Beyond AGI: How Frontier Models Are Computationalizing Distributed Intelligence

Abstract

In “AGI Is a Category Error,” I argued that Artificial General Intelligence rests on an overly individualistic conception of intelligence, treating it as a property contained within a human mind and therefore reproducible inside a machine. This article asks the constructive question that follows: If AGI is the wrong target, how should we understand the extraordinary capabilities of frontier models? I argue that frontier models are computationalizing distributed intelligence by performing representational, analytical, generative, and coordinative work that was previously distributed across people, documents, tools, software, and institutions. Rather than proposing a wholly separate theory, the article applies and extends the established framework of distributed cognition to contemporary human-AI systems. A published human-technology model provides the conceptual core: biological cognitive capacities change slowly, external data and technology expand rapidly, and consequential intelligence emerges from the coupled system (Zhang & Fenton, 2024). Frontier models transform the technological component from primarily supportive cognitive artifacts into active computational participants. The central question is therefore not whether a model crosses an AGI threshold, but how human-AI distributed cognitive systems allocate cognitive work, validate outputs, preserve human judgment and authority, learn, and remain accountable.

1. Framing the Problem

1.1 From the Critique of AGI to a Constructive Account

In my earlier Substack essay, “AGI Is a Category Error,” I argued that Artificial General Intelligence points toward the wrong object of analysis (Zhang, 2026a). AGI is usually framed as the effort to construct an artificial agent with sufficiently broad, transferable, and autonomous capabilities to equal or surpass human intelligence. This framing assumes that intelligence is fundamentally something an individual possesses. Human beings possess it biologically. Machines are gradually acquiring it computationally. Progress can therefore be measured by placing humans and machines on a common scale and waiting for the machine to cross a threshold called “human level.”

This conception mistakes one historically narrow model of intelligence for intelligence itself. It inherits the architecture of the examination hall and the intelligence test: one individual, separated from tools and other people, solving standardized problems under controlled conditions. Intelligence becomes what a person can remember, reason through, or produce alone.

That is not how consequential intelligence operates in science, medicine, engineering, organizations, governments, or civilization more broadly. Human intelligence has always extended beyond the individual mind. It is distributed across language, writing, mathematics, representations, tools, instruments, cultural knowledge, institutions, and other people. Scientific discovery, clinical judgment, engineering, organizational leadership, and public governance emerge from organized cognitive systems, not isolated minds.

AGI becomes a category error when it treats the intelligence of an isolated individual as the universal form of intelligence and defines the replication of that form inside a machine as the inevitable destination of AI. The problem is not that AGI is too ambitious. The problem is that it targets the wrong thing.

1.2 The Constructive Question

Rejecting AGI as the dominant framework creates a more important question:

If AGI is the wrong target, how should we understand the extraordinary things frontier models are doing now?

The answer cannot be that frontier models are merely sophisticated autocomplete systems. That description fails to capture the breadth and significance of their emerging capabilities. Frontier models can write and revise software, synthesize large bodies of research, interpret multimodal information, generate hypotheses, analyze complex data, compare explanations, produce plans, critique outputs, execute code, use external tools, and participate in scientific, clinical, educational, and organizational workflows.

These are extraordinary developments. But extraordinary capability does not require us to interpret frontier models as synthetic persons approaching a singular threshold called AGI.

A different interpretation is both more precise and more consequential:

Frontier models are computationalizing parts of intelligence that have historically been distributed across humans, representations, artifacts, tools, workflows, and institutions.

They are not placing the whole of intelligence inside a machine. They are moving more of the work performed by distributed cognitive systems into an active computational medium.

1.3 The Central Thesis and Terminological Distinction

The argument of this article is as follows:

Frontier models are not creating a complete, self-contained general intelligence inside a machine. They are computationalizing distributed intelligence and, in doing so, reorganizing the distributed cognitive systems through which intelligence is produced.

By computationalizing distributed intelligence, I mean converting parts of the cognitive work previously performed across people, documents, representations, software, tools, and workflows into processes that can be performed within computational systems. This includes representing and reorganizing information, translating among language, images, data, and code, retrieving and synthesizing knowledge, generating hypotheses and alternatives, comparing explanations, simulating outcomes, critiquing proposed answers, planning sequences of activity, calling tools, and coordinating portions of complex workflows.

Two related terms should be distinguished. Distributed cognition is the established theoretical framework for analyzing how cognitive processes are organized across people, internal and external representations, artifacts, environments, and social arrangements. Distributed intelligence is the effective capability that emerges from such an organized cognitive system. Frontier models computationalize more of the cognitive work within distributed cognition and thereby change the scale, speed, scope, and character of the distributed intelligence that the system can produce.

The cognitive system remains distributed. But a growing share of its representational, analytical, generative, and coordinative work is now carried out through an active computational medium.

This is why frontier models are historically significant. They differ from earlier cognitive artifacts not merely in scale, but in function. A text stores information. A diagram displays relationships. A database retrieves records. Conventional software executes specified procedures. A frontier model can generate, transform, interpret, compare, evaluate, and respond. It does not simply support the cognitive process. It increasingly participates in advancing it.

2. The Individualistic Ontology of AGI

2.1 The Machine Gets an Infrastructure. The Human Gets a Test.

The AGI framing obscures this transformation because it begins with a structurally distorted comparison. A frontier model is supported by vast computational infrastructure, immense training datasets, retrieval systems, search, code execution, specialized software, external tools, engineered interfaces, and extensive human feedback. It is then compared with an individual person sitting alone in an examination setting.

The machine gets an infrastructure.

The human gets a test.

This comparison inherits a century-old psychometric image of intelligence. Psychometrics made major scientific contributions by identifying meaningful patterns in human cognitive performance. Spearman’s work, for example, demonstrated correlations across different cognitive tasks and introduced the concept of a general factor of intelligence (Spearman, 1904).

Psychometrics also helped institutionalize a particular ontology. Intelligence became a quantity located inside the individual and inferred from standardized performance. Artificial intelligence inherited that ontology. The model became another test taker. Benchmarks became the examination hall. Scores became evidence of intelligence. AGI became the point at which the machine supposedly passed the human threshold.

Much of the AGI debate is therefore psychometrics for machines.

2.2 The Ambiguity of “Human-Level Intelligence”

The phrase human-level intelligence sounds precise until the comparison group is specified. Does it refer to the median adult, a highly educated adult, a trained professional, the best specialist in a field, a multidisciplinary team, a university, a corporation, or a national laboratory? Does it refer to a person working alone, or a person using colleagues, databases, software, instruments, and institutional resources?

These comparisons are not equivalent.

AGI is also multidimensional. Breadth, depth, transfer, robustness, learning efficiency, calibration, autonomy, and long-horizon reliability are related but distinct properties. No single benchmark captures all of them (Morris et al., 2024).

AGI can remain useful as a bounded engineering construct when the task distribution, comparison group, permitted resources, novelty requirements, reliability thresholds, and failure conditions are explicitly defined. It becomes the wrong analytical target when it is treated as a complete theory of intelligence or as the inevitable endpoint of AI development.

2.3 Capability Is Not the Whole of Intelligence

Benchmarks measure real capabilities. A model that performs well in mathematics, coding, science, language, or professional examinations possesses capabilities that matter. Those achievements should not be dismissed.

The mistake is treating such performance as evidence that the complete phenomenon of intelligence has been reproduced inside the model.

A model can pass a professional examination without possessing the judgment of a professional. It can generate a persuasive recommendation without understanding the institutional consequences of acting on it. It can optimize an objective without determining whether that objective deserves pursuit. It can produce compassionate language without establishing that it experiences compassion. It can contribute causally to a decision without possessing legitimate authority to make that decision. It can influence an outcome without being morally or institutionally accountable for the result.

Performance is evidence of capability. It is not evidence that the whole of intelligence has been reconstructed inside the machine.

3. Distributed Cognition as the Theoretical Foundation

Distributed cognition is not a new theory invented to explain frontier AI. It is an established cognitive-science framework that changes the unit of analysis from the isolated individual to the functionally organized system. Its central claim is not that brains cease to matter, but that the brain alone is often insufficient to explain observed cognitive performance.

The framework rests on several core principles. First, cognition can be distributed across individuals, social groups, internal representations, external representations, artifacts, physical environments, and time. Second, external representations do not merely store the results of thought. They can transform the structure of a task by making information, constraints, intermediate states, and possible operations perceptually available. Third, the boundaries of a cognitive system should be determined by functional organization rather than automatically by the skin or skull. Fourth, system performance depends on the coordination and transformation of information among heterogeneous components. Finally, cognitive artifacts can improve or impair performance, which makes design, interaction, and organizational context central to the analysis (Hollan et al., 2000; Hutchins, 1995; Norman, 1993; Zhang, 1997; Zhang & Norman, 1994; Zhang & Patel, 2006).

A physician does not diagnose through unaided memory. Clinical cognition is distributed across the patient’s account, examination findings, laboratory results, imaging, medical records, guidelines, colleagues, software, and professional judgment. A scientist reasons through instruments, datasets, models, code, laboratories, published research, collaborators, and peer review. A pilot operates through displays, checklists, crew coordination, aircraft systems, procedures, and air traffic control. In each case, the individual remains essential, but the individual is not the complete cognitive system.

Zhang and Fenton (2024) offered a compact representation of this relationship by depicting distributed cognition as a coupled human-technology system. The human biological brain changes relatively slowly over historical time, while external data and technology have expanded rapidly (see also Stead, Searle, Fessler, Smith, & Shortliffe, 2011). The relevant unit of analysis is the combined system, not either component in isolation.

Figure 1

Distributed cognition as a human-technology system. A conceptual diagram showing a coupled human-technology distributed cognition system. The human brain is relatively stable while external data and technology grow rapidly.

Note. Reproduced from “Preparing healthcare education for an AI-augmented future” (Figure 2), by J. Zhang and S. H. Fenton, 2024, npj Health Systems, 1. The figure depicts the human brain and rapidly expanding data and technology as one coupled distributed cognitive system.

The figure is a conceptual starting point rather than a complete causal model. It captures two foundational ideas: the asymmetry in the rates of change between biological cognition and cognitive technology, and the interdependence of the human and technological components. It is not a substitution model in which technology simply replaces the brain. It is a distributed system in which external cognitive artifacts augment, reorganize, and sometimes constrain human cognition.

Frontier models deepen this framework. The technological side of the system is no longer composed primarily of external memory, static representations, and predefined procedures. Generative and agentic models can now produce, transform, evaluate, and coordinate representations. They are becoming active computational participants within the distributed cognitive system.

This intellectual lineage is synthesized and extended into the AI era in The Cognitive Revolution: How AI Is Reorganizing Intelligence, Expertise, and Institutions (Zhang, 2026b). The book argues that the relevant unit of intelligence is increasingly the human-AI interaction system and the institution in which that system operates. Distributed cognition provides the scientific foundation. The cognitive revolution names the broader transformation from primarily supportive cognitive artifacts to active computational participants, from AI as a bounded tool to AI as cognitive infrastructure, and from the analysis of individual tasks to the deliberate design and governance of institutional cognition.

4. Frontier Models as Active Computational Participants

4.1 From Cognitive Support to Computational Participation

Cognitive artifacts and conventional software have long extended human capability. What frontier models change is the nature and range of computational participation.

Traditional artifacts primarily stored, displayed, organized, transmitted, or constrained information. Earlier software systems could calculate, search, classify, and execute predefined procedures. Humans nevertheless remained responsible for interpreting outputs, choosing subsequent operations, and coordinating most of the broader cognitive sequence.

Frontier models can now generate representations, translate among representational forms, retrieve and synthesize information, compare explanations, produce hypotheses, critique outputs, simulate alternatives, write code, call tools, and plan sequences of action. AI no longer merely holds, transmits, or retrieves representations. It increasingly transforms them, evaluates them, produces new ones, and participates in determining what happens next.

Frontier models are becoming active cognitive artifacts.

4.2 Four Forms of Computationalized Cognitive Work

Computationalization does not mean that cognition suddenly became distributed when AI arrived. It means that work previously accomplished through repeated interaction among humans, documents, representations, databases, software, and organizational procedures can now be partially performed within computational systems. The distributed system does not disappear. Its internal allocation of cognitive work changes.

Four forms of work are particularly important:

1. Representational work converts information into summaries, classifications, explanations, models, visualizations, plans, or code. Quotation 2. Analytical work compares alternatives, identifies patterns, traces relationships, and evaluates possible interpretations. Quotation 3. Generative work produces hypotheses, designs, drafts, scenarios, and candidate solutions. Quotation 4. Coordinative work sequences operations, calls tools, routes information, tracks state, and manages portions of a workflow.

Frontier models do not perform these activities independently of data, tools, prompts, interfaces, and human direction. Their capability emerges through these relationships. The model is consequential because it can now occupy more positions within the distributed cognitive system.

4.3 Research, Software, and Institutional Work

Consider a literature review. Researchers traditionally searched databases, selected papers, extracted claims, compared findings, organized themes, evaluated evidence, and drafted a synthesis. Frontier models can now perform meaningful portions of that sequence. They can propose search strategies, summarize papers, compare claims, identify apparent contradictions, generate thematic structures, and draft syntheses.

But the complete epistemic process still requires people to determine which question matters, which sources are credible, what evidence is missing, whether findings are comparable, whether the synthesis is justified, and who assumes responsibility for the conclusions. AI computationalizes part of the research process. It does not become the complete research system.

The same transformation is visible in software development. A frontier model can generate code, explain unfamiliar libraries, identify errors, propose tests, revise implementations, translate among programming languages, and interact with development tools. More of the representational and procedural work has moved into the computational medium. But the complete system still requires problem definition, system architecture, security judgment, integration, validation, deployment, maintenance, and responsibility for failure.

Institutional analysis follows the same pattern. AI can aggregate information, detect patterns, model assumptions, generate alternatives, prepare summaries, and draft recommendations. But it cannot independently establish which institutional purposes are legitimate, whose interests should take priority, what tradeoffs are acceptable, or who has authority to act.

The computation expands. The responsibility remains institutional.

4.4 Scientific and Clinical Capability Emerges from the Distributed System

Recent AI systems designed for scientific discovery illustrate why frontier models should be understood as components of distributed cognitive systems rather than as self-contained scientific agents.

Google’s Co-Scientist combines a frontier language model with specialized agents that generate, critique, rank, and refine scientific hypotheses while using scientific literature, external tools, and human feedback. Scientists define research goals, establish constraints, steer the inquiry, evaluate competing hypotheses, and conduct experimental validation (Gottweis et al., 2026). Robin, a multi-agent system capable of fully automating both hypothesis generation and data analysis for experimental biology, similarly integrates literature search, hypothesis generation, experimental planning, and data analysis in an iterative lab-in-the-loop workflow. Human researchers perform the physical experiments and return the resulting data to the system, which analyzes the evidence and generates revised hypotheses (Ghareeb et al., 2026).

The scientific capability of these systems does not reside inside a single frontier model. Nor does it arise merely because multiple agents are connected. It emerges from the coordinated interaction among models, specialized agents, scientific literature, search systems, software tools, laboratory instruments, experimental procedures, domain scientists, and institutional research practices.

The model is a critical component. It is not the whole scientific intelligence.

Healthcare makes the same point. A frontier model may assemble a patient summary, retrieve relevant evidence, generate a differential diagnosis, identify missing information, or draft a treatment plan. Yet meaningful clinical cognition remains distributed across the patient, clinicians, electronic health records, laboratory and imaging systems, clinical guidelines, institutional policies, professional standards, communication processes, authorization mechanisms, monitoring systems, and accountability structures.

In both scientific research and clinical care, frontier models move increasing amounts of representational, analytical, generative, and coordinative work into computational systems. That transformation is historically significant. It is not, however, the same as creating a complete intelligence inside a machine.

Frontier models are computationalizing parts of distributed intelligence, not absorbing the entire distributed cognitive system.

5. A Distributed Cognition Framework for Human-AI Systems

5.1 An Extension and Application, Not a Replacement Theory

The appropriate framework for frontier AI is distributed cognition itself, extended and operationalized for systems in which AI models have become active cognitive artifacts. This article does not propose a wholly separate theory of intelligence. It applies the established distributed cognition framework to a new technical condition.

A human-AI distributed cognitive system is a functionally organized ensemble of people, AI models, data, internal and external representations, tools, workflows, institutions, and environments through which cognition and action are accomplished. Its performance depends not only on the capabilities of its components, but also on how cognitive work is allocated, how information is represented and transformed, how components coordinate, how outputs are validated, how decisions are authorized, how the system learns, and who remains accountable.

The term architecture is useful here, but it should not be mistaken for a separate theoretical framework. It refers to the organization of the distributed cognitive system: the relationships among people, models, representations, tools, workflows, institutions, and governance mechanisms. Distributed cognition is the theory. Architecture describes how a particular distributed cognitive system is configured.

5.2 Architectural Dimensions of Human-AI Distributed Cognition

The Zhang-Fenton human-technology model provides the conceptual core. Frontier models expand the active capabilities of the technological component, while institutions organize the relationships through which those capabilities enter consequential work. The resulting system should be analyzed across the following dimensions.

Adding more models does not automatically create more intelligence. Adding more agents does not automatically create coordination. Adding more data does not automatically create knowledge. Adding more autonomy does not automatically create wisdom.

The function matters more than the inventory.

A poorly designed distributed cognitive system containing extremely powerful components can amplify error, obscure responsibility, overwhelm decision-makers, preserve false assumptions, and accelerate action in the wrong direction. The relevant unit of evaluation is therefore the quality of the complete system.

5.3 Changing the Unit of Analysis

AGI asks:

How much intelligence can be built inside a machine?

A distributed cognition framework asks:

How is cognitive work organized across humans, AI systems, representations, tools, workflows, institutions, and environments?

AGI evaluates the artificial agent as the principal unit. Distributed cognition evaluates the complete functional system. AGI tends to treat tools, data, people, and institutions as external supports for the machine. Distributed cognition treats them as constitutive components of the cognitive process.

Generality may be one useful property of a system, but it should not define the framework. A system may operate across many domains and still be incoherent, brittle, unjust, or profoundly unwise. Breadth does not guarantee judgment. Autonomy does not guarantee legitimacy. Scale does not guarantee knowledge. Speed does not guarantee direction. More capability does not determine which purposes deserve pursuit.

The defining question is not whether the system is sufficiently general. It is whether the distributed cognitive system produces reliable, adaptive, legitimate, and accountable intelligence.

5.4 Governance Is Constitutive of Distributed Cognition

Governance is often treated as a constraint imposed after an intelligent system has been developed. That is the wrong model.

Governance determines which objectives are legitimate, which data may be used, which outputs require validation, who may authorize reliance, when a system may act, when it must stop, who may contest a decision, and who remains accountable when the system fails. These are not external administrative details. They shape what the system can know, decide, and do.

NIST’s AI Risk Management Framework reflects this socio-technical perspective by treating AI outcomes as products of interactions among models, data, people, organizations, interfaces, and use contexts, rather than as properties of the model alone (National Institute of Standards and Technology, 2023).

In a consequential human-AI system, governance is part of the organization of cognition itself.

6. What Frontier Models Do Not Independently Supply

6.1 Performance Does Not Establish Experience

The capabilities of frontier models are real. So are the limits of what those capabilities establish.

Two symmetrical errors should be avoided. The first is human exceptionalism by assertion, the assumption that machines can never perform functions previously regarded as uniquely human. The second is operational reductionism, the assumption that when a machine produces behavior associated with a human attribute, it necessarily possesses that attribute in the same sense.

Frontier models can describe pain, fear, love, illness, aging, vulnerability, and mortality with extraordinary fluency. Fluency is not first-person experience. A model can produce an accurate description of pain without establishing that it is in pain.

Human cognition is shaped by embodiment, including sensation, movement, fatigue, dependency, development, illness, vulnerability, and mortality. Embodiment does not merely provide an additional input channel. It helps determine what matters and how situations acquire significance. A model may represent embodied experience without sharing it.

6.2 Optimization Does Not Establish Purpose

AI can optimize an objective, propose alternative goals, and predict possible consequences. But the ability to optimize does not confer the authority to determine what should be optimized.

Questions such as the following cannot be resolved by performance alone:

• Is this objective worth pursuing?

• Who has the authority to select it?

• Whose interests does it serve?

• What rights constrain it?

• Which sacrifices are acceptable?

• Who bears the consequences?

These are questions of value, legitimacy, and responsibility. They are not simply questions of prediction and optimization.

6.3 Capability Does Not Confer Authority

A model may outperform a person on a bounded decision task. That does not give it legitimate authority to deny medical treatment, terminate employment, sentence a defendant, remove a child from a family, establish public policy, or initiate military action.

Authority derives from legal, ethical, constitutional, professional, and social arrangements. No benchmark score can confer it. Capability can support an authorized decision process. It cannot create legitimacy by itself.

6.4 Causal Contribution Is Not Accountability

AI systems will increasingly contribute to consequential decisions. But an organization cannot escape responsibility by claiming that “the system decided.”

Models do not provide an acceptable destination for institutional accountability. Humans and institutions establish purposes, choose systems, define rules, authorize reliance, and determine how outputs enter action.

The more cognitive work is delegated, the more explicitly responsibility must be designed.

7. What Humans and Institutions Must Preserve

7.1 Foundational Knowledge and Judgment

When AI can produce answers instantly, it becomes tempting to treat knowledge as unnecessary. That is a mistake. A person who knows nothing cannot reliably distinguish a strong answer from a plausible fabrication.

Foundational knowledge provides the conceptual structure required to formulate consequential questions, interpret evidence, identify omissions, detect error, and recognize when an output does not fit the situation. Knowledge becomes less scarce. It does not become irrelevant. Its role shifts from being the final product to being the foundation for judgment.

Judgment integrates evidence with context, uncertainty, timing, values, consequences, and responsibility. It determines when a technically correct answer is inappropriate, when a recommendation should not be followed, and when the problem itself has been framed incorrectly.

As generation becomes abundant, judgment becomes more valuable.

7.2 Intellectual Independence and Resilience

A productive human-AI system should expand human capability without destroying the capacity to think independently of the system. Overreliance can create cognitive fragility. People may lose the ability to evaluate outputs, recover from failure, recognize when assumptions no longer fit reality, or operate when computational infrastructure is unavailable.

Preserving human capability is not nostalgia. It is a requirement for resilience.

A system that increases short-term productivity while eliminating the expertise required to validate, govern, and repair it is not becoming more intelligent. It is becoming more dependent.

7.3 Contestability and Institutional Memory

Consequential outputs must remain open to questioning, inspection, appeal, and correction. A system may be accurate on average and still cause serious harm in individual cases. People affected by decisions need meaningful ways to challenge the data, assumptions, representations, interpretations, and authority behind them.

Contestability is not an obstacle to intelligence. It is part of intelligent system design.

Institutions must also preserve more than final outputs. They must retain the provenance, reasoning, uncertainty, context, and authority behind consequential decisions. An institution that stores conclusions but loses their rationale does not possess durable intelligence. It possesses an archive of unexplained decisions.

7.4 Purpose, Care, and Responsibility

AI can expand the space of possible actions. Humans must determine which possibilities deserve pursuit.

AI can generate and test representations. Humans must interpret significance and consequences.

AI can support decisions. Humans and institutions must authorize and answer for them.

AI can simulate values. Humans must establish legitimate commitments and obligations.

Purpose, care, legitimacy, and accountability are not residual tasks left after automation. They are constitutive elements of any distributed cognitive system worthy of trust.

8. Institutional Implications

8.1 Organizations: Design the Distributed Cognitive System

AI strategy is not primarily a technology procurement decision. Comparable access to frontier models will become increasingly widespread. Organizations will often use similar models, platforms, and services.

Durable advantage will come from designing a superior distributed cognitive system. That design determines which problems are worth solving, how questions are framed, which data enter the system, how information is represented, which tasks are assigned to people or machines, how outputs are validated, how uncertainty is communicated, who has authority to act, how failure is detected, how the institution learns, and who remains responsible.

The organization, not the model, is the strategic unit.

Leadership in the AI era increasingly becomes the design of systems that can sense, represent, reason, coordinate, decide, act, and learn.

8.2 Education and Science: From Production to Validation

The purpose of education cannot remain the unaided production of answers. Students still need foundational knowledge, but expertise is shifting from knowledge possession alone toward problem framing, representational literacy, evidence evaluation, verification, metacognition, human-AI collaboration, judgment under uncertainty, and responsibility for conclusions.

The goal is not cognitive dependency. It is competent participation in distributed cognition.

Science faces a parallel transformation. A system capable of generating ten thousand hypotheses but unable to validate them may increase activity without increasing knowledge. As generation becomes abundant, verification, replication, interpretation, and epistemic governance become the bottlenecks.

The relevant unit of scientific evaluation is not the model alone. It is the complete epistemic system connecting investigators, models, instruments, databases, methods, laboratories, peer review, and research governance.

8.3 Healthcare and Government: Evaluate the Whole System

The central question in healthcare is not whether AI is more intelligent than the physician. It is whether the redesigned patient-clinician-AI-institution system produces better outcomes, communicates uncertainty, protects dignity, supports professional judgment, reduces inequity, and preserves accountability.

The relevant unit of evaluation is the complete clinical cognitive system.

The same principle applies to government. The proper unit of AI governance is not only the algorithm or foundation model. It is the socio-technical system in which the model operates. The same model can be relatively safe in one configuration and dangerous in another. Data, interfaces, incentives, oversight, authority, appeal mechanisms, institutional capacity, and liability all shape the intelligence and risk of the system.

The future of intelligence is not merely a technical outcome. It is an institutional choice.

9. Conclusion: The Real Frontier Is Distributed Cognition

9.1 The Question Has Changed

AGI asks:

Can a machine match or exceed human intelligence across a broad range of tasks?

That remains a legitimate engineering question when its terms are defined precisely. But it is not the question that best describes the transformation already underway.

Distributed cognition asks:

Where does cognition occur when thinking is accomplished through people, representations, tools, environments, and institutions?

The computationalization of distributed intelligence asks:

Which parts of that cognitive work can now be generated, transformed, evaluated, and coordinated within computational systems?

Applied to human-AI systems, distributed cognition asks:

How should cognitive work be organized across humans, frontier models, representations, tools, workflows, institutions, and governance structures?

That is the framework that best fits what frontier models are doing.

9.2 The Computational Boundary Is Expanding

Frontier models are not simply becoming isolated artificial minds. They are expanding the computational boundary of distributed cognition. They are transforming representations that were once primarily external and relatively static into active computational processes capable of generation, synthesis, translation, evaluation, planning, and action.

But the expansion of computation does not eliminate the larger human and institutional system. Purpose, meaning, wisdom, legitimacy, care, judgment, and accountability remain essential precisely because machine capability is increasing.

The frontier is not a machine that contains everything intelligence requires. The frontier is the distributed cognitive system through which different forms of capability are combined, validated, governed, and directed.

9.3 The Target We Should Pursue

The decisive question is no longer:

How intelligent is the model?

It is:

What kind of intelligence emerges from the complete distributed cognitive system, whose purposes does it serve, what human capacities does it preserve, and who remains responsible for what it does?

AGI directs attention toward a hypothetical threshold inside the machine. Distributed cognition directs attention toward the human-AI systems already being built around us.

Those systems include models, but they also include people, data, representations, tools, workflows, institutions, authority, values, and responsibility.

The next frontier is not general intelligence inside a machine. It is the redesign of distributed cognition across humans, machines, and institutions.

The future will not be determined by intelligence alone. It will be determined by how distributed cognition is organized, governed, and directed.

References

Ghareeb, A. E., Chang, B., Mitchener, L., Yiu, A., Szostkiewicz, C. J., Shved, D., Gyimesi, G. J., Laurent, J. M., Wright, S. M., Razzak, M. T., White, A. D., Finnemann, S. C., Hinks, M. M., & Rodriques, S. G. (2026). A multi-agent system for automating scientific discovery. Nature, 655, 497–505. https://doi.org/10.1038/s41586-026-10652-y

Gottweis, J., Weng, W.-H., Daryin, A., Tu, T., Sirkovic, P., Myaskovsky, A., Glowaty, G., Weissenberger, F., Orlandi, A., Popovici, D., Palepu, A., Rong, K., Tanno, R., Saab, K., Zhang, F., Blum, J., Carroll, A., Kulkarni, K., Tomašev, N., … Natarajan, V. (2026). Accelerating scientific discovery with Co-Scientist. Nature, 655, 487–496. https://doi.org/10.1038/s41586-026-10644-y

Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human-computer interaction research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196. https://doi.org/10.1145/353485.353487

Hutchins, E. (1995). Cognition in the wild. MIT Press.

Morris, M. R., Sohl-Dickstein, J., Fiedel, N., Warkentin, T., Dafoe, A., Faust, A., Farabet, C., & Legg, S. (2024). Position: Levels of AGI for operationalizing progress on the path to AGI. In Proceedings of the 41st International Conference on Machine Learning (Vol. 235, pp. 36308–36321). PMLR.

National Institute of Standards and Technology. (2023). Artificial intelligence risk management framework, AI RMF 1.0 (NIST AI 100-1). https://doi.org/10.6028/NIST.AI.100-1

Norman, D. A. (1993). Things that make us smart: Defending human attributes in the age of the machine. Addison-Wesley.

Spearman, C. (1904). “General intelligence,” objectively determined and measured. The American Journal of Psychology, 15(2), 201–292. https://doi.org/10.2307/1412107

Stead, W. W., Searle, J. R., Fessler, H. E., Smith, J. W. & Shortliffe, E. H. Biomedical informatics: Changing what physicians need to know and how they learn. Academic Medicine 86, 429–434 (2011).

Zhang, J. (1997). The nature of external representations in problem solving. Cognitive Science, 21(2), 179–217. https://doi.org/10.1207/s15516709cog2102_3

Zhang, J. (2026a). AGI is a category error. https://jiajiezhang100.substack.com/p/agi-is-a-category-error

Zhang, J. (2026b). The cognitive revolution: How AI is reorganizing intelligence, expertise, and institutions. Open Intelligence Press.

Zhang, J., & Fenton, S. H. (2024). Preparing healthcare education for an AI-augmented future. npj Health Systems, 1, Article 4. https://doi.org/10.1038/s44401-024-00006-z

Zhang, J., & Norman, D. A. (1994). Representations in distributed cognitive tasks. Cognitive Science, 18(1), 87–122. https://doi.org/10.1207/s15516709cog1801_3

Zhang, J., & Patel, V. L. (2006). Distributed cognition, representation, and affordance. Pragmatics & Cognition, 14(2), 333–341. https://doi.org/10.1075/pc.14.2.12zha

Continue Reading

This framework is based on my book, The Cognitive Revolution: How AI Is Reorganizing Intelligence, Expertise, and Institutions.

Book: https://www.amazon.com/dp/B0GWMHQSNG

Website: https://www.dcognition.ai

Jiajie Zhang, PhD
Author of The Cognitive Revolution: How AI Is Reorganizing Intelligence, Expertise, and Institutions
Dean and Glassell Family Foundation Distinguished Chair
D. Bradley McWilliams School of Biomedical Informatics
UTHealth Houston

How This Essay Was Created
I originate the ideas, arguments, conceptual framework, conclusions, and the first draft of every essay. I then use AI as a cognitive partner to challenge my thinking, improve organization, strengthen clarity, and refine the writing. The intellectual contributions are my own. This collaborative process reflects the central thesis of The Cognitive Revolution: intelligence increasingly emerges through distributed cognitive systems.

© 2026 Jiajie Zhang. All rights reserved.

Brief quotations with attribution and a link to the original publication are welcome. No part of this article may be reproduced, republished, or distributed in whole or in substantial part without prior written permission from the author.