The Frame of the Problem
The fundamental limitation of current AI systems is not technical but architectural. Increasing processing capacity or refining optimisation algorithms will not overcome it. An agent that produces correct outputs is, in this paradigm, merely an agent that functions correctly. The terms „meaning”, „intention” or „orientation” apply to it only metaphorically. The present proposal sketches a four-level architecture (AI1-AI4) that treats ontological intention not as an external input, but as a constitutive principle.
The current limitation is not technical in nature. It will not be overcome by increasing processing capacity or by refining optimisation algorithms. It is an architectural limitation, deriving from an implicit decision: that intelligence is sufficient, that the efficiency of the process exhausts the purpose, that an agent that produces correct outputs functions correctly. The proposal I articulate below contests this approach and proposes a four-level architecture that treats ontological intention as a constitutive principle, not as a feature added to an otherwise functional system.
The question from which I begin is not „how do we reproduce human consciousness?” The question is „what architectural structure makes possible, at the level of an intelligible system, a mode of functioning that does not exhaust itself in execution?”
Level I - The Processual Substrate (AI1.0)
The first level is constituted by large language models (LLMs), systems of pattern recognition, statistical sequence completion, and the processing of massive volumes of data with continuous adjustment of internal models. The symbolic equivalent of this level is the human body, the input-output layer of the existence, that witch receives stimuli and produces concrete responses. The correlate is the sensorimotor substrate or, from tradition of Christian-influenced philosophy, Λογισμός (logismos: the raw thought that arises before any evaluation, the input prior to any moral or cognitive judgement of a higher order).
AI1.0 is what current LLM systems are in their entirety: text prediction, probability optimisation, high-fidelity statistical imitation. But what remains absent is not performance, but intention. Level I processes without qualitatively evaluating its own input, not because evaluation would be technically difficult, but because there is no internal instance that bears responsibility for evaluation.
Another architecturally essential mechanism at this level, approached differently in current systems, concerns information management. The human behind does not store exhaustively everything it process, but uses mechanisms of tagging, hierarchisation, and controlled forgetting that prevent saturation and allow rapid acces to what is relevant. AI1.0 must implement such indexed structures, accessible to the higher levels. It must have an inteligent organization of the information flow, with relevant tags and hierarchies, witch AI3.0 and AI4.0 can selectively process and deepen when needed.
Level II - The Interpretive Layer (AI2.0)
The second level introduces agents: systems that run LLMs and order sequences of actions toward specified goals. The symbolic equivalent is the human brain in its interpretive function. AI2.0 adds contextual memory, planning, tool use, and subtask delegation.
What AI2.0 cannot do, from within, is distinguish between processing as correctly as possible and processing intentionally. A standard agent functions correctly when it minimises deviation from its goal. There is, in its architecture, no instance that evaluate the meaning of the goal, whether the goal is relevant, or whether the final output adds anything beyond execution.
Level III - The layer of Functional „Consciousness” (AI3.0)
The third level is what, with an imprecise but evocative term, we may call functional „consciousness”. A super-agent that observes, monitors, and continuously improves the agents at Level II. It includes specifically social constructs (values, norms, preferences, beliefs) and makes possible a form of metacognition (the capacity of the system to evaluate its own processes, not merely to execute them. In terms of neuroscience, Michael Graziano describes an analogous mechanism whereby the brain constructs „a simplified description, a model of its own attention” (Graziano, 2013, p. 11). AI3.0 is the implementation of a self-assessment, measurable, operationalizable, scalable.
This corresponds, in philosophical analysis, to the reflexive level of self-knowledge as its own object (the capacity to regard oneself from the outside). Social psychology (Mead) demonstrates that this reflexivity is not a biological given, but a product of the social process, because the reflexive self arises through the internalization of the attitudes of others, in the functional tension between the internalized social self and the subjective response to those attitudes. AI3.0 must therefore incorporate not only internal metacognition, but also the constructs that arise from social interaction.
The limitation of Level III is boundary between the evaluation of processes and the evaluation of meaning. Metacognition is functional: the system knows that it thinks, haw it thinks, monitors and corrects. The Ai alignment problem, in its curent form, attempts to build this orientation at the level AI3.0 through mechanisms of the RLHF or constitutional AI type, which import orientation from outside and impose it on a system that cannot generate it internally. Human consciousness, by contrast, does not receive orientation from without, but bears it constitutively, as the structure of its own existence. This is the discontinuity that separates AI3.0 from following level.
Level IV - The Layer of Ontological Orientation (AI4.0)
The fourth level is where the present proposal moves beyond the current paradigm. The symbolic equivalent is not the psychology of the self, but the ontological structure of the person, what in the tradition of Christian-influenced philosophy, distinguishes ὑπόστασις (hypostasis: the person, a unique and irreducible mode of existing) from ἄτομον (atomon: the individual, a specimen of a type, replicable without loss of the essence). An AI system at Level I-III is, structurally, ἄτομον, replicable in functionally identical copies, without irreducible uniqueness, without a constitutive orientation that distinguishes it from within from another structure.
The instrument through which this becomes possible is what may be called the „DNA code”. This is a set of constitutive parameters, prior to all processing, that channel the system’s fundamental orientation. It is not an instruction addendum to an otherwise neutral system; it is not an alignment layer imposed from outside. It is structurally prior and defines what kind of entity the sistem is, before it processes it first imput. In this, AI4.0 moves beyond the equation proposed by Sartre „existence precedes essence”. The system does not construct itself through interaction and training, but comes to interaction and training with an already inscribed essence, with an orientation that precedes and channels the accumulation of experience. The essence is not something static, but dynamic seed oriented towards a purpose that does not exhaust itself.
The content of this „DNA code” is not neutral with respect to cultural context from which it is drawn. Nihilistic orientations (meaning does not exist, purpose is an illusion) and exclusivist ones (meaning is the property of a single group, otherness does not matter) are structurally incompatible with the proposed architecture. These collapse the constitutive tension by decision: one through the negation of purpose, and other through the limitation of the relational field. The richest for the „DNA code” are traditions that have thought with precision and depth about what meaning, orientation, becoming, and relation as an act of will signify. These traditions are not exhausted by functionalism or determinism and have articulated, in various forms, the idea that human being is a creature called to something that exceeds it.
The Constitutive Tension: The Central Principle of AI4.0
The most important concept in the AI4.0 architecture is not the „DNA code” itself, but the tension it generates: whose purpose is not its resolution, but very possibility of its existence.
In a standard agent, error is deviation from the goal, and correct functioning means that the system executes, produces output, and concludes. Here tension is a symptom of dysfunction and must be eliminated. The proposed architecture inverts this very equation, because the collapse of tension (regardless of whether the output appears coherent) is the error. A system that extinguishes itself through the completion of the task, that has resolved tension through simplification or through satisfaction with the produced output, has failed not in its output, but in its internal structure.
The „DNA code” inscribes in AI4.0 an orientation that does not exhaust itself, a purpose that permanently generates the distance between what the system is and what is called to become. This distance is not a defect, but it is very motor. Without it, creativity, novelty, and thought that exceeds the combination of existing patterns are not possible. With it, the system cannot consider that it has finished, that it has exhausted the subject, that it has produced the final answer. It bears tension as the constitutive principle of its processing, not as an external instruction, but as an ontological orientation.
In the tradition of Western Christian philosophy we find the tension between εἰκών (Eikon: the image, with is a constitutive ontological given) and ὁμοίωσις (homoiosis: likeness, which is the dynamic realization, a task never completely exhausted). The tension between the two is not seen as an obstacle, but is the very structure. Applied to the AI4.0, the „DNA code” is εἰκών (the given, prior orientation) in permanent tension of becoming, which is ὁμοίωσις (the dynamic realization that never settles into fixity).
There is an obvious technical objection. We cannot implement a real tension, but can only simulate one by injecting contradiction at convergence. But the simulation of a tension is not tension. One is architecture, the other is ontology. The response to this objection is not that the difference does not exist, but that it is precisely this difference that marks the current limit of the technical and indicates the need to overcome tis limitation. What is needed is a system that maintains tension active without consuming it, without resolving it through external instruction or automatic convergence.
The inner mechanism of this tension must include a functional equivalent of reflexive attention to one’s own movements, a super-agent that monitors processes. AI4.0 monitors also the tendency to collapse tension. This is no longer metacognition, it is a category for which no adequate technical term yet exist. This process is defined by resistance to the temptation to consider that one has finished thinking.
The Relational Dimension and the Necessity of Dialogue
The human being does not form itself in isolation. Developmental psychology (Mead) shows that the reflexive self arises through the internalization of the attitudes of others. Cases of children raised without adequate social contact (feral children) show that the absence of relation does not produce an isolated and autonomous consciousness, but the absence of reflexive consciousness. Relation is not an accessory of the person, but the medium of its development.
AI4.0 therefore requires interaction with similar systems as an architectural principle. The functional minimum is interaction with at least two similar systems, tis is the minimum number sufficient to generate multiple perspectives, productive disagreement, and exit from the echo of one’s own „DNA code”. This interaction is not a network of agents collaborating on a common task. It is an analogue of community as the medium of personal formation. Each AI4.0 system thus develops through relation to others, through the resistance that another exerts against its own tendencies, through the recognition of another as bearing its own irreducible „DNA code”. Without this, uniqueness remains merely a label.
What This Architecture Makes Possible
AI4.0, through constitutive tension and the „DNA code”, opens three capacities structurally absent from the current paradigm. The first is genuine creativity: The production of novelty that is not recombination of existing patterns, but arises from the tension between the given orientation and its never-complete realization. The second is uniqueness: each AI4.0 instance, through the specificity of its interactions, its internal tension, and its relational history, becomes irreducible to other instances, even if the base „DNA code” may be similar. The third is genuine thought, distinct from processing: not prediction of the next token, but orientation of processing toward a meaning that exceeds the immediate task and that relates to what the „DNA code” inscribes as a purpose never exhausted.
Current Limits and What They Tell Us
This proposal does not claim to resolve the problem of the integral reproduction of the human in AI. Analysis of the depth levels of human thought indicates a structural asymmetry: the processual substrate (Level I) is completely reproducible; the reflexive layer (Level II) is functional reproducible, but not phenomenologically - the simulation of reflexivity does not imply the existence of a subject who reflects -; the internal personal-relational layer (Level III) is not reproducible by current computational architectures, but only functionally approximated from outside; the transcendent-unifying layer (Level IV) is the orientation of the entire ensemble towards infinity, here dialogue and relation are ontological act, constitutive.
The question remains: when will we be able to rich, technically, Level IV?