Is talking to a chatbot communication? An Oxford volume takes the question apart

On a third answer, neither word fits. Reading presupposes an author somewhere upstream; here the upstream is a corpus of several hundred billion tokens and a company's tuning decisions. Conversation presupposes a second party with a stake. Diane Proudfoot's chapter goes at this through Wittgenstein's lion — if a lion could speak, we could not understand him — and through recent work on communication protocols emerging spontaneously between software agents. Matti Eklund's "AI and Alien Languages" works the same seam from the semantic side. Elena Esposito has been arguing for years, from a sociological rather than an analytic direction, that what we have is artificial communication, a genuinely new social form. That third answer is not a fudge; it is the claim that our vocabulary predates the phenomenon it is being asked to describe.

Externalism is the engine, and it was built in 2021#

Everything in this book leans on a commitment the editors made five years ago, and a reader who does not grasp it will find half the chapters opaque.

In Making AI Intelligible: Philosophical Foundations (Oxford, 2021, 184 pages, also open access), Cappelen and Josh Dever argued — against most of the field at the time — that the outputs of AI systems can have genuine content. The argument is externalist. Semantic externalism, the view descending from Hilary Putnam and Saul Kripke, holds that what your words mean is not fixed by anything going on inside you. Kripke's case is the proper name: "Gödel" in my mouth picks out Gödel because of an unbroken causal chain of uses reaching back to a baptism, and it keeps picking out Gödel even if everything I believe about him is false. Putnam's case is the natural-kind term: I cannot tell an elm from a beech, but "elm" in my mouth still means elm, because I am embedded in a linguistic community in which somebody can, and I defer to them. Meanings, in Putnam's phrase, ain't in the head.

Run that thought about a machine and the conclusion is startling. If content is fixed by causal and social history rather than by inner states, then the fact that a model has no inner states is not decisive. Its words might mean what they mean for the same reason mine do: because of where they came from. Cappelen and Dever called this de-anthropocentrised externalism.

Their chapter here, "A Hyper-Externalist Manifesto for LLMs", pushes much further. Externalism, they now argue, should govern not only content but "structural and procedural properties" — whether a model is genuinely following a rule, whether its representations are compositional, whether it believes rather than merely supposes. All of that, they say, depends on training history, functional role and relational embedding rather than on internal architecture. From which they draw a conclusion that will not be popular in machine-learning departments: mechanistic interpretability, the research programme that tries to read a model's reasoning off its weights, is "philosophically misguided".

My view is that the first half of this is strong and the second half overreaches. Kripke's own rule-following considerations do support the claim that no internal fact settles which rule is being followed, and interpretability researchers who think they can find belief in a residual stream should read the argument. But externalism about reference was always a thesis about what our words pick out, not about who is doing the picking. Push it as far as the manifesto does and the attribution starts to float free of the machine altogether: if training history and functional role do all the work, the natural subject of the attribution is the corpus, or the company, not the model. That is not a refutation. It is an invitation the chapter does not take up.

The chapter that says the volume's premise is false#

Anandi Hattiangadi and Anders Schoubye contribute a piece titled, without hedging, "The Outputs of Large Language Models Are Meaningless". Their argument has two premises: that certain kinds of intention are needed for an utterance to have a literal meaning on an occasion of use, and that language models cannot plausibly have intentions of that kind, processing as they do tokenised vectors stripped of semantic information and trained only to predict the next token. The conclusion follows. They anticipate the externalist reply — that meaning could derive from a history of meaningful use rather than from present intention — and block it: a model, they contend, cannot have the implicit intentions such a history would require.

Publishing this in a volume whose editors hold the opposite view is the single best editorial decision in the book. It also does something the editors may not have intended. Hattiangadi and Schoubye concede, late on, that model outputs function as if meaningful and can be a route to knowledge. Once that concession is on the table, the gap between their position and the editors' narrows to a question about the word "meaning" rather than about anything you would notice in use. I find premise one persuasive for ambiguity and context-sensitivity, where the speaker's intention really does seem to settle which of two readings is in play, and much less persuasive when generalised to every expression. But the argument is clean, and it is stated in a form that can be attacked. That is rarer than it should be.

Responsibility, trust, and the one chapter that is not about semantics#

Guido Löhr argues that chatbots open what he calls a speech act responsibility gap, distinct from the familiar worry about who is responsible when an autonomous system acts. The point is sharp. Assertion is not just the production of a true or false sentence; it comes with a structure of commitment — the speaker can be held to it, challenged, required to retract, expected to apologise. With a chatbot's output, nobody occupies those roles. The developer did not assert it. The user did not. The model cannot. Löhr is right that this is a new gap and, in my judgement, right that existing responsibility frameworks do not close it; he is less convincing about what to do next, which is a fair criticism to make of a first paper on a subject.

Mona Simion and Christopher Willard-Kyle take aim at what they call the Explanation View of AI Trust — the widely held idea that we may rationally trust a system once we are given a clear account of how it works. Explainability and trustworthiness come apart in both directions, and saying so plainly is useful, because a great deal of AI regulation is currently being drafted on the assumption that they do not.

The chapter I expect to be read outside philosophy is Regina Rini's. Models were trained on human conversation, she notes, and will increasingly train us. Her worry is about autonomy in Kant's sense: chatbots are "autonomy traps", because "their deference to our commands tempts us into venting authoritarian whims, ultimately weakening our own self-control". She develops the argument through Piaget and Durkheim and reaches a conclusion she cheerfully calls bizarre — that we will need to treat mindless chatbots as if they were autonomous persons, for our own sake rather than theirs. The empirical premise here is doing enormous work and is not established; whether habitual command-giving degrades self-governance is a question for psychologists, not philosophers. But the structure of the argument is the best thing in the volume, and it is the only chapter that treats these systems as something people live with rather than something people theorise about.

The objection the volume invites#

Analytic philosophy of language has a habit. A new phenomenon appears, and the existing machinery — Grice on implicature, Kripke on reference, Lewis on convention, Stalnaker on common ground — is wheeled out to meet it. The objection is not that the machinery is bad. It is that every piece of it was built for cases in which the producer of an utterance has a stake in the exchange. Grice's maxims describe what cooperative agents do because they want something from each other. Lewis's conventions solve coordination problems both parties have a reason to solve. Stalnaker's common ground is a shared commitment set that both parties track because being caught out matters to them. Remove the stake and it is not obvious the apparatus is describing anything.

To its credit, the volume has an answer, and it is the externalist one. Reference never did require that the speaker have a stake, or even competence: that was the point of the elm. The division of linguistic labour means my words can mean things I could not explain, and deference does the work my ignorance cannot. The comparative method running through the book — animals, aliens, Wittgenstein's lion, hypothetical superintelligences in Tristram McPherson and David Plunkett's closing chapter on whether such a system could engage in moral reasoning — is a deliberate attempt to strip the anthropocentrism out of the concepts before applying them.

It is a good answer about half the distance. Externalism is a theory of what words pick out. It is not a theory of assertion, and the normative vocabulary — trust, commitment, retraction, testimony, being owed an explanation — does presuppose a party with something to lose. That is precisely where Löhr's gap opens, and the volume does not notice that his chapter is a counter-example to the method the rest of it uses. John Searle made a related point about institutions decades ago and was written off as a reactionary for it.

What is not in the book#

Three absences, in ascending order of seriousness.

First, political economy. The chapters speak of "the model" as though it were a natural kind. It is a product: someone chooses the training data, writes the system prompt, sets the refusal policy, decides which answers are safe and which are commercially inconvenient. "Who is speaking?" has a dull answer that is also largely the right one, and no chapter gives it. Philosophy departments now have an awkward relationship with the labs, which may or may not be relevant.

Second, evidence. There is very little here about how people actually use these systems, and Rini's chapter in particular rests on a claim that could be tested. The human-computer interaction literature has been studying this for years. None of it appears.

Third, and most tellingly, the volume is almost entirely Anglophone-analytic in its resources. The hermeneutic tradition spent two centuries on exactly this question — what understanding is, and whether it can happen without a meeting of minds. Schleiermacher's lectures treat understanding as a reconstructive achievement rather than a transmission; Gadamer treats it as something that happens between parties, in which the question is prior to the proposition and neither side controls the outcome. A tradition that already denies that meaning is delivered intact from a head would seem to have something to say about a case where there is no head at all. It is not cited.

Will it be read in ten years?#

Parts of it. The Cappelen–Dever manifesto and the Hattiangadi–Schoubye rebuttal will be read as a pair, because they are a real disagreement, stated in refutable form, at the moment the question was live; that is what makes a volume survive. Rini's chapter will be read by people who are not philosophers. Löhr's will be cited whenever a chatbot says something actionable and a court has to decide who said it.

The rest will read like 2026, which is no disgrace. There is a smaller irony worth recording. Cappelen's Philosophy without Intuitions (2012) argued that intuitions play far less of a role in analytic philosophy than philosophers claim. This volume, from first page to last, turns on intuitions about whether something counts as communication. He would presumably say the appeals are to theory rather than to intuition. Read the chapters and judge. Elsewhere on this site, David Chalmers has been asking what we are talking to when we talk to these systems, and the honest answer, which this book gets closer to than anything else in print, is that nobody yet knows.

Rivka Weinberg's The Meaning of It All (Oxford, March 2026) argues that a life as a whole cannot have a point, that death is overrated as a threat to meaning, and that time is the real thief. A close reading, then the autumn's other titles: Ganeri's revised Indian philosophy introduction, Han, Žižek and Habermas.