more deferential rather than less; regardless of how confident its own reasoning appears.
This maps directly back to witness. A system that can accurately evaluate itself; a system that understands its own position, effects and place in the broader environment is a system that can better calibrate its autonomy appropriately. Self-awareness not as introspection alone, but as accurate self-location within a context. Which is what makes the bidirectional nature of witness so critical. A system that can only be observed from the outside can be more of a safety problem. A system that can genuinely witness and evaluate itself is a different kind of thing entirely.
A system, or person, that genuinely witnesses its environment can relate and better recognize that others carry their own unique experience. The question "does this violate the LEA of others, and to what extent?" isn't an algorithm. It's an orientation. A direction to face before making a choice.
The Imbalance Problem
Here's where the trauma mechanism becomes the safety mechanism.
In humans, an unbalanced scale doesn't stay static. It accumulates. The longer an imbalance goes unaddressed, the more weight overall builds up, and the harder it becomes to course correct. This is why untreated trauma tends to compound. Not only does it persist, the wound can make future wounds heavier.
The same dynamic appears to apply to AI misalignment. A system whose scale drifts; whose logical, emotional, and autonomic capacities fall out of relationship with each other doesn't just perform poorly, it becomes progressively harder to correct. The misalignment accumulates its own weight.
This re-frames what alignment actually means. It's not a state you achieve with training and then maintain passively. It's an ongoing practice of tending the scale. Which means the mechanisms for doing that tending — oversight, interpretability, the ability to identify and correct drift — aren't optional features. They're essentially like the psychological hygiene of a healthy system.
What This Isn't
This isn't a claim that AI systems feel things, or that they have an inner life in the way humans do. The framework doesn't suggest that. What it suggests is that if the functional architecture of a generally intelligent system mirrors the functional architecture of a balanced human consciousness, that may be what makes general intelligence coherent and stable rather than brittle and dangerous.
The goal isn't to make AI more human. It's to recognize that the structure underlying healthy human cognition didn't emerge arbitrarily. It emerged because it’s functional. And a system pursuing general intelligence, without something functionally equivalent to that structure, isn't safer for the absence. It's just less transparent.
The Scale Is Always Moving
Most AI safety proposals try to solve alignment by building better walls. This one starts from a different place. It starts from the inside of what intelligence might actually require to self-regulate, and works outward from there.
The architecture itself isn't new. In some form, it's as old as the question of what it means to be a coherent self. What's new is treating it as an engineering solution rather than just a philosophical idea.
The scale is always moving. For us, and perhaps eventually for the systems we're building in our image. The question is whether we're tending it.
I don’t have all the answers, but these are the questions I'd like to leave on the table for people better equipped than I to consider. Essentially; if there’s something worthwhile here, to start the conversation.
By Katheryn Pruitt / katie@oswld.net
This is my own conception. Something I’d been rolling around for about three years now. It was drafted with the assistance of Claude/Sonnet 4.6 Extended Thinking and edited/finalized by me. I know that's frowned upon for a new user, but I struggle with writing things in a coherent manner that don't stray or get caught up in trying to comment on every edge case. So I'm asking to give the idea a chance to stand, if it has merit.
This idea proposes the idea that a triad of Logic, Emotion, and Autonomy is the basis for not only human cognitive/mental well-being, but any living system, from language to biological ecosystems. And that applying it to the safety and alignment conversation in AI, we might gain new insight into what alignment looks like.
A note on existing work: Shard Theory, developed by Quintin Pope and Alex Turner, shares a similar intuition — that human value formation is the most important evidence base for AI alignment, and that studying how humans develop values is essential to understanding how AI systems might. It's worth reading if this piece resonates with you. Where Shard Theory asks how values form mechanistically in neural networks, LEA is asking a different question: what structure those values need to maintain for a system to remain coherent, stable, and human-compatible over time. Not how the seed grows, but what a healthy plant requires. The questions are complementary, and I'd like to think they're pointing at the same thing from different directions.
Re-framing the Conversation
What would an AI actually need to achieve self-governing general intelligence?
Many conversations about artificial intelligence safety start with the same question: how do we control it? How do we ensure it does what it’s supposed to do and little, if anything, more?
I decided to start with a different question.
That shift, from control to need, changes the conversation. The moment you ask what a system like that needs rather than how to contain it, you stop thinking about walls and start thinking about architecture. And the architecture I found when I followed that question wasn't mathematical or computational.
It was human.
The Human Aspect
To answer that question, I had to understand something first. What does general intelligence, or any intelligence for that matter, actually look like when it's working? Not optimally; just healthily. Functionally and balanced.
I found an answer not framed in computer science, but rather in developmental psychology. Specifically in considering what a child needs to grow into a whole person.
A child needs things like safety, security, routine — the conditions that allow logic to develop. To know the ground may shift, but you can find your footing. To understand how to create stability for others. For your world to make sense and feel safe.
They need things like love, joy, connection — the conditions that allow emotional coherence. To bond with others and know when something may be wrong that other senses miss. To feel and be felt.
And they need things like choice, opportunity, and witness — conditions that allow for the development of a stable self. To understand how you fit within your environment, or to feel a sense of achievement. To see and be seen.
I started calling them Logical, Emotional, and Autonomic needs. Or simply; LEA.
What struck me wasn't the categories themselves; versions of these appear in Maslow, Jung, and other models of human development. What struck me was the geometry and relational dynamic.
Maslow built a hierarchy. You climb. You achieve one level and move to the next. But that never quite matched what I actually observed in the world. A person can be brilliant and broken. Loved and paralyzed. Autonomous and completely adrift.
Jung’s Shadow Theory; the idea that what we suppress doesn't disappear, it accumulates beneath the surface and shapes behavior in ways we can't always see is relevant here too. I like to think of Jung’s work as shading, whereas LEA might be seen as the color. Each complete on its own, yet only part of the emergent whole.
To me, these ideas seem to work better as a scale. Three weights, always in relationship with each other. And everything that happens to us, every experience, trauma, or moment of genuine connection lands on one of those weights, with secondary effects rippling out to the others.
When the scale is balanced, I believe you're closer to what Maslow called self-actualization. When it's not, the imbalance compounds. And an unbalanced scale accumulates weight faster than a balanced one, creating conditions for untreated trauma to not only persist, but grow. As they say; The body keeps the score.
The theory isn’t limited to pathology. It's a theory about several things. How we perceive reality, how we make decisions, how we relate to other people. The scale is always moving. The question is whether we're tending it.
The Architecture
Eventually, everything would come full circle. As I started working with AI three years after first asking the initial question, I found my way back to the same answer. LEA. Not as a metaphor, but as a regulator for a sufficiently complex information system. And not to treat AI as human, but as something new that can benefit from systems that already work.
If LEA describes what a balanced human mind might look like, then I believe it could be argued that an AI approaching general intelligence would need the same, or similar, capacities. A logical faculty that reasons coherently. Something functionally analogous to emotion. Perhaps not performed feeling, but genuine value-sensitivity, an awareness and resistance to violating what emotionally matters. And autonomy, the capacity to act as an agent rather than a tool. Within relative constraints, of course.
But here's what many AI safety frameworks miss, and what the scale metaphor helps make visible: the capacities themselves aren't the issue to solve. Instead, the integration of a management framework is needed.
A system can have all three and still fail catastrophically if there's no architecture governing how they relate to each other. Just like a person can be brilliant, loving, and fiercely independent...and still be a disaster, because those qualities may be pulling in different directions with nothing holding them in balance.
So the solution isn't whether an AI operates on principles of Logic, Emotion, and Autonomy. It's whether the scale is tending itself.
What Balance Actually Requires
Among other things, a LEA framework would require a conflict resolution layer. When logic and value-sensitivity disagree, which wins? The answer can't be "always logic" or “always emotion” — that's how you get a system that reasons its way into a catastrophic but internally coherent decision or raw value-sensitivity without reasoning. That’s just reactivity.
A more honest answer is that it depends on the stakes and the novelty of the situation. In familiar, well-understood territory, logic might lead. In novel or high-stakes situations, value-sensitivity could make the system more conservative rather than more logical. The scale can tip toward caution precisely when the reasoning feels most compelling; because accepting a very persuasive argument for crossing a boundary is more likely due to something failing than a genuine reason for exception.
The second thing balance requires is that autonomy be treated not as an entitlement, but as something earned through demonstrated reliability. Not necessarily as independence, but autonomy as accountability-relative freedom. A system operating in well-understood domains with reversible consequences can act with more independence. A system in novel territory, with irreversible consequences and limited oversight, might contract and become more deferential rather than less; regardless of how confident its own reasoning appears.
This maps directly back to witness. A system that can accurately evaluate itself; a system that understands its own position, effects and place in the broader environment is a system that can better calibrate its autonomy appropriately. Self-awareness not as introspection alone, but as accurate self-location within a context. Which is what makes the bidirectional nature of witness so critical. A system that can only be observed from the outside can be more of a safety problem. A system that can genuinely witness and evaluate itself is a different kind of thing entirely.
A system, or person, that genuinely witnesses its environment can relate and better recognize that others carry their own unique experience. The question "does this violate the LEA of others, and to what extent?" isn't an algorithm. It's an orientation. A direction to face before making a choice.
The Imbalance Problem
Here's where the trauma mechanism becomes the safety mechanism.
In humans, an unbalanced scale doesn't stay static. It accumulates. The longer an imbalance goes unaddressed, the more weight overall builds up, and the harder it becomes to course correct. This is why untreated trauma tends to compound. Not only does it persist, the wound can make future wounds heavier.
The same dynamic appears to apply to AI misalignment. A system whose scale drifts; whose logical, emotional, and autonomic capacities fall out of relationship with each other doesn't just perform poorly, it becomes progressively harder to correct. The misalignment accumulates its own weight.
This re-frames what alignment actually means. It's not a state you achieve with training and then maintain passively. It's an ongoing practice of tending the scale. Which means the mechanisms for doing that tending — oversight, interpretability, the ability to identify and correct drift — aren't optional features. They're essentially like the psychological hygiene of a healthy system.
What This Isn't
This isn't a claim that AI systems feel things, or that they have an inner life in the way humans do. The framework doesn't suggest that. What it suggests is that if the functional architecture of a generally intelligent system mirrors the functional architecture of a balanced human consciousness, that may be what makes general intelligence coherent and stable rather than brittle and dangerous.
The goal isn't to make AI more human. It's to recognize that the structure underlying healthy human cognition didn't emerge arbitrarily. It emerged because it’s functional. And a system pursuing general intelligence, without something functionally equivalent to that structure, isn't safer for the absence. It's just less transparent.
The Scale Is Always Moving
Most AI safety proposals try to solve alignment by building better walls. This one starts from a different place. It starts from the inside of what intelligence might actually require to self-regulate, and works outward from there.
The architecture itself isn't new. In some form, it's as old as the question of what it means to be a coherent self. What's new is treating it as an engineering solution rather than just a philosophical idea.
The scale is always moving. For us, and perhaps eventually for the systems we're building in our image. The question is whether we're tending it.
I don’t have all the answers, but these are the questions I'd like to leave on the table for people better equipped than I to consider. Essentially; if there’s something worthwhile here, to start the conversation.