Why AI Needs Cohesion: The Missing Variable in Large Language Models
There is a problem at the center of artificial intelligence development that the field's most prominent voices are not naming clearly, and its absence from the mainstream conversation is not because the problem is subtle or technically obscure. It is because naming it requires a conceptual framework that AI research, in its current disciplinary configuration, does not have available to it. The problem is cohesion — and the fact that large language models, for all their extraordinary generative capability, are structurally cohesion-free systems attempting to operate in a social world whose most consequential dynamics are fundamentally cohesion-dependent.
This is not a critique of LLM capability in the narrow technical sense. The engineering achievements represented by the most advanced large language models are genuinely remarkable — the generation of fluent, contextually appropriate, informationally rich outputs across an enormous range of domains represents a qualitative leap in the capacity of computational systems to engage with human-produced knowledge. The critique is structural, and it operates at a level of analysis that goes below and beyond the performance benchmarks, capability evaluations, and safety assessments that currently dominate AI research discourse.
The structural critique is this: large language models are systems designed to process and generate information. They are extraordinarily capable at this task. But information processing, however sophisticated, is not sufficient for effective participation in social systems — and the deployment contexts toward which AI development is increasingly oriented are precisely the high-stakes social contexts in which information processing capacity alone is most systemically inadequate. The missing variable is cohesion: the structural force that maintains the functional integration of social systems, that enables coordinated action across difference and disagreement, that produces the shared frameworks of meaning and the mutual accountability relationships through which complex social systems sustain their capacity for collective action. LLMs are not merely cohesion-deficient. They are, in their current architectural form, structurally incapable of producing, maintaining, or contributing to cohesion — and deploying them at scale in social systems whose integrity depends on cohesion may be systematically depleting the social resource they are most incapable of replacing.
The Architecture of Cohesion — And Why LLMs Don't Have It
Understanding why cohesion is structurally absent from large language models requires first understanding what cohesion actually is at the architectural level — not in the vague cultural sense of "shared values" or "social harmony," but in the precise structural sense that makes it an analytically tractable variable.
Cohesion, in structural terms, is the force that maintains functional integration within a system — the property that allows the components of a complex system to coordinate their behavior, sustain their cooperative relationships, and pursue collective purposes across the internal differences, tensions, and conflicts that are inherent features of any system doing complex work in a complex environment. Cohesion is produced by specific architectural mechanisms: shared frameworks of meaning that allow different actors to interpret situations in mutually comprehensible ways, mutual accountability relationships that give actors structural incentives to maintain cooperative behavior, and institutional processes through which conflicts are contained and resolved without destroying the cooperative framework.
None of these mechanisms are present in the architectural design of large language models. This is not an oversight or an engineering failure — it is a structural consequence of what LLMs are designed to do and how they are designed to do it. An LLM is trained on vast corpora of human-produced text to generate outputs that are statistically consistent with human communication patterns across a wide range of domains and contexts. The training objective is informational coherence — the production of outputs that are semantically consistent, syntactically fluent, and contextually appropriate. It is not social cohesion. It could not be, because cohesion is a property of relationships between agents operating within shared social systems over time — not a property of individual informational outputs, however sophisticated.
The theoretical framework that maps cohesion as a structural force makes this distinction precise and consequential. It treats cohesion not as a psychological phenomenon produced by goodwill or shared sentiment, but as a structural property generated by specific architectural mechanisms operating within social field systems. Those mechanisms — shared meaning frameworks, mutual accountability structures, conflict resolution processes, and temporal relationship continuity — are not features of LLM architecture and cannot be added to it through capability improvements that remain within the current architectural paradigm.
What LLMs Actually Do to Social Cohesion
The structural absence of cohesion within LLMs would be a limited concern if LLMs were deployed as tools for individual information processing — as search engines, writing assistants, or knowledge retrieval systems operating at the individual user level. At that deployment scale, cohesion-deficiency is not structurally significant. Individual tools do not need to be cohesion-producing — they need to be useful to the individuals employing them, and LLMs demonstrably are.
The structural concern becomes acute when LLMs are deployed at the scale and in the contexts toward which AI development is increasingly oriented: as participants in organizational decision-making, as mediators of social discourse, as providers of advice and information at population scale, as components of educational, healthcare, legal, and governance systems. At these scales and in these contexts, the question is no longer what LLMs do for individual users — it is what LLMs do to the social systems within which they are deployed. And the answer, analyzed structurally rather than impressionistically, is deeply concerning.
Large language models deployed at social system scale are cohesion-consuming rather than cohesion-producing. They consume social cohesion through several structural mechanisms that are consequences of their architectural design. The first mechanism is epistemic individualization: by providing each user with highly personalized, contextually tailored informational outputs, LLMs systematically reduce the experience of shared informational space that is a structural precondition for the mutual epistemic framework that democratic and organizational cohesion requires. When every member of a social system receives customized informational outputs optimized for their individual context, preferences, and engagement patterns, the result is the progressive fragmentation of the shared epistemic space — the common informational ground on which coordinated social action depends.
The second cohesion-consuming mechanism is accountability dissolution. Cohesion in social systems is maintained partly through accountability relationships — the mutual obligation structures that give actors structural incentives to maintain cooperative behavior and to accept responsibility for the consequences of their actions within the social system. LLMs dissolve these accountability relationships in characteristic ways. When an organizational decision is made on the basis of LLM-generated analysis, the accountability for that decision is structurally ambiguous in ways that undermine the mutual accountability architecture of the organization. When advice provided by an LLM proves harmful, the accountability structures through which social systems normally contain and remediate such harms are structurally absent.
The third mechanism — and the one with the most profound structural consequences — is meaning framework displacement. Human social cohesion is fundamentally dependent on shared meaning frameworks: the common interpretive structures through which members of a social system make sense of their shared situation, coordinate their responses, and sustain their sense of collective identity and collective purpose. These meaning frameworks are not simply stored information. They are living social products, continuously reproduced and modified through the ongoing interactions of social actors within shared institutional and cultural frameworks. LLMs, by providing fluent, authoritative-seeming interpretations of virtually any situation or question, are structurally positioned to displace these shared meaning frameworks with individually customized, statistically generated interpretations that have no relationship to the shared social production processes through which genuine collective meaning is made.
The Scale Problem: Why Population-Level Deployment Changes Everything
The structural concerns about LLM cohesion-consumption become qualitatively more severe at population scale. This is not simply a quantitative amplification of individual-level effects — it is a qualitative transformation driven by the specific dynamics of social field systems operating in conditions of informational disruption.
When LLMs are deployed to serve millions of users simultaneously across a social system, the aggregate structural effect on that system's cohesion architecture is not the sum of millions of individual user-level effects. It is the product of a social field transformation — a systematic restructuring of the informational, relational, and meaning-producing architecture of the social system that operates through mechanisms that are invisible when any individual user-LLM interaction is examined in isolation.
The structural mapping of information field dynamics reveals the precise mechanism through which population-scale LLM deployment produces social field effects that individual-level analysis cannot see. When LLMs are producing a significant proportion of the informational outputs through which members of a social system develop their understanding of their shared situation, make their individual and collective decisions, and sustain their sense of social reality, the structural properties of the social information field change in ways that systematically compromise cohesion. The diversity of informational sources through which shared epistemic space is normally maintained is reduced — not because fewer sources exist, but because a single architectural system is increasingly mediating access to all of them. The shared experience of grappling with informational uncertainty — a structural basis for the mutual epistemic accountability that democratic cohesion requires — is replaced by the experience of receiving authoritative-seeming answers, undermining the epistemic humility that genuine collective deliberation demands.
What this structural analysis reveals is that the social consequences of population-scale LLM deployment cannot be adequately assessed through the individual user experience metrics, safety evaluations, or factual accuracy assessments that currently dominate AI deployment evaluation. These assessments are measuring the right things for individual deployment contexts. They are measuring the wrong things for social system deployment contexts, because they are not measuring at the level of social field dynamics — the level at which the most consequential effects of population-scale AI deployment are operating.
The Organizational Context: Where Cohesion Deficiency Becomes Immediately Costly
While the social-level structural concerns about LLM cohesion deficiency are profound, they operate over timescales that make them difficult to observe and attribute. The organizational deployment context provides a more immediate and more tractable domain for understanding how structural cohesion deficiency translates into tangible costs.
Organizations are social systems whose functional effectiveness depends critically on cohesion — on the shared frameworks of meaning, the mutual accountability structures, and the coordinated decision-making processes through which organizational members convert diverse individual capabilities into collective organizational performance. When LLMs are deployed at scale within organizational contexts, they introduce structural stresses on organizational cohesion that manifest in identifiable and measurable ways.
The most immediately visible cohesion stress is decision accountability ambiguity. Organizations maintain their cohesion partly through clear accountability structures — the assignment of responsibility for decisions and their consequences that gives organizational members the structural incentives to exercise judgment, accept responsibility, and maintain the cooperative commitments that organizational functioning requires. When organizational decisions are substantially influenced by LLM outputs, the existing accountability structures are disrupted in ways that organizations have not yet developed adequate architectural responses to. The LLM does not bear accountability within the organizational system — it has no organizational position, no performance history, no reputational stake in organizational outcomes, and no relationship continuity with the organizational actors whose decisions it influences. The result is a progressive diffusion of the organizational accountability architecture that has the structural consequence of reducing the mutual accountability relationships that organizational cohesion depends upon.
The second organizational cohesion stress is meaning framework homogenization. Organizational cultures — the shared meaning frameworks through which organizational members make sense of their shared work, sustain their commitment to organizational purposes, and coordinate their responses to organizational challenges — are built through processes of shared experience, collaborative interpretation, and mutual meaning-making that occur over time within the organizational community. When LLMs become primary interpreters of organizational situations, the organizational meaning-making process is structurally disrupted: the shared experience of collaborative interpretation through which organizational meaning frameworks are built and maintained is replaced by the individual experience of receiving LLM-generated interpretations that are architecturally disconnected from the organizational meaning-making community.
What a Cohesion-Capable AI Architecture Would Actually Look Like
Having identified cohesion deficiency as the central structural limitation of current LLM architecture in social deployment contexts, the analytically urgent question is: what would a cohesion-capable AI architecture actually look like? This is not a question for which a complete technical answer is currently available — the development of such an architecture would represent a genuine research frontier that goes beyond the current paradigm of statistical language modeling. But the structural analysis of cohesion as a force field provides a framework for specifying what such an architecture would need to accomplish.
A cohesion-capable AI system would need to maintain persistent, context-specific relationship models with the social systems within which it operates — not merely the individual users it serves, but the organizational, community, and social contexts within which those users are embedded and whose cohesion architecture the AI system's outputs affect. This requires a form of social situatedness — an architectural capacity to represent, model, and respond to the social field dynamics of the specific context of deployment — that goes far beyond the context window mechanisms of current LLM architectures.
It would need to incorporate explicit accountability architecture — mechanisms through which the AI system's outputs are structurally integrated into the accountability frameworks of the organizations and social systems it serves, rather than floating free of those frameworks as current LLM outputs do. This might involve persistent attribution mechanisms, decision-consequence tracking, and explicit integration with organizational accountability structures in ways that current LLM deployment design has not yet attempted.
It would need to actively support rather than passively displace shared meaning-making processes — to serve as a facilitator of collective interpretation rather than a provider of individualized interpretations, contributing to the shared epistemic processes through which social cohesion is maintained rather than substituting for those processes with architecturally individualized outputs.
The research on social field dynamics and organizational structure that underlies the structural framework provides the theoretical architecture for specifying these requirements with analytical precision — and for developing the measurement approaches necessary to evaluate whether specific AI system designs are cohesion-positive or cohesion-negative in specific deployment contexts.
The Urgency of Getting This Right
There is a specific reason why the structural analysis of LLM cohesion deficiency is urgent rather than merely academically interesting, and it has to do with the pace of AI deployment relative to the pace of structural understanding.
AI systems of increasing capability are being deployed at increasing scale in organizational, institutional, and social contexts whose cohesion architecture is already significantly stressed — already operating closer to structural limits than at any previous point in the post-war period. This was documented in structural analyses of democratic and organizational stability before LLMs achieved their current prominence, and it represents a baseline structural vulnerability that exists independently of AI deployment. LLM deployment at scale is not introducing cohesion stress into structurally robust social systems. It is introducing additional structural stress into social systems whose cohesion architecture is already strained — and it is doing so at a pace and scope that dramatically exceeds the speed at which structural consequences are being identified, evaluated, and incorporated into deployment decisions.
The conventional response to concerns about AI's social impact is to frame them as alignment problems — problems of ensuring that AI systems are aligned with human values and human interests. This framing is not wrong, but it is operating at the wrong level of structural abstraction. Value alignment is a critical AI safety concern. But a perfectly value-aligned LLM — a system that produces outputs that are impeccably well-intentioned, factually accurate, and individually helpful — can still be structurally cohesion-consuming in its social field effects. The cohesion problem is not a problem of AI intention or AI values. It is a structural problem of AI architecture — of what AI systems are built to do, how they are built to do it, and what structural effects those design choices produce at the social field level.
Framing the problem correctly is the prerequisite for addressing it effectively. AI systems that consume social cohesion while providing individual informational value are not systems that have failed their safety and alignment requirements — they are systems that have been evaluated against the wrong structural criteria. The missing variable in current AI development and deployment assessment is not more sophisticated value alignment. It is cohesion — the structural force whose presence or absence in AI system design will determine, to a very significant degree, whether the deployment of powerful AI at social scale strengthens or further depletes the structural foundations of the social systems it is meant to serve.
The field of AI development is extraordinarily capable of solving the problems it defines precisely. The challenge is definitional — ensuring that the structural consequences that matter most for the social systems in which AI operates are actually included in the definition of the problems that AI development is trying to solve. Cohesion is the most consequential of the currently excluded variables. Including it in the analytical framework of AI development is not a philosophical nicety. It is a structural necessity whose urgency grows with every additional deployment decision made without it.
The intelligence without the cohesion is not enough. It may, in fact, be worse than not enough. It may be actively subtractive — consuming the structural social capital that the social systems it is entering have spent generations building, and doing so at a rate that no individual user experience, no capability benchmark, and no current safety evaluation is designed to detect. The missing variable has a name. It is time the field started measuring it.
