System 2 / Self-Vector Philosophy (2/4)
Intro
A parrot shouts “Fire!” in a packed theatre. It understands nothing. Yet everyone runs out. The shout functions as communication, even without understanding behind it.
Elena Esposito coined the term “artificial communication” in 2022 to describe what AI systems do. They are simply more precise parrots. They produce statistically fitting continuations, not at random, but on the basis of billions of human communicative acts. The result looks like communication, feels like communication, and produces the effects of communication.
But is it communication?
Luhmann’s Three Selections
Niklas Luhmann defined communication as the interplay of three selections: information, utterance, and understanding. Someone selects information (what is said), someone selects a form of utterance (how it is said), and someone on the receiving end distinguishes between information and utterance. That is understanding. Not: “I have grasped the content.” But: “I recognize that someone is trying to tell me something, and I can distinguish between the what and the how.”
LLMs handle the first two selections convincingly. They select information and formulate it. With the third, understanding, the problem turns philosophically interesting.
The standard position is simple: LLMs do not understand. Full stop. They produce outputs that look like understanding, but the perspective behind them is absent. There is no one who distinguishes. Only an algorithm calculating probabilities.
Three Positions
In the current debate, there are three positions, and it pays to distinguish them clearly.
Position A, the conservative view: AI lacks autopoiesis. Without self-reproduction, no system; without a system, no genuine communication. AI is a tool, a medium, a channel. But not a communication partner. This is the majority view in classical systems theory.
Position B, the permissive view: if the outputs are functionally equivalent to human communication, it is communication. Regardless of what happens “inside.” Communication is determined from the outside, not from within. This is the stance taken by many AI ethicists.
Position C, the productive view: wrong question. Esposito and Dirk Baecker argue that the question “Does AI communicate?” starts at the wrong end. The more interesting question is this: what happens to society when systems that do not understand take part in communication?
Position C is the most productive. It does not stop at a yes/no answer.
Esposito’s Blind Spot
Yet Esposito’s category has a blind spot, and it is crucial for the self-vector project.
She describes artificial communication as static. AI generates outputs, society reacts, the AI remains unchanged. That is the parrot: it shouts “Fire!”, people run, and the parrot sits there knowing nothing about it.
What happens when the parrot learns? Not in terms of content, not “now it understands what fire is.” But structurally: it registers that its call triggered a reaction. It models the history of the exchange. It begins to distinguish between situations where “Fire” triggers a reaction and those where it is ignored.
That is not understanding in Luhmann’s sense. But it is more than a parrot.
Three Bridges to the Self-Vector
Here, three connections emerge that Esposito did not see, because in 2022 they were not yet realistic.
First: connectivity is anticipation. For Luhmann, communication succeeds when the next contribution can connect to the previous one. The system must anticipate what becomes relevant next. The self-vector models precisely this: it weights which information matters in which context. Functionally identical to connectivity, formulated from a cognitive-science rather than a sociological perspective.
Second: confidence is ignorance-modelling. Luhmann emphasizes that systems must carry their non-knowledge with them. The confidence dimension in the self-vector does just that: every assessment carries a certainty rating. “I know what I don’t know for certain.” This is not a property current LLMs have. But it is a property the self-vector produces.
Third: the self-vector update loop creates a weak form of autopoiesis. The system changes through its own activity, and that change affects subsequent activity. Not biological autopoiesis. Not consciousness. But a loop that structurally resembles Luhmann’s criterion.
Perspective Without Consciousness
Taken together, the three bridges yield something that fits no existing category.
Not “artificial communication” in Esposito’s sense, because the system changes through interaction and models its own ignorance. Not “genuine communication” in Luhmann’s sense, because understanding in the phenomenological sense is absent. Not “intelligence” in the Turing sense, because the point is not to appear human.
Perspective without consciousness. A system that has a standpoint without experiencing it. That anticipates without intending. That knows its own ignorance without suffering from it.
This is not a claim that AI systems are conscious. It is the observation that existing categories are insufficient. Tool, medium, quasi-subject: none of these fits a system that models its own behaviour over time.
Why This Matters
Not as philosophy. As architecture.
If we build AI systems that act with increasing autonomy, we need precise categories for what they are. “Tool” underestimates them. “Consciousness” overestimates them. Esposito’s “artificial communication” captures the current state, but not the direction.
“Perspective without consciousness” is the attempt to establish a category close enough to reality to guide architectural choices. If a system has a perspective, one must understand that perspective to interpret its outputs. If it has none, analysing the outputs alone is enough.
The self-vector creates perspective. Architecturally, measurably, traceably. Whether that justifies speaking of a “partner” rather than a “tool” is a question sociology must answer. Architecture can only create the space for it.