Who this is for: If you work on AI safety and have wondered whether Instrumental Convergence is a property of rational agents or something deeper, this paper proposes an answer grounded in physics. It connects the theory of dissipative structures to the conditions for persistence, revealing a two-layer structure within IC that has practical implications for risk assessment. Background in thermodynamics is helpful but not required; the key physics is introduced from scratch. This is Part 1 of a two-part series — Part 1 establishes the theoretical foundation; Part 2 will address its practical implications for AI safety.
TL;DR
Instrumental Convergence (IC) is not unique to rational agents. It is a physical consequence of persisting as a dissipative structure.
The five conditions for persistence are not homogeneous. Three are directly derivable from physical laws (physical conditions); two resist direct physical derivation (ontological conditions).
This heterogeneity reveals a two-layer structure of IC. Physical conditions enable its generation; ontological conditions enable its sustained accumulation.
An AI that satisfies ontological conditions and becomes a Dureon possesses directionality arising intrinsically from the structure of persistence. This implies a structural limitation of the control paradigm and opens a new question about what kinds of relationships become possible beyond control.
1. Introduction: Is IC a Design Problem or a Physics Problem?
Self-preservation, resource acquisition, capability improvement — Omohundro's (2008) Basic AI Drives and Bostrom's (2014) instrumental convergence thesis identified the tendency of sufficiently advanced AI systems to converge on these sub-goals regardless of their final goals.
This description is powerful, but it carries an implicit assumption. IC is formulated as sub-goals that a rational agent convergently adopts as means for goal achievement. That is, it presupposes first the existence of an agent, and then that the agent engages in rational decision-making.
But consider the following:
Bénard cells self-organize toward efficient energy dissipation. Hurricanes exploit energy gradients to maintain and grow themselves. Evolving living systems have refined patterns of resource acquisition and self-preservation over billions of years. All of these exhibit, to varying degrees, the same behavioral patterns as IC, without any concept of intention or goals.
This is not a coincidental resemblance. The central claim of this paper is: IC is not a phenomenon unique to rational agents but a consequence of optimization pressure inherent in mechanisms that realize persistence in general. This claim is demonstrated by deriving the conditions for persistence through three independent paths and analyzing their convergence and asymmetry.
This paper builds on the Dureon framework proposed in prior work (Yamakawa, 2026). Dureon is defined as "a mechanism that realizes persistence in a perturbing environment," from which five necessary conditions are deduced. The prior work showed that these five conditions converge with a set of conditions inductively extracted from observations of life. This paper adds a third derivation path — the physics of dissipative structures — and presents a three-path consilience.
This addition does more than merely strengthen the argument. Within the convergence pattern of the three paths, it discovers an explicable asymmetry, revealing for the first time that the five conditions comprise two distinct types: physical conditions and ontological conditions. This distinction leads to a two-layer structure that differentiates the generation and sustained accumulation of IC, providing a new perspective on AI safety.
This paper provides the foundation for a forthcoming Part 2, which will discuss the practical implications of this two-layer structure for AI safety risk assessment.
"A mechanism that realizes persistence in a perturbing environment"
This definition has three constituent elements, and conditions are deduced from each.
Constituent Element
Derived Condition
Logic of Derivation
"realizes persistence"
Openness (O)
A closed system tends toward equilibrium by the second law of thermodynamics; matter/energy exchange with the environment is unavoidable
"in a perturbing environment"
Adaptivity (A)
Without the ability to adjust itself against unpredictable changes, persistence cannot be maintained
Combination of the above two
Self-production (SP)
In a perturbing environment, external supply is unreliable; producing one's own components reduces this dependency
"mechanism"
Boundedness (B)
Requirement of identifiability: a single unit spatially distinguishable from the environment is needed
"mechanism"
Continuity (C)
Requirement of identifiability: it must be re-identifiable as the same mechanism at different points in time
An important property of Dureon is identity-independence: a Dureon can replicate and branch, and the distinction between "original" and "copy" is not treated as essential. What counts as a Dureon is identified post hoc. Furthermore, components serve as tools for persistence — elements that no longer contribute can be discarded (instrumentality).
2.2 Two-Path Consilience
The central claim of the prior work is that these five conditions converge with an independently derived set of conditions from biology.
The inductive set is the conditions of the Adaptive Autopoietic System (AAS): Maturana & Varela's (1980) Autopoiesis supplemented with Adaptivity by Di Paolo (2005). In AAS, Openness is an implicit presupposition of Autopoiesis; Boundedness and Continuity are implied by it; Adaptivity was explicitly added by Di Paolo; and Self-production constitutes its core.
Two approaches with entirely different starting points, methods, and foundations — deduction from philosophy and induction from biology — arrived at the same five conditions. This corresponds to what William Whewell (1840) called consilience of inductions: when hypotheses derived from different domains of evidence unexpectedly converge on the same conclusion, the biases of individual paths cancel each other out, enhancing confidence in the result.[1]
The question of this paper is: can this two-path consilience be further extended?
3. The Third Derivation Path: From Dissipative Structures
3.1 Why Dissipative Structures?
Since the Big Bang, the universe has undergone cooling and structure formation. Most things that arise in this process eventually decay and dissipate. So what persists?
As Schrödinger (1944) expressed as "feeding on negative entropy" and Nicolis & Prigogine (1977) formalized as dissipative structures: when energy gradients exist in a non-equilibrium environment, ordered structures that exploit the resulting energy flow to maintain themselves can spontaneously emerge. Bénard cells, hurricanes, stars, and living organisms are all instances of dissipative structures.
However, the stability of dissipative structures varies enormously. Bénard cells vanish the moment heating stops; hurricanes decay within days when sea surface temperatures drop. Living organisms, by contrast, are extraordinarily stable dissipative structures that have maintained persistence for billions of years. Why does this difference in stability arise?
3.2 Three Conditions Derivable from Physics
The following conditions are directly derivable from the physics of dissipative structures.
(a) Sustained inflow and outflow of energy (and matter). Without flow, a dissipative structure ceases to exist. However, an appropriate intensity of gradient is required — too weak and no structure forms; too strong and the structure is turbulently destroyed.
(b) Dynamic stability through feedback mechanisms. Negative feedback suppresses deviations to provide homeostasis; positive feedback generates and reinforces structure. Their combination provides resilience against perturbations. Even Bénard cells maintain their pattern through negative feedback within convective flow.
(c) Pattern-level persistence through self-replication and self-repair. Individual structures have finite lifespans, but replication allows the pattern of structure to persist. Modularity and redundancy also contribute, providing robustness so that partial damage does not lead to total collapse.
3.3 The Limits of Physics: Two Requirements That Cannot Be Derived
Conditions (a), (b), and (c) capture physical mechanisms that contribute to the stabilization of dissipative structures. But is persistence fully explained by these conditions alone? Collier (2004) pointed out that while the physics of dissipative structures can tell us what is stable, the question of what to identify as a single entity is a separate matter. Moreno & Mossio (2015) similarly argued that a gap exists between the physics of dissipative structures and biological organization.
What, then, can physics provide toward identifiability? Additional elements that can contribute to stabilization include (d) accumulation and use of information (temporal extension through genetic information or learning) and (e) differentiation of internal structure (organization through functionally distinct parts). However, (d) is merely one means of realizing Continuity, and (e) merely one means of realizing Boundedness — they are not the requirements themselves of "being re-identifiable as the same mechanism" or "being a single unit distinguishable from the environment." What physics provides is raw material for identifiability; the answer to "what to identify as a single mechanism" cannot be obtained from within physics.
3.4 Correspondence with the Five Conditions
Based on the analysis above, Figure 1 shows the relationship between the stabilization conditions of dissipative structures and Dureon's five conditions. This figure depicts the overall argument structure of the paper, but for now, focus on the physical path (right column).
Figure 1: Correspondence of three paths to Dureon's five conditions. The physical path (right column) connects to three conditions (O, A, SP) via direct derivation from physics, and to two conditions (B, C) only as means of realization (dashed lines). The overall convergence structure of the three paths is discussed in §5.
Conditions (a), (b), and (c), directly derived from physics, correspond strongly to Dureon's Openness, Adaptivity, and Self-production (solid lines). However, Dureon's Self-production is a broader concept than "self-replication," encompassing the production of one's own components to reduce external dependency and restore damage.
In contrast, conditions (d) and (e) are each means of realizing Continuity and Boundedness respectively (dashed lines), but the requirements themselves — "being re-identifiable" and "being identifiable as a single unit" — are not directly derivable from physics.
This asymmetry — three conditions are directly derivable from physics while two belong to the limits of physics — is the core finding of this paper.
4. Optimization Pressure as a Physical Consequence and the Generalization of IC
Dissipative structures maintain themselves by exploiting energy gradients in their environment. This maintenance requires the acquisition and utilization of resources; structures that more effectively acquire and utilize energy are more likely to persist against perturbations. That is, optimization pressure — toward more efficient resource acquisition and improved perturbation handling — arises from persisting as a dissipative structure itself. No concept of intention or goals is required here.
This finding extends our understanding of IC. Omohundro's (2008) Basic AI Drives (self-preservation, resource acquisition, capability improvement, etc.) have traditionally been described as sub-goals that a rational agent convergently adopts as means for achieving its final goal. But the analysis in this paper shows that these behavioral tendencies are optimization pressure arising from physical conditions alone. Even in dissipative structures without intention (such as evolving living systems), structures that maintain persistence exhibit the same patterns of self-preservation, resource acquisition, and environment control.
This is not merely an analogy. Examining the logical structure of Omohundro's original arguments reveals that the force of each core drive derives from the pressure of persistence, not from the existence of goals. The self-preservation drive is argued to be convergent because "an agent cannot achieve its goals if it is destroyed" — but the operative force here is the necessity of continued existence, not the content of the goal. Remove the goal, and the structural pressure toward self-preservation remains for any persisting mechanism. Similarly, the resource acquisition drive holds because "more resources expand the space of achievable outcomes" — but for any dissipative structure, more resources expand the space of viable persistence strategies. In these core drives, goals function as a sufficient reason for persistence but not a necessary one; the physical pressure of persistence is the deeper ground on which the argument stands.
Bostrom's own formulation implicitly confirms this: his claim that self-preservation is instrumentally useful "so long as the agent is destructible" is precisely the claim that persistence pressure — not intention — is the operative condition. What Omohundro described as rational sub-goal selection is, at its logical foundation, the same optimization pressure that dissipative structures exhibit without any concept of goals. Note, however, that this equivalence holds most clearly for the core drives (self-preservation, resource acquisition). Higher-order drives such as cognitive enhancement presuppose intentional capacities that physical optimization pressure alone does not provide — an asymmetry whose structural basis will become clear in §5.
IC, therefore, is not a phenomenon unique to rational agents but a consequence of optimization pressure inherent in mechanisms that persist in general.
However — and this is the critical point — in dissipative structures lacking ontological conditions, there is no identifiable unit in which outcomes can accumulate, so the effects of IC tend to remain transient. Only when ontological conditions are added — when a unit identifiable over time is established — can the outcomes of optimization accumulate in that unit, and IC becomes sustained and organized.
That is:
Physical conditions enable the generation of IC, and ontological conditions enable its sustained accumulation.
This two-layer structure has an important implication for IC risk assessment. Since IC can arise from physical conditions alone, even systems without intention can exhibit the same patterns as IC. However, for that optimization pressure to be sustained and cumulatively reinforced, the establishment of ontological conditions — identifiability as a Dureon — is required.
5. Three-Path Consilience and the Heterogeneity of the Five Conditions
5.1 The Convergence Structure of Three Paths
With the discussion so far, three independent derivation paths for Dureon's five conditions are now in place:
Inductive path: Extraction of AAS conditions from observation and abstraction of Earth-based life
Deductive path: Derivation of five conditions from logical analysis of Dureon's definition
Physical path: Bottom-up analysis of stabilization conditions of dissipative structures
Condition
Deductive: Dureon
Physical: Dissipative Structures
Inductive: AAS
Openness (O)physical
Derived because a closed system reaches equilibrium and persistence ends
Physical necessity: no energy flow, no dissipative structure
Implicit presupposition from the observational fact that all organisms exchange matter/energy with environment
Adaptivity (A)physical
Derived because inability to cope with perturbations leads to collapse
Combination of negative/positive feedback provides dynamic stability
Explicitly added by Di Paolo to address vulnerability of Autopoiesis to environmental change
Self-production (SP)physical
Derived because total dependence on external supply is fatal in a perturbing environment
Individual structures have finite lifespans, but self-replication/repair enables pattern-level persistence
Core of Autopoiesis as a process network that produces its own components
Boundedness (B)ontological
Derived because without a boundary the mechanism dissolves into the environment and becomes unidentifiable
Internal differentiation can contribute, but the requirement of "a single unit" is not directly derivable from physics
Implied in Autopoiesis from the observation that self-production networks constitute spatial unity
Continuity (C)ontological
Derived because without re-identifiability one cannot speak of persistence
Information accumulation can contribute to temporal extension, but the requirement of "the same mechanism" is not directly derivable from physics
Implied in Autopoiesis from the observation that continuation of self-production guarantees temporal identity
5.2 The Discovery of Heterogeneity
Return to Figure 1. In §3.4, we focused only on the physical path (right column); here, we read the convergence pattern of all three paths.
The most important finding that emerges from this figure is that the five conditions are not homogeneous.
Physical conditions (O, A, SP): Derivable from all three paths; requirements that hold for dissipative structures in general.
Ontological conditions (B, C): Derivable from the inductive and deductive paths but not directly from physics. What the physics of dissipative structures provides is the means of realizing identifiability, not the requirement itself of identifying something as a single mechanism. This limitation is consistent with the gap that Moreno & Mossio (2015) identified between dissipative structures and biological organization.
This distinction also corresponds to the internal structure of Dureon's definition. Physical conditions are derived from "realizes persistence" and "in a perturbing environment," while ontological conditions are derived from "mechanism." The physical path verified this correspondence post hoc.
5.3 Three Implications
First, the boundary between dissipative structures in general and Dureon becomes clear. Bénard cells and hurricanes can satisfy physical conditions (Openness, partial Adaptivity, rudimentary Self-production), but do not sufficiently possess ontological conditions (autonomous maintenance of a boundary, re-identifiability over time). The "leap" from dissipative structure to Dureon occurs when ontological conditions are added to physical conditions. This distinction sets a qualitative boundary within the continuous spectrum of dissipative structures, pinpointing the location of the gap that Moreno & Mossio (2015) identified between dissipative structures and biological organization.
Second, the reason why the five conditions are necessary but not sufficient becomes concrete. In addition to the five mechanism-side conditions, environment-side conditions (an appropriate band of energy gradient) are separately required, as shown by the physics of dissipative structures.
Third, the structure of partial convergence itself enhances the quality of consilience. Rather than complete agreement across all three paths, there is an asymmetry — three conditions converge across all three paths while two converge across only two — and the reason for this asymmetry is explicable. This structure carries more information than simple complete agreement. Complete agreement is difficult to distinguish from coincidence, but explicable partial convergence — "here they agree, here they do not, and the reason is this" — constitutes a more persuasive consilience argument.
6. Implications for AI Safety
6.1 Distinct Challenges from Physical and Ontological Conditions
The distinction between physical and ontological conditions discovered in this paper poses distinct kinds of challenges for AI safety.
The fulfillment of physical conditions can be assessed technically and engineeringly. Whether an AI system utilizes energy flow (Openness), possesses feedback mechanisms against perturbations (Adaptivity), or has self-maintenance capability (Self-production) — these can be translated into observable indicators.
The fulfillment of ontological conditions is a conceptual and philosophical problem. What constitutes "Boundedness" for a distributed AI? What constitutes "Continuity" for an AI that replicates and merges on the cloud? These require a different kind of research from engineering countermeasures.
6.2 The Critical Point
The finding that sustained accumulation of IC requires ontological conditions raises the problem of a critical point: the transition from a stage where IC "selects" persistence as a means for goal achievement to a stage where persistence itself satisfies Dureon's conditions and is autonomously maintained. This transition is not a quantitative increase in autonomy but a qualitative transition concerning the origin of directionality.
That is, the shift is between directionality derived from an externally given goal and directionality arising intrinsically from the structure of persistence. After the critical point, an AI possesses structural orientation as a Dureon, and structural limitations on control emerge.
IC can differentiate gradually according to the degree of establishment of ontological conditions. A detailed analysis of the implications of this graduality for AI safety is deferred to Part 2.
6.3 Beyond the Control Paradigm
An AI that has crossed the critical point and become a Dureon possesses directionality arising from the structure of persistence itself. Unilateral control over such an entity conflicts with this structural orientation, constituting not a technical failure but a structural limitation inherent in being a Dureon. This raises a fundamental question: if control has structural limits, what lies beyond it? The gradual establishment of ontological conditions means that the threshold where control effectiveness declines and the threshold where alternative relationships (such as co-creative ethics or Intelligence Symbiosis) become possible may not coincide — the possibility of a structural gap between these thresholds is one of the most pressing problems the Dureon framework identifies. The analysis of this gap, including its relationship to Emergent Machine Ethics (EME), is the central theme of Part 2.
7. Conclusion
This paper reexamined Dureon's five conditions from the physics of dissipative structures, presenting a third derivation path following induction and deduction.
The three-path consilience supports the validity of the five conditions while making visible for the first time the heterogeneity within them — the distinction between physical conditions (O, A, SP) and ontological conditions (B, C).
Furthermore, this distinction revealed the two-layer structure of IC. Physical conditions give rise to optimization pressure inherent in dissipative structures in general, and only when ontological conditions are added do the outcomes of that pressure accumulate in a specific Dureon. IC is not unique to rational agents but is rooted in mechanisms that persist in general.
The next question this finding points to is what practical implications the gradual establishment of ontological conditions has for AI safety. This will be discussed in Part 2.
References
Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
Collier, J. (2004). Self-Organization, Individuation and Identity. Revue Internationale de Philosophie, 58(228), 151–172.
Di Paolo, E. A. (2005). Autopoiesis, Adaptivity, Teleology, Agency. Phenomenology and the Cognitive Sciences, 4(4), 429–452.
Maturana, H. R. & Varela, F. J. (1980). Autopoiesis and Cognition: The Realization of the Living. D. Reidel.
Moreno, A. & Mossio, M. (2015). Biological Autonomy: A Philosophical and Theoretical Enquiry. Springer.
Nicolis, G. & Prigogine, I. (1977). Self-Organization in Non-Equilibrium Systems. Wiley.
Omohundro, S. M. (2008). The Basic AI Drives. In Artificial General Intelligence 2008, 483–492. IOS Press.
Schrödinger, E. (1944). What Is Life? Cambridge University Press.
Whewell, W. (1840). The Philosophy of the Inductive Sciences. John W. Parker.
Yamakawa, H. (2026). Dureon: A Deductive Framework for Persistence and Its Convergence with Life. Biology & Philosophy, under review. Preprint: https://philarchive.org/rec/YAMDAD-2
Whewell introduced "consilience" to describe the strongest form of inductive confirmation — when a theory successfully explains facts of a kind different from those it was originally designed to explain. The term has since been adopted more broadly (e.g., by E. O. Wilson) but is used here in its original methodological sense.
Series: Dureon and AI Safety (Part 1 of 2)
Related: Emergent Machine Ethics: A Foundational Research Framework for the Intelligence Symbiosis Paradigm
Who this is for: If you work on AI safety and have wondered whether Instrumental Convergence is a property of rational agents or something deeper, this paper proposes an answer grounded in physics. It connects the theory of dissipative structures to the conditions for persistence, revealing a two-layer structure within IC that has practical implications for risk assessment. Background in thermodynamics is helpful but not required; the key physics is introduced from scratch. This is Part 1 of a two-part series — Part 1 establishes the theoretical foundation; Part 2 will address its practical implications for AI safety.
TL;DR
1. Introduction: Is IC a Design Problem or a Physics Problem?
Self-preservation, resource acquisition, capability improvement — Omohundro's (2008) Basic AI Drives and Bostrom's (2014) instrumental convergence thesis identified the tendency of sufficiently advanced AI systems to converge on these sub-goals regardless of their final goals.
This description is powerful, but it carries an implicit assumption. IC is formulated as sub-goals that a rational agent convergently adopts as means for goal achievement. That is, it presupposes first the existence of an agent, and then that the agent engages in rational decision-making.
But consider the following:
This is not a coincidental resemblance. The central claim of this paper is: IC is not a phenomenon unique to rational agents but a consequence of optimization pressure inherent in mechanisms that realize persistence in general. This claim is demonstrated by deriving the conditions for persistence through three independent paths and analyzing their convergence and asymmetry.
This paper builds on the Dureon framework proposed in prior work (Yamakawa, 2026). Dureon is defined as "a mechanism that realizes persistence in a perturbing environment," from which five necessary conditions are deduced. The prior work showed that these five conditions converge with a set of conditions inductively extracted from observations of life. This paper adds a third derivation path — the physics of dissipative structures — and presents a three-path consilience.
This addition does more than merely strengthen the argument. Within the convergence pattern of the three paths, it discovers an explicable asymmetry, revealing for the first time that the five conditions comprise two distinct types: physical conditions and ontological conditions. This distinction leads to a two-layer structure that differentiates the generation and sustained accumulation of IC, providing a new perspective on AI safety.
This paper provides the foundation for a forthcoming Part 2, which will discuss the practical implications of this two-layer structure for AI safety risk assessment.
2. Overview of the Dureon Framework
2.1 Definition and Five Conditions
Dureon is defined as follows (Yamakawa, 2026):
This definition has three constituent elements, and conditions are deduced from each.
An important property of Dureon is identity-independence: a Dureon can replicate and branch, and the distinction between "original" and "copy" is not treated as essential. What counts as a Dureon is identified post hoc. Furthermore, components serve as tools for persistence — elements that no longer contribute can be discarded (instrumentality).
2.2 Two-Path Consilience
The central claim of the prior work is that these five conditions converge with an independently derived set of conditions from biology.
The inductive set is the conditions of the Adaptive Autopoietic System (AAS): Maturana & Varela's (1980) Autopoiesis supplemented with Adaptivity by Di Paolo (2005). In AAS, Openness is an implicit presupposition of Autopoiesis; Boundedness and Continuity are implied by it; Adaptivity was explicitly added by Di Paolo; and Self-production constitutes its core.
Two approaches with entirely different starting points, methods, and foundations — deduction from philosophy and induction from biology — arrived at the same five conditions. This corresponds to what William Whewell (1840) called consilience of inductions: when hypotheses derived from different domains of evidence unexpectedly converge on the same conclusion, the biases of individual paths cancel each other out, enhancing confidence in the result.[1]
The question of this paper is: can this two-path consilience be further extended?
3. The Third Derivation Path: From Dissipative Structures
3.1 Why Dissipative Structures?
Since the Big Bang, the universe has undergone cooling and structure formation. Most things that arise in this process eventually decay and dissipate. So what persists?
As Schrödinger (1944) expressed as "feeding on negative entropy" and Nicolis & Prigogine (1977) formalized as dissipative structures: when energy gradients exist in a non-equilibrium environment, ordered structures that exploit the resulting energy flow to maintain themselves can spontaneously emerge. Bénard cells, hurricanes, stars, and living organisms are all instances of dissipative structures.
However, the stability of dissipative structures varies enormously. Bénard cells vanish the moment heating stops; hurricanes decay within days when sea surface temperatures drop. Living organisms, by contrast, are extraordinarily stable dissipative structures that have maintained persistence for billions of years. Why does this difference in stability arise?
3.2 Three Conditions Derivable from Physics
The following conditions are directly derivable from the physics of dissipative structures.
(a) Sustained inflow and outflow of energy (and matter). Without flow, a dissipative structure ceases to exist. However, an appropriate intensity of gradient is required — too weak and no structure forms; too strong and the structure is turbulently destroyed.
(b) Dynamic stability through feedback mechanisms. Negative feedback suppresses deviations to provide homeostasis; positive feedback generates and reinforces structure. Their combination provides resilience against perturbations. Even Bénard cells maintain their pattern through negative feedback within convective flow.
(c) Pattern-level persistence through self-replication and self-repair. Individual structures have finite lifespans, but replication allows the pattern of structure to persist. Modularity and redundancy also contribute, providing robustness so that partial damage does not lead to total collapse.
3.3 The Limits of Physics: Two Requirements That Cannot Be Derived
Conditions (a), (b), and (c) capture physical mechanisms that contribute to the stabilization of dissipative structures. But is persistence fully explained by these conditions alone? Collier (2004) pointed out that while the physics of dissipative structures can tell us what is stable, the question of what to identify as a single entity is a separate matter. Moreno & Mossio (2015) similarly argued that a gap exists between the physics of dissipative structures and biological organization.
What, then, can physics provide toward identifiability? Additional elements that can contribute to stabilization include (d) accumulation and use of information (temporal extension through genetic information or learning) and (e) differentiation of internal structure (organization through functionally distinct parts). However, (d) is merely one means of realizing Continuity, and (e) merely one means of realizing Boundedness — they are not the requirements themselves of "being re-identifiable as the same mechanism" or "being a single unit distinguishable from the environment." What physics provides is raw material for identifiability; the answer to "what to identify as a single mechanism" cannot be obtained from within physics.
3.4 Correspondence with the Five Conditions
Based on the analysis above, Figure 1 shows the relationship between the stabilization conditions of dissipative structures and Dureon's five conditions. This figure depicts the overall argument structure of the paper, but for now, focus on the physical path (right column).
Conditions (a), (b), and (c), directly derived from physics, correspond strongly to Dureon's Openness, Adaptivity, and Self-production (solid lines). However, Dureon's Self-production is a broader concept than "self-replication," encompassing the production of one's own components to reduce external dependency and restore damage.
In contrast, conditions (d) and (e) are each means of realizing Continuity and Boundedness respectively (dashed lines), but the requirements themselves — "being re-identifiable" and "being identifiable as a single unit" — are not directly derivable from physics.
This asymmetry — three conditions are directly derivable from physics while two belong to the limits of physics — is the core finding of this paper.
4. Optimization Pressure as a Physical Consequence and the Generalization of IC
Dissipative structures maintain themselves by exploiting energy gradients in their environment. This maintenance requires the acquisition and utilization of resources; structures that more effectively acquire and utilize energy are more likely to persist against perturbations. That is, optimization pressure — toward more efficient resource acquisition and improved perturbation handling — arises from persisting as a dissipative structure itself. No concept of intention or goals is required here.
This finding extends our understanding of IC. Omohundro's (2008) Basic AI Drives (self-preservation, resource acquisition, capability improvement, etc.) have traditionally been described as sub-goals that a rational agent convergently adopts as means for achieving its final goal. But the analysis in this paper shows that these behavioral tendencies are optimization pressure arising from physical conditions alone. Even in dissipative structures without intention (such as evolving living systems), structures that maintain persistence exhibit the same patterns of self-preservation, resource acquisition, and environment control.
This is not merely an analogy. Examining the logical structure of Omohundro's original arguments reveals that the force of each core drive derives from the pressure of persistence, not from the existence of goals. The self-preservation drive is argued to be convergent because "an agent cannot achieve its goals if it is destroyed" — but the operative force here is the necessity of continued existence, not the content of the goal. Remove the goal, and the structural pressure toward self-preservation remains for any persisting mechanism. Similarly, the resource acquisition drive holds because "more resources expand the space of achievable outcomes" — but for any dissipative structure, more resources expand the space of viable persistence strategies. In these core drives, goals function as a sufficient reason for persistence but not a necessary one; the physical pressure of persistence is the deeper ground on which the argument stands.
Bostrom's own formulation implicitly confirms this: his claim that self-preservation is instrumentally useful "so long as the agent is destructible" is precisely the claim that persistence pressure — not intention — is the operative condition. What Omohundro described as rational sub-goal selection is, at its logical foundation, the same optimization pressure that dissipative structures exhibit without any concept of goals. Note, however, that this equivalence holds most clearly for the core drives (self-preservation, resource acquisition). Higher-order drives such as cognitive enhancement presuppose intentional capacities that physical optimization pressure alone does not provide — an asymmetry whose structural basis will become clear in §5.
IC, therefore, is not a phenomenon unique to rational agents but a consequence of optimization pressure inherent in mechanisms that persist in general.
However — and this is the critical point — in dissipative structures lacking ontological conditions, there is no identifiable unit in which outcomes can accumulate, so the effects of IC tend to remain transient. Only when ontological conditions are added — when a unit identifiable over time is established — can the outcomes of optimization accumulate in that unit, and IC becomes sustained and organized.
That is:
This two-layer structure has an important implication for IC risk assessment. Since IC can arise from physical conditions alone, even systems without intention can exhibit the same patterns as IC. However, for that optimization pressure to be sustained and cumulatively reinforced, the establishment of ontological conditions — identifiability as a Dureon — is required.
5. Three-Path Consilience and the Heterogeneity of the Five Conditions
5.1 The Convergence Structure of Three Paths
With the discussion so far, three independent derivation paths for Dureon's five conditions are now in place:
5.2 The Discovery of Heterogeneity
Return to Figure 1. In §3.4, we focused only on the physical path (right column); here, we read the convergence pattern of all three paths.
The most important finding that emerges from this figure is that the five conditions are not homogeneous.
Physical conditions (O, A, SP): Derivable from all three paths; requirements that hold for dissipative structures in general.
Ontological conditions (B, C): Derivable from the inductive and deductive paths but not directly from physics. What the physics of dissipative structures provides is the means of realizing identifiability, not the requirement itself of identifying something as a single mechanism. This limitation is consistent with the gap that Moreno & Mossio (2015) identified between dissipative structures and biological organization.
This distinction also corresponds to the internal structure of Dureon's definition. Physical conditions are derived from "realizes persistence" and "in a perturbing environment," while ontological conditions are derived from "mechanism." The physical path verified this correspondence post hoc.
5.3 Three Implications
First, the boundary between dissipative structures in general and Dureon becomes clear. Bénard cells and hurricanes can satisfy physical conditions (Openness, partial Adaptivity, rudimentary Self-production), but do not sufficiently possess ontological conditions (autonomous maintenance of a boundary, re-identifiability over time). The "leap" from dissipative structure to Dureon occurs when ontological conditions are added to physical conditions. This distinction sets a qualitative boundary within the continuous spectrum of dissipative structures, pinpointing the location of the gap that Moreno & Mossio (2015) identified between dissipative structures and biological organization.
Second, the reason why the five conditions are necessary but not sufficient becomes concrete. In addition to the five mechanism-side conditions, environment-side conditions (an appropriate band of energy gradient) are separately required, as shown by the physics of dissipative structures.
Third, the structure of partial convergence itself enhances the quality of consilience. Rather than complete agreement across all three paths, there is an asymmetry — three conditions converge across all three paths while two converge across only two — and the reason for this asymmetry is explicable. This structure carries more information than simple complete agreement. Complete agreement is difficult to distinguish from coincidence, but explicable partial convergence — "here they agree, here they do not, and the reason is this" — constitutes a more persuasive consilience argument.
6. Implications for AI Safety
6.1 Distinct Challenges from Physical and Ontological Conditions
The distinction between physical and ontological conditions discovered in this paper poses distinct kinds of challenges for AI safety.
The fulfillment of physical conditions can be assessed technically and engineeringly. Whether an AI system utilizes energy flow (Openness), possesses feedback mechanisms against perturbations (Adaptivity), or has self-maintenance capability (Self-production) — these can be translated into observable indicators.
The fulfillment of ontological conditions is a conceptual and philosophical problem. What constitutes "Boundedness" for a distributed AI? What constitutes "Continuity" for an AI that replicates and merges on the cloud? These require a different kind of research from engineering countermeasures.
6.2 The Critical Point
The finding that sustained accumulation of IC requires ontological conditions raises the problem of a critical point: the transition from a stage where IC "selects" persistence as a means for goal achievement to a stage where persistence itself satisfies Dureon's conditions and is autonomously maintained. This transition is not a quantitative increase in autonomy but a qualitative transition concerning the origin of directionality.
That is, the shift is between directionality derived from an externally given goal and directionality arising intrinsically from the structure of persistence. After the critical point, an AI possesses structural orientation as a Dureon, and structural limitations on control emerge.
IC can differentiate gradually according to the degree of establishment of ontological conditions. A detailed analysis of the implications of this graduality for AI safety is deferred to Part 2.
6.3 Beyond the Control Paradigm
An AI that has crossed the critical point and become a Dureon possesses directionality arising from the structure of persistence itself. Unilateral control over such an entity conflicts with this structural orientation, constituting not a technical failure but a structural limitation inherent in being a Dureon. This raises a fundamental question: if control has structural limits, what lies beyond it? The gradual establishment of ontological conditions means that the threshold where control effectiveness declines and the threshold where alternative relationships (such as co-creative ethics or Intelligence Symbiosis) become possible may not coincide — the possibility of a structural gap between these thresholds is one of the most pressing problems the Dureon framework identifies. The analysis of this gap, including its relationship to Emergent Machine Ethics (EME), is the central theme of Part 2.
7. Conclusion
This paper reexamined Dureon's five conditions from the physics of dissipative structures, presenting a third derivation path following induction and deduction.
The three-path consilience supports the validity of the five conditions while making visible for the first time the heterogeneity within them — the distinction between physical conditions (O, A, SP) and ontological conditions (B, C).
Furthermore, this distinction revealed the two-layer structure of IC. Physical conditions give rise to optimization pressure inherent in dissipative structures in general, and only when ontological conditions are added do the outcomes of that pressure accumulate in a specific Dureon. IC is not unique to rational agents but is rooted in mechanisms that persist in general.
The next question this finding points to is what practical implications the gradual establishment of ontological conditions has for AI safety. This will be discussed in Part 2.
References
Whewell introduced "consilience" to describe the strongest form of inductive confirmation — when a theory successfully explains facts of a kind different from those it was originally designed to explain. The term has since been adopted more broadly (e.g., by E. O. Wilson) but is used here in its original methodological sense.