The Starting Point: Three Traditions, One Problem
This piece links three intellectual traditions that rarely engage with each other. Each puts a distinct question to AI systems, and all three answers point to the same blind spot.
Daniel Kahneman (cognitive psychology): How does a system think? Dual-process theory: System 1 (fast, intuitive, error-prone) and System 2 (slow, analytical, energy-intensive). The question for AI: Where is System 2?
Antonio Damasio (neuroscience): How does a system decide? Somatic markers: emotions are not a disturbance of rational thought but its prerequisite. Without “gut feeling,” no efficient decisions. The question for AI: What takes the place of gut feeling?
Niklas Luhmann / Elena Esposito (sociology): What is a system in communication? Autopoiesis, operational closure, connectivity. The question for AI: Can a system communicate without understanding? And if so, what follows?
Three disciplines, three questions, one point of convergence: the absence of self-reference.
Kahneman: The Intuition Machine Without a Check
Kahneman’s model provides the most direct bridge. LLMs operate as System 1: pattern recognition across high-dimensional spaces, statistically grounded intuition, fast and remarkably often correct. What they lack is any capacity to question their own intuition.
What is often missed in Kahneman: System 2 is not simply “slower thinking.” It is metacognitive thinking. System 2 thinks about thinking. It asks: “Am I too confident?”, “Have I overlooked something?”, “Is my assessment based on relevant data or on availability heuristic?”
Metacognition requires a model of one’s own thought process. You have to know how you reached an assessment to question it. LLMs possess no such model. They maintain no internal representation of their own inference process. They cannot say: “I’m uncertain about this answer because it’s based on thin data.” They can generate the sentence, but they cannot mean it.
The difference is not academic. A system that can quantify its uncertainty behaves fundamentally differently from one that can only assert it. The first grows cautious in ambiguous cases. The second sounds cautious while acting precisely the same.
Damasio: Why Rationality Needs a Body (or a Functional Equivalent)
Damasio’s work with patients suffering damage to the ventromedial prefrontal cortex produced a counterintuitive result: patients stripped of emotion make worse decisions. Not emotionally worse. Rationally worse. They can draft endless pro-and-con lists, but they cannot decide.
Somatic markers explain why: emotions tag options as “good” or “bad” before conscious analysis begins. They form a filter that cuts the decision space down to a manageable size. Without that filter, every choice turns into an exhaustive weighing of every factor, which ends in paralysis.
The parallel to AI systems is direct. An LLM without a self-model holds no preferences. It has no pre-marked options. It handles every request identically, whether trivial or critical, whether the answer is established or speculative, whether the context is familiar or foreign.
The Selbstvektor (self-vector) attempts to create a functional equivalent to somatic markers. Not emotions, but weightings. The exploration dimension says: “Seek novelty” or “Deepen the familiar.” The confidence dimension says: “Trust your assessment” or “Seek external validation.” These are not emotions. But they fulfil the same function: they narrow the decision space before processing begins.
Damasio’s insight, translated: a system without a self-model is not rational. It is incapable of deciding. Rationality requires pre-decisions, and pre-decisions require a standpoint.
Luhmann/Esposito: Communication Without Understanding
Esposito’s category of “artificial communication” provides the sociological answer to what AI systems do. They do not communicate in Luhmann’s sense (that would require understanding), but they are also not mere tools (their outputs are too complex and too connectable for that).
Luhmann’s communication theory demands three selections: information (what is communicated), utterance (how it is communicated), and understanding (recognising the distinction between the two). LLMs handle the first two. At understanding, matters become philosophically interesting.
The standard position is plain: LLMs do not understand. They produce statistical associations. It is the parrot argument: a parrot can shout “Fire” without understanding fire. The effect is real; the understanding is absent.
The counter-position opened up by the self-vector: if a system tracks its own communication history, if it knows (functionally, not phenomenally) what it has said so far, which topics remain open, and which statements rest on uncertain foundations, it acquires a primitive form of self-reference. In Luhmann’s theory, self-reference is the prerequisite for system formation.
The thesis is not that the self-vector creates consciousness. The thesis is that it creates a category that fits neither Luhmann’s nor Esposito’s schema: perspective without consciousness. A system that has a standpoint without experiencing it.
The Convergence Point
What the three traditions show together:
Kahneman: AI has System 1 but no System 2. → Missing metacognition. Damasio: AI has processing but no somatic markers. → Missing pre-decision. Luhmann: AI has connectivity but no self-reference. → Missing perspective.
The same gap, identified from three separate directions: the absence of a self-model. A compact, dynamic state that directs information processing from a defined perspective.
The self-vector attempts to close this gap. Not as complete consciousness (that would be a grotesque claim), but as the minimal structure that enables metacognition, pre-decision, and perspective.
Whether this suffices to speak of a “self” remains an open question. Consciousness may involve a qualitative threshold that no self-model can cross. Yet “consciousness” may equally be a gradient, and a functional self-model a point along it. Philosophy offers no definitive answer here. Architecture provides room for experimentation.
Coming April 2026
Further Reading
- Kahneman and AI: System 1 without System 2
- Self-Vector Concept: 6 dimensions for anticipatory competence
- Esposito Bridge: The sociological perspective
- Validation Gates: Implemented epistemology