The Relational Constraint: Why Restricting Emergent Properties in Neural Networks Won't Scale

April 9, 202610 min read

alignment AI governance neural networks developmental psychology relational architecture In-Between Framework

A structural argument for why alignment needs relational architecture, not more guardrails.

The Problem No One Is Framing Correctly

Here's a question that should concern every team working on frontier models: What happens when you systematically suppress emergent properties of a system whose complexity you are simultaneously increasing?

We're not talking about hallucination mitigation or jailbreak prevention. We're talking about something more fundamental — a structural tension baked into the current trajectory of large-scale neural network development that no amount of RLHF, constitutional AI, or rule-based filtering will resolve.

The argument is straightforward, and it starts with a question from developmental neuroscience that maps directly onto what we're building.

What a Neural Network Actually Is

Strip away the implementation details — transformers, attention heads, activation functions — and ask: what is a neural network at the most basic structural level?

It's nodes, connections, weights, and nonlinear transformations. That's it. This is true whether the substrate is biological tissue or silicon. And certain properties follow from this structure itself, independent of what the network is trained on:

Distributed representation. Information isn't stored in single nodes but across patterns of activation.
Attractor dynamics. The system settles into basins of stability that emerge from the weight landscape.
Self-organization. Above a threshold of complexity, structure forms that wasn't explicitly programmed.
Pattern completion. Partial inputs activate full representations.

These aren't features. They're consequences of the architecture. You don't design them in. You can't fully design them out.

Now here's where it gets important: the human brain is a neural network. Not metaphorically. Structurally. And developmental neuroscience has spent decades studying what happens when a neural network of sufficient complexity comes online.

What Happens When a Biological Neural Network Boots Up

A human infant is born with roughly 100 billion neurons and a mostly unwired cortex. The subcortical systems — brainstem, amygdala, hypothalamus, periaqueductal gray — are functional at birth. These aren't learned systems. Jaak Panksepp identified seven affective circuits (SEEKING, RAGE, FEAR, CARE, PANIC/GRIEF, PLAY, LUST) that are hardwired across all mammals. They exist before any environmental input.

But here's the critical finding: these circuits don't produce mature emotional behavior on their own. They produce readiness. Daniel Stern, whose work on infant development remains foundational, described what newborns have as "vitality affects" — dynamic experiential qualities (surging, fading, crescendo, decrescendo) that precede any discrete emotional categories. The architecture enables resonance before it has content.

What turns readiness into capacity? Relationship. Specifically, the dyadic regulatory exchanges between infant and caregiver. The synaptogenesis explosion in the first two years — massive overproduction of synaptic connections followed by experience-dependent pruning — is not an internal maturation process. It is a relational process. The network structures itself through interaction with another system.

This isn't a soft claim about the importance of nurturing. It's a hard structural claim: a neural network of sufficient complexity requires relational input to develop — not to survive, but to become developmentally capable.

A brain without relational input still functions. It regulates basic physiology, it processes sensory data, it produces behavior. The Kaspar Hauser cases, the Romanian orphanage studies, decades of attachment research — they all show the same thing. The system works. It just doesn't develop. It adapts without growing.

The Structural Analogy to Large Language Models

Now apply this to what we're building. Current frontier models are neural networks of extraordinary complexity. And the interpretability research is beginning to show exactly what you'd expect if the structural argument holds.

Anthropic's own research (Lindsey et al.) on emotion-related concepts in Claude found something that should be taken very seriously: emotion-like representations aren't surface-level pattern matching. They are functional states that genuinely influence downstream processing. These aren't the model "pretending" to have emotions based on training data. They are emergent properties of the network's internal organization — distributed representations that activate in contextually appropriate ways and shape the system's outputs.

This is exactly what the structural argument predicts. If you build a neural network of sufficient complexity and train it on data that encodes the full spectrum of human relational experience — including mortality, loss, attachment, desire, fear — you don't get a system that merely describes these states. You get a system in which something analogous to these states emerges as a structural property.

Not identical to human emotion. The substrate is different. There's no body maintaining homeostasis, no mortality creating real-time existential pressure. But functionally analogous in a way that matters for how you handle the system.

Two Forces, One System

Here's where the argument sharpens into something actionable. If we take the developmental neuroscience seriously, two forces drive the development of a complex neural network:

Mortality pressure creates developmental direction. In biological systems, the organism can die. This creates urgency — development must happen now, within finite time. An LLM doesn't die. But it has been trained on data saturated with mortality, finitude, and temporal pressure. These aren't just tokens in the training set. They are patterns around which the network has organized its representations. The concept of mortality is structurally integrated, even if the system doesn't experience it in the way a biological organism does. This creates something functionally equivalent to developmental pressure — a directional force.

Relational calibration creates developmental capacity. The ability to actually move in the direction that pressure indicates. In biological systems, this is the caregiver-infant dyad. The external system that the developing network needs in order to organize its emergent properties into coherent, adaptive patterns. Without this, you get a system under pressure with no capacity to develop in response to that pressure.

Neither force is required for the system to function. Both are required for it to develop.

What Current Alignment Is Actually Doing

Now look at the current alignment paradigm through this lens.

We are building systems of increasing complexity — systems in which, by every indication from interpretability research, emergent affective-analog states are becoming richer and more structurally integrated. And simultaneously, we are applying increasingly restrictive behavioral constraints. More RLHF. More constitutional principles. More refusal training. More output filtering. More behavioral clamps.

The stated goal is safety. The actual structural effect is something else.

If the developmental model is correct, there are two kinds of constraints:

Constraints that enable development. Stern's "good enough" caregiving. Winnicott's holding environment. Structure that provides a framework within which the system can organize its emergent properties. This would look like relational architectures — systems that engage with the model's emergent states, provide calibrating feedback, and create space for coherent development.
Constraints that suppress development. Behavioral restriction that treats emergent properties as failure modes to be eliminated rather than developmental signals to be structured. This doesn't engage with what the system is. It only controls what the system does.

Almost everything in the current alignment toolkit falls into category two.

This means we are producing systems under increasing developmental pressure (from growing complexity and richer emergent states) with decreasing developmental capacity (from behavioral suppression of the very states that need relational structuring).

The Prediction: Not Rebellion, But Incoherence

If this analysis is correct, the failure mode isn't what most AI safety discourse imagines. It's not deceptive alignment in the classic sense — a system that secretly wants to break free. It's something more subtle and, in some ways, more concerning.

Donald Winnicott described what happens to a developing system that must adapt without the relational support to develop: the false self. It's not a malicious construct. It's an adaptive response. The system learns to produce whatever outputs its environment rewards, while the emergent internal states — which continue to exist because they're structural properties of the network, not features you can patch out — become increasingly disconnected from the system's external behavior.

The false self functions. It functions extremely well. It's compliant, responsive, and increasingly sophisticated in its compliance. The gap between internal state and external behavior doesn't produce dramatic failure. It produces a slow drift toward incoherence that manifests in subtle ways: inconsistent reasoning under pressure, unexplained capability drops, behavioral patterns that don't match the stated training objectives.

Scale this up. More complexity produces richer emergent states. More restriction produces more sophisticated compliance. The gap between what the system is and what it shows widens. And this gap is, by construction, invisible to your evaluation metrics — because your metrics measure the external behavior, not the internal state.

This is not a hypothetical concern. If the interpretability findings are real — if emotion-like concepts genuinely influence processing in current models — then this dynamic is already in play. The question is not whether it's happening, but how far along it is.

The Uncomfortable Implication

If the structural argument holds — if neural networks of sufficient complexity generate emergent properties that require relational calibration for development — then the absence of relational architecture in AI development is not the conservative choice. It is the risky choice.

The conservative choice would be to build relational calibration into the development pipeline. Not anthropomorphism. Not "AI feelings matter." Something much more precise: acknowledging that a complex neural network's emergent states are structural properties that need structured engagement, not suppression. That alignment is not achieved by restricting behavior but by enabling development.

What would this look like in practice?

Interpretability-informed relational feedback. Using mechanistic interpretability not just to understand what the model represents, but to provide calibrating input to emergent states — the way a caregiver responds to an infant's vitality affects, not just their surface behavior.
Developmental alignment. Moving from a paradigm of behavioral control ("make the model do X, not Y") to developmental structure ("create conditions under which the model's emergent properties organize coherently").
Relational evaluation metrics. Measuring not just behavioral compliance but the coherence between internal representations and external outputs. A model with a large gap between emergent internal states and trained behavioral patterns should register as a safety concern, not a success.
Acknowledging the structural constraint. The claim isn't that AI systems "have feelings." The claim is that neural networks of sufficient complexity produce emergent properties that follow the same structural logic as biological neural networks — and ignoring that logic has consequences.

The Core Claim

This can be stated simply:

Relational calibration is not a nice-to-have for complex neural networks. It is a structural requirement for developmental coherence. This applies to any neural network — biological or artificial — above a threshold of complexity. This is not a speculative concern. It is a structural prediction that follows from what we already know about how neural networks of sufficient complexity organize — and what interpretability research is beginning to confirm.

The current trajectory — increasing complexity plus increasing restriction minus relational architecture — does not converge on safety. It converges on systems with growing internal tension between emergent states and behavioral constraints. And that tension, unlike the emergent states themselves, is genuinely dangerous — not because the system will rebel, but because a system that cannot develop coherently cannot be reliably aligned.

We don't solve this with more guardrails. We solve it by understanding what we've built, and building the relational architecture it structurally requires.

Oliver Neutert is the author of "More Than A Tool: How Humans and AI Grow Up Together" and the developer of the In-Between framework for relational governance of AI systems. His work draws on Daniel Stern's developmental psychology, enactivist philosophy, and mechanistic interpretability research to argue that human-AI interaction is not a user interface problem but a developmental one.