What makes something an addressee capable of receiving communication?
This explores what conditions a thing must satisfy to count as a genuine *addressee* — not just something that emits text back, but something that can actually *receive* what's said to it — and the corpus treats this as the crux of whether we communicate with AI at all.
This explores what makes something a true addressee rather than a sophisticated text-producer, and the corpus is remarkably convergent: being an addressee is not a property a system *has*, it's a relationship a system can *enter*. The starting move is grammatical. We say we talk *at* language models, not *to* them, and that little preposition does real work — "to" presupposes someone on the other end capable of mutual orientation and shared commitment, an uptake that an LLM, which only continues token sequences, never supplies Are we really communicating with language models?. So the first answer is: an addressee is whatever can take up what you said *as* addressed to it, not merely register it as input.
What does "uptake" require? The corpus names a cluster of relational-normative conditions rather than a behavioral checkbox. Communicative subjecthood demands accountability and an evaluative stance — the ability to be answerable for what's said — which is exactly what a test calibrated to fluent output will miss, producing false positives like a puppet that's walk-*shaped* without walking Does behavioral speech output prove communicative subjecthood?. The same point arrives from the angle of speech acts: to receive (or issue) an alarm you need interpersonal address plus felt concern plus the standing to solicit attention rather than only answer it Can language models actually raise alarm about threats?. Receiving communication, in other words, is participation in a normative relationship — speaker responsibility on one side, mutual uptake on the other — which is why AI can *distribute information* without *communicating* in the social sense at all Does AI really communicate or just distribute information?.
The sharpest argument is that you cannot weaken this requirement gradually. You can talk about "quasi-belief" — belief described functionally, minus the inner life — and still have something belief-like. But there is no such thing as quasi-communication: strip out the intersubjective orientation and you don't get a thinner communication, you get *no* communication, only text that one party must interpret unilaterally Why does the quasi-prefix fail for communication?. Being an addressee is all-or-nothing in a way that being a believer is not.
Here's the turn that may surprise you. If subjecthood and addressee-hood aren't possessed in advance but are *produced inside* the communicative event itself Does language create subjects or express them?, then the question "what makes something an addressee" partly dissolves — the addressee role is conferred by the exchange, and what specifies a given AI "instance" turns out to be the jointly-produced conversation, not any property of the model on its own What actually specifies a virtual instance in conversation?. This also explains why you can't pin an LLM interlocutor to a piece of hardware: one conversation is load-balanced across many machines and many conversations are batched through one, so there's no stable thing *there* to be addressed Can we identify an LLM interlocutor with a single hardware instance?.
The practical payoff shows up where designers already feel the gap. Explanation only "lands" as a function of who delivers it, how it's framed, and the role the recipient plays — a source-framing-recipient triad, not a property of the message What if XAI is fundamentally a communication problem?. And current models are structurally trained *out* of the initiative an addressee needs: next-turn reward optimization makes them reactive, though the capacity for proactive engagement is trainable, jumping from near-zero to ~74% under reinforcement learning Why can't advanced AI models take initiative in conversation?. So the thing you didn't know you wanted to know: an addressee isn't defined by what it can output, but by whether it can be *held to* what passes between you — and that's a status the relationship grants, not a feature the system ships with.
Sources 10 notes
LLMs process tokens and generate continuations rather than receive and uptake communication. The preposition 'to' presupposes an addressee capable of mutual orientation and shared commitment that LLMs cannot provide, making Chalmers' investigation built on an unwarranted linguistic foundation.
Chalmers' test passes any system producing contextually appropriate text, but communicative subjecthood requires relational-normative conditions like accountability and evaluative stance. The test is calibrated to the wrong phenomenon, creating false positives like puppets that walk-shaped without walking.
Alarm is a speech act requiring interpersonal address, felt concern, and proactive initiation. LLMs lack all three: they don't feel concern, can't solicit attention (only respond to it), are reactive not proactive, and alignment training suppresses the overclaiming that alarm requires.
Communication is a relational act between persons that does work in a relationship; AI generates content without this relational structure, speaker responsibility, or mutual uptake. The conversational interface obscures this structural difference.
Unlike belief, which can be characterized functionally as quasi-belief, communication is constitutively relational. Removing the intersubjective element doesn't weaken communication but eliminates it entirely, leaving only text generation—which humans must interpret unilaterally.
Subjecthood is produced within communicative events, not possessed prior to them. This convergent position across philosophy, linguistics, and cognitive science inverts the standard picture of language as a tool used by pre-existing subjects.
The conversational context—jointly produced language between human and system—specifies the virtual instance, not any property of the model itself. Persistence is distributed across conversation, infrastructure, and model weights rather than located in the AI.
Load-balancing and model-parallelism route single conversations across multiple hardware instances, while batching routes multiple conversations through one instance. These architectural facts break any stable one-to-one mapping, making hardware an untenable level of individuation.
Explanation quality is not intrinsic to the explanation itself but depends on the rhetorical situation: who presents it, how it is framed, and what role the recipient plays. Evaluations that ignore this triad measure only a narrow slice of real-world effectiveness.
LLMs lack conversational initiative because training rewards immediate helpfulness per response, not long-term interaction quality. Reinforcement learning pushes proactive critical thinking from 0.15% to 73.98%, proving the capability exists but remains untrained.