At the table someone asks, "Can you pass the salt?" You do not say "Yes"; you pass the salt. The literal wording asked about the function of your arm, and you skipped that layer completely.
The issue on whether meaning can be computed covered how the literal meaning gets composed. This one is about the other half: computing the literal meaning is only step one, and immediately after it you also have to work out why the person said it in that way, in order to land on what they actually want you to do.
The machine has a name. Grice in 1975 called it the cooperative principle. Its maxims are humble — what you say should be true, be enough, be relevant, be clear. These are not orders; they are defaults you assume the other person is holding to. Because you assume they are, whenever they seem to break one, you automatically supply the missing sentence that would put them back inside the rules. That supplied sentence is the conversational implicature.
This issue follows three examples: why questions are often not questions, why the word some carries a silent "not all" inside it, and why irony survives being flipped word-for-word.
Searle in 1975 gave sentences like this a name — indirect speech acts: the grammar is a question, a statement, or a command, and what the sentence is really doing is something else. Understanding it takes one extra step: first see what the wording asks, then ask what asking that at this moment could accomplish, and finally land on that accomplishment.
Clark and Schunk in 1980 ran a clean test of how cleanly the step is taken. They compared people's replies to indirect questions at different politeness levels. When asked Can you tell me where Jordan Hall is?, an answer that included both Yes, I can and the directions was judged more polite than one that only gave the directions. That sounds trivial — of course saying more is nicer — but what it demonstrates is stronger: the listener computed both layers, the literal one (whether you have the ability) and the request (tell me), so as to be able to answer the first layer at all. Understanding is not "try literal first, back off to implicature only if literal fails." Both run in parallel. Strength of evidence: several materials replicated, the direction is clear, and it has been cited widely since.
Languages leave different amounts of room on this ramp. Japanese hardly permits the bare imperative in adult contexts; requests default to shells like 〜てもらえますか (-te moraemasu ka, roughly "could I have you do it"), and children are drilled onto that track from age two and a half. Mandarin 你倒是说话啊 (nǐ dào shì shuō huà a, literally "you actually speak") reads as a complaint about someone being cold, not as an urge to talk. English Would you mind closing the window? is literally a question about willingness — where the literal answer Yes, I would is in fact a refusal, and second-language learners who miss the convention keep tripping on it. Not every indirect request is a memorised phrase: Gibbs in 1986 showed that a freshly invented indirect question ("You wouldn't happen to know where the bathroom is, would you") is understood the first time, which means the inference route is general, not a lookup table.
"Some of my classmates have been to Japan" — the moment you hear it you infer almost automatically that not all of them have. Where does that inference come from? The literal some commits only to at least one and stays true all the way up to all. But you assume the speaker is holding to Grice's maxim of quantity: say as much as you know. Had they known all of them had gone, they should have said all; saying the weaker some tells you obliquely that they could not have said the stronger word. Inferences of this shape are called scalar implicatures.
Bott and Noveck in 2004 turned this inference into a stopwatch experiment. They asked people to judge sentences like Some elephants are mammals — literally true, but implicating something false — as fast as they could. Literalists snap-answer true; implicature users answer false, because had the speaker known all elephants are mammals, that is what they would have said. Neither answer is more correct — but the implicature users took reliably longer. The slide from literal to implicated meaning is not free; it costs a measurable extra step in the tens to hundreds of milliseconds. That overturned the earlier default that such inferences were near-instant. Strength of evidence: dozens of replications have confirmed the effect; recent work isolates the slowdown to the step of reading the speaker's intent, not to knowing the semantics.
Across languages, the scale is not always the same shape, and neither is the force of the implicature. English or can be exclusive (either this or that, not both) or inclusive (or, both allowed); Japanese か (ka) leans more strongly exclusive; Latin split them into two different words, aut and vel. Implicatures live in the contrast between two or more available expressions; what your language stocks is what your implicatures can be shaped by. So the inference "I heard what they did not say" is bankrolled by the local inventory of ways to say it.
You are drenched in the rain and a colleague greets you with "Lovely weather today." The wording pins a good sky onto a bad one, and yet you do not misread them as filing a weather report. The old view was that you understand irony by mentally flipping the literal sentence. Sperber and Wilson in 1981 offered a different account, the echoic mention theory: irony is not a reversal of the literal, it is a quotation of some prior expectation — what the forecast promised, what everyone was assuming, what you yourself said yesterday — with a layer of "look how well that held up" attached. Understanding irony, then, is not doing negation; it is recognising a borrowed expectation and the mocker's stance toward it.
Winner and Leekam in 1991 built an experiment that could tell irony apart from lying. Children watched short stories in which a character said something contrary to the facts. The stories came in two versions — one made it obvious the speaker knew the truth was otherwise (that makes the counter-factual utterance irony), the other made the speaker seem to believe their own words (that makes it a lie). Below age five, children routinely fused the two; around age six they began to split them reliably. The dividing line is not vocabulary or hearing. It is whether the child can represent both what the speaker believes and what the speaker wants the child to recognise them as believing — irony requires tracking a second-order intention: the speaker not only believes the truth is the opposite of what they said, they also want you to recognise that they do. Drop the second-order intention and irony collapses into a lie. Strength of evidence: the direction has been replicated in several languages and by several teams; the exact age of mastery fluctuates and depends on whether prosodic cues are supplied in the stimulus.
Irony also leans on tone of voice and context. The same "you're really something" delivered with rising pitch and a drawn-out vowel reads as clear irony; delivered flat, in a sincere context, it reads as praise. Mandarin's sentence-final particles hang on the same chain — 你可真行啊 (nǐ kě zhēn xíng a, "you really are something") almost forces the ironic reading through 可…啊 (kě … a), while the bare 你真行 (nǐ zhēn xíng) is much more neutral. Bryant and Fox Tree in 2005 measured English speakers' vocal cues for irony and found reliable falling pitch, slower rate, and vowel lengthening — a recognisable "irony voice" — but no single cue was necessary; it is a stack of cues that gives the listener a probabilistic signal. So understanding irony is three things running together: catching the violated expectation, hearing the irony voice, and reading the second-order intention. Miss any one and your foot goes through the floor.
The cooperative principle works because both sides assume the other is holding to it. But how does that assumption get set up between strangers on the first sentence? When the other person is a novice — a child, a second-language learner, an AI — is your default quietly getting dialled down?
"Some of my classmates have been to Japan — in fact, the whole class has" reads fine. Replace some with the semantically explicit not all, and appending "in fact, the whole class has" becomes a direct contradiction. Cancellability is the fingerprint of an implicature; should the semantic boundary sit exactly on this line?
Echoic theory needs the listener to summon a prior expectation (the forecast, common sense, what everyone was saying). But some ironic remarks have no clear precedent, and you still get them. What is being summoned in that case? Can the "expectation" itself be built on the fly?
Implicature spares one round of open conflict, but it shifts the interpretive cost to the other person. When is "one more literal layer" the more responsible move, and when is "one more implicit layer" the more respectful one? Can you sketch a criterion?