Robert Stalnaker's common ground and the machines that borrow it

Some papers go unused for a generation and then find a readership that was not born when they were written. Robert Stalnaker's 'Assertion', published in 1978, is one of them. Fifteen pages long, addressed to philosophers of language, it proposed that a conversation can be modelled as the steady narrowing of a set of possibilities, and it has sat near the centre of formal semantics ever since. What is new is where the vocabulary has started to appear. In 2026 researchers measuring whether chatbots understand the unsaid have reached for Stalnaker's terms — common ground, context, presupposition — to say what they are testing. Stalnaker himself, who is eighty-six and emeritus at the Massachusetts Institute of Technology, has taken no part in this. His concepts are being handled by people asking a question he never posed: whether they describe anything going on inside a machine that only predicts the next word.

The borrowing is real but looser than the headlines suggest, and the looseness is the story. Before coming to the machines it is worth being clear about what Stalnaker actually built, because most of what is interesting turns on getting that right.

A simple idea: talking is ruling worlds out#

Stalnaker's model rests on an ordinary intuition. When two people talk, there is a body of information they both treat as settled — each takes it as given, each takes the other to take it as given, and so on. That shared stock is the common ground. The set of possible worlds in which everything in the common ground is true, he called the context set: the ways the world might still be, for all that has so far been agreed. To make an assertion is to propose cutting that set down. Say 'the meeting is cancelled', and if no one objects, every possibility in which the meeting goes ahead is struck out; the conversation proceeds inside the smaller set that remains. Informative speech shrinks the context set. To assert what is already common ground is to say nothing; to assert what contradicts it is to misfire.

The reach of the model comes from how much it ties together. Presupposition is just whatever is already in the common ground, taken for granted rather than proposed — which is why 'my sister is coming' presupposes, without asserting, that I have a sister. Implicature can be read off the same structure: a listener works out an unstated meaning by asking why the speaker narrowed the set in this particular way and not another. Stalnaker returned to the machinery in 'Common Ground' in 2002, where he drew a distinction that matters here. The common ground is built not from what the parties believe but from what they accept for the purposes of the exchange. That is what lets us lie, bargain, suppose things 'for the sake of argument', and follow a work of fiction: we feed the context set propositions we do not hold true at all.

The engine underneath is possible-worlds semantics. On this way of doing things the content of a sentence is identified with the set of possible worlds in which it is true, so that meaning and information become set-theoretic objects one can calculate with. A context set is itself a set of worlds; an assertion is a proposition, another set of worlds; and to accept the assertion is simply to intersect the two, discarding every world the new proposition excludes. The appeal is that vague talk about 'what a conversation has established' becomes something precise and tractable: a region of logical space that shrinks as the exchange goes on. The cost, which Stalnaker spent much of his career managing, is that very fine distinctions — between logically equivalent sentences, or between necessary truths — tend to disappear, because sentences true in exactly the same worlds come out with exactly the same content.

Where the apparatus came from, and where it went#

Stalnaker did not arrive at conversation first. His route ran through the logic of conditionals. 'A Theory of Conditionals' (1968) proposed that 'if A, then B' is true when B holds in the most similar possible world in which A holds — the nearest world where the antecedent is satisfied. The Stalnaker conditional became, alongside David Lewis's close variant, the standard semantics for the subject, and the two men's disagreement over whether there is always a single nearest world set the terms of debate for years. We touched on their contrast when David Lewis came up in our 'this day' column: Lewis took possible worlds to be concrete universes as real as our own, while Stalnaker held them to be ways the actual world might be — tools for representing content, not a second cosmos to believe in. That moderate line runs through Inquiry (1984), Context and Content (1999), Ways a World Might Be (2003) and Mere Possibilities (2012), and was restated late, on a deliberately austere footing, in Propositions: Ontology and Logic (2022).

Within linguistics the context-set idea spread until it became a programme. Irene Heim's file-change semantics, Hans Kamp's discourse representation theory and Frank Veltman's update semantics — the family now called dynamic semantics — all treat meaning as a change to an evolving context rather than a static matching of words to the world. One more thread reaches into philosophy of mind: the two-dimensional, 'diagonal' apparatus Stalnaker used in 'Assertion' to handle the way an utterance's truth depends on facts about the context of utterance is among the sources David Chalmers later drew on in his own two-dimensional framework for thinking about meaning and necessity. A paper written for a narrow audience turned out to be unusually portable.

2026: the machines enter, with caveats#

The portability is now being tested on systems no one imagined in 1978. Large language models are trained only to predict the next token; they carry no explicit representation of what the parties to a conversation mutually accept. Yet they often behave as if they grasp the unsaid — handling presupposition, implicature and indirect requests with some fluency. That is a genuine puzzle for a theory that located common ground in the minds of speakers, and it is the puzzle a cluster of recent studies has tried to pin down. What those studies actually establish, and what they merely gesture at, needs separating with some care.

The most substantial of them, 'The Pragmatic Mind of Machines', was released as a preprint in 2025 and accepted to the 2026 meeting of the European Chapter of the Association for Computational Linguistics. Kefan Yu, Rob Voigt and colleagues built a dataset they call ALTPRAG, in which each item pairs two equally plausible continuations of an utterance and asks a model both to infer what the speaker meant and to explain why one continuation would be preferred. Testing twenty-two models across pre-training, supervised fine-tuning and preference optimisation, they report that even base models show a measurable sensitivity to pragmatic cues, that this scales with model and data size, and that fine-tuning adds further gains; pragmatic competence, they conclude, is an 'emergent and compositional' property of training. It is careful work. It is also worth saying clearly that the paper does not build on Stalnaker, and does not use the phrase common ground; its frame is the theory of pragmatic alternatives, closer to Grice than to the context set. Reports that tie it directly to Stalnaker are reading a lineage into it that the authors did not claim.

The one study in the current wave that does cite Stalnaker is 'On the Same Wavelength?' by Cedegao Zhang, Joshua Tenenbaum and Roger Levy, which evaluates pragmatic reasoning in language models across a broad range of concepts and lists his 1978 'Assertion' among the formal foundations of pragmatics it builds on — while taking its working apparatus from the Rational Speech Act model, a Bayesian picture of speakers and listeners reasoning about one another. So the genuine borrowing exists, but it is a citation in a background section, not a theory of common ground implemented and probed.

A third paper, posted to arXiv in August 2026, looks at the same terrain from the direction of security. 'Pragmatic Attack Surface' argues that because safety training attends to what a prompt literally contains rather than to what a reader would infer from it, harmful intent can be smuggled in through implicit, context-dependent language that explicit instructions would not get past. The observation is sharp, and it rhymes with Stalnaker's point that what is quietly accepted, rather than asserted, is the efficient part of human talk — and now, in a machine, the exposed part. But as far as we could confirm the paper does not invoke Stalnaker's framework either. The affinity is one a reader may draw; it is not a result the authors report.

The philosophical question underneath#

Set the citations aside and a real question remains, and it is the one a philosopher should care about. Stalnaker defined common ground in terms of the mental states of the parties — what they mutually accept. That makes his theory, in part, a psychological claim. If a system with no mental states at all can behave as though it is keeping track of common ground, one of two things follows. Either common ground is not, after all, something mental, but a structural regularity encoded in the statistics of language itself — in which case a next-token predictor might carry it without anyone's believing anything. Or the model is not tracking common ground at all, and is reproducing the surface of text produced by people who were. The empirical findings cut both ways. A sensitivity present in base models leans towards the first reading; competence that improves sharply with fine-tuning on human feedback leans towards the second. The data sharpen the question rather than settling it, which is the honest thing to say about them.

This is where Stalnaker's model meets an older argument. John Searle's Chinese Room held that a system manipulating symbols by their shape alone can never thereby mean anything, because syntax does not add up to semantics. Stalnaker locates content not in syntax but in an utterance's effect on a set of possibilities — its content just is the way it narrows the context set. Read strictly, that tempts a deflationary verdict: a system that narrows the set correctly could be said to assert, inner life or no. Read the other way, it cuts against the machine, because if nothing is genuinely accepted there is no common ground to narrow, and what looks like assertion is only its outward form. Stalnaker's own instinct, one suspects, would be the undramatic one: look at what the theory predicts and test it, rather than ask the machine to confess. His Propositions treats the objects of assertion as logical, not mental, items — which may make the apparatus travel into non-human systems more easily than the psychological gloss would suggest, and which is exactly why the travelling needs watching.

Why this matters beyond the labs#

AI debate usually casts philosophy in one of two roles: ethical chaperone, or keeper of the question of whether the machine is conscious. Stalnaker's case points to a third. Researchers who want to measure what a language model does with the unsaid have to define what they mean by pragmatic competence, and a model a philosopher of language built out of possible worlds half a century ago has turned out to be a convenient way to say it. We reported earlier this year on the strange arithmetic by which AI companies are hiring philosophers even as university departments shrink, and on the wider conversation about what we are really doing when we talk to these systems. The afterlife of 'Assertion' is a small, precise instance of the same pattern: a piece of work written for no practical end at all becomes useful precisely because it was trying only to be true. The thing to resist is the shortcut that turns a borrowed vocabulary into a finding. That a 2026 paper says common ground does not mean a machine has it.