What Participants Mean But Never Say: Implicature in Interview Answers (2026)

Ask a customer Would you recommend us to a colleague? and get back I would recommend you to a certain kind of colleague. Nothing in that sentence is negative. Every word of it is true. And every human reader instantly understands it as a no.

That gap -- between what a sentence states and what it conveys -- is called conversational implicature, and it is the single largest source of silent data loss in interview analysis. A literal reader codes the answer above as a qualified yes. A competent reader codes it as a polite refusal. They are reading the same transcript.

The practical test is one line long, and it is the only thing from this article you strictly need: try to cancel the inference out loud. If you can append a contradiction without the sentence becoming nonsense, what you were about to code as a finding was an implicature, not a statement. I would recommend you to a certain kind of colleague -- in fact, to anyone is odd but coherent. That oddness-without-contradiction is the signature of an implicature.

The four maxims, and what flouting one tells you

Paul Grice's account starts from a cooperative principle: contribute what is required by the accepted purpose of the conversation. Under it sit four maxims. The Stanford Encyclopedia of Philosophy states them compactly: Quality -- make your contribution true; so do not convey what you believe false or unjustified; Quantity -- be as informative as required; Relation -- be relevant; and Manner -- be perspicuous; so avoid obscurity and ambiguity, and strive for brevity and order.

The useful part for research is not the maxims themselves. It is the difference between breaking one quietly and breaking one obviously. When a speaker violates a maxim quietly, the listener is simply misled. When a speaker flouts a maxim -- breaks it so visibly that both parties notice -- the listener is meant to notice, and to work out what the speaker meant instead. A flout is a deliberate signal. In an interview it is almost always the most informative moment in the transcript, and almost always the moment a literal coding pass throws away.

Flouting Quantity: the answer that is too short

Grice's own illustration is a reference letter. A professor writes that a candidate's command of English is excellent and that his attendance at tutorials has been regular -- and by saying remarkably little about the candidate's philosophy, implicates that he is unsuitable for a philosophy post. Nothing critical is written down. The criticism is carried entirely by the omission.

Customers do this constantly. Asked what they think of a new dashboard, a participant who answers the loading speed has really improved has told you the visual design and the information architecture are not worth mentioning. Code that as positive sentiment about performance and you have recorded the opposite of the message.

Flouting Quality: saying the false thing on purpose

Irony, sarcasm, and overstatement flout Quality. Oh, the migration went beautifully after a described outage is not a positive data point. Sentiment scoring built on lexical polarity is structurally unable to see this, because every word in the sentence is positive. This is one of the reasons sentiment scores and human coders diverge most sharply on the accounts that churn.

Flouting Relation: the answer to a different question

You ask about pricing; the participant answers about the sales process. A literal reader records a non-response. A pragmatic reader records that the participant has declined the pricing question and substituted the thing they actually want on record -- which is usually the stronger finding.

Flouting Manner: the conspicuously long way round

When someone stops saying our admin and starts saying the person who currently has the admin seat, the detour is doing work. Obscurity and prolixity are expensive; speakers pay that cost when plain phrasing would commit them to something they do not want to commit to.

The some and all problem, where literal reading loses most

The most measurable implicature is scalar. Using a weaker term on a scale implicates that the stronger one does not hold: some implicates not all, sometimes implicates not always, it works on most screens implicates it does not work on all of them.

Strictly, some is logically compatible with all. A reader who insists on the logical reading is not making an error of grammar; they are making an error of cooperation. Ira Noveck's 2000 study in Cognition, memorably titled around the finding that children are sometimes more logical than adults, put numbers on it: only 41 percent of adults accepted pragmatically infelicitous statements of this kind, against 89 percent of eight-year-olds and 85 percent of ten-year-olds. The children accepted the literal reading. The adults heard the implicature and rejected it.

That result is worth sitting with, because the literal reader in your pipeline is usually not a person. When a participant says some of my team logged in this week, a literal coder records something consistent with full adoption. Every adult listener in the room heard not all of them, and I am telling you that on purpose.

The cancellation test, and why it is not the negation test

Implicatures are cancelable. That is their defining property and it is also your diagnostic.

Running the test

Take the inference you are about to write into the report. Append an explicit contradiction of it to the participant's sentence. Then ask whether the result is a contradiction or merely strange.

• Some of my team logged in -- in fact, all of them did. Strange, not contradictory. The not all was an implicature. It is real, it is intended, and it is not a fact the participant asserted.
• Only some of my team logged in -- in fact, all of them did. Now it is a flat contradiction. Only makes the exclusion part of what was said.

The first belongs in your analysis as an inference you are drawing. The second belongs in it as a claim the participant made. Writing them into a report in the same register is the error this test exists to prevent.

What a failed cancellation means for your evidence

If the cancellation produces a genuine contradiction, you are holding an assertion, and you can quote it and count it. If it only produces oddness, you are holding an inference -- well grounded, usually correct, and still yours rather than theirs. Report it as a reading of the evidence, not as the evidence. When several analysts independently draw the same implicature from the same passage, that agreement is worth recording; when only one does, you have found a place where the interview needed a follow-up that never happened.

Where this stops and the presupposition article starts

These two look similar and behave in opposite ways, which is exactly why they get confused.

A presupposition lives in the question and survives denial. An implicature lives in the answer and dies under cancellation. Negation is the test for the first; cancellation is the test for the second. If you want the question-side problem -- what your own wording smuggles into the participant's head before they speak -- that is presupposition, and it is a different article on purpose. This one is about what comes back.

Two further boundaries. Response timing is its own signal and is handled separately; a pause before an answer is not an implicature. And the analyst's own hedging in a written finding is a different question from the participant's hedging in an answer.

How Koji handles this

Reading implicature requires a listener who can act on it during the conversation, which is precisely what a static questionnaire cannot do. Koji's AI interviewer is built around that gap:

• Follow-ups fire on the flout, not on a timer. When an answer is conspicuously short, conspicuously off-topic, or scalar where a definite answer was expected, Koji asks the natural next question -- when you say some of your team, roughly how many is that? -- and turns an inference into an assertion while the participant is still in the room. That is the single highest-leverage thing an interview can do with an implicature, and it is unavailable to SurveyMonkey, Typeform, or any other form-based tool once the respondent has clicked submit.
• Structured questions remove the ambiguity where you do not want it. Koji supports six question types -- open_ended, scale, single_choice, multiple_choice, ranking, and yes_no. Use the open_ended type where implicature is the point and you want the participant's own framing; use scale or single_choice for the quantities you refuse to infer. A some that matters should be a number, and the structured questions guide covers how to mix the two in one study.
• Analysis codes meaning, not wording. Koji's transcript analysis is instructed to code the meaning of an answer rather than its surface text, and to record a confidence level on each extraction. Low confidence on a short answer is frequently a flouted maxim sitting in plain sight.
• Quotes stay attached to their context. Every theme in a Koji report links back to the passage it came from, so a reviewer can check whether a coded positive was ironic before it reaches a roadmap decision.

Frequently asked questions

What is conversational implicature in a user interview?

It is what an answer conveys without stating it. The loading speed has really improved implicates that nothing else improved. The test is cancellation: if you can append a contradiction and get strangeness rather than a flat contradiction, you are looking at an implicature rather than an assertion.

What is the difference between flouting and violating a maxim?

A violation is quiet and leaves the listener misled. A flout is obvious, and the obviousness is the point -- the speaker intends you to notice the break and to work out what they meant instead. Flouts are signals. In an interview transcript they mark the most informative passages.

Can I quote an implicature as a finding?

Quote the words; report the inference separately. If the cancellation test yields only oddness, the participant did not assert what you inferred, and a report that presents the two identically overstates its evidence. Flag it as a reading, or go back and ask.

Why do sentiment tools miss this?

Because flouted Quality is positive on the surface. The migration went beautifully, said after a described outage, scores as positive on every word. Lexical sentiment scoring has no access to the mismatch between the sentence and the situation, which is where the meaning lives.

Does this mean some always means not all?

No, and that is why it is an implicature rather than a fact. It is defeasible: context, or the speaker, can cancel it. Treat some as a strong signal of not all that is worth one clarifying question, and let Koji ask that question rather than resolving it yourself in the spreadsheet.

How does Koji help with implicature specifically?

By probing during the interview instead of after it. Koji's AI follow-ups target short, evasive, and scalar answers and convert them into explicit statements, so your analysis holds assertions rather than inferences. Voice and text interviews both support this, and the resulting quotes stay linked to their transcript context.

Related Resources

• Structured questions in AI interviews -- the six question types, and when to force a number instead of inferring one.
• Presupposition in interview questions -- the mirror image: what your question asserts before the participant answers.
• Probing questions in user interviews -- how to follow a flouted maxim without importing a premise.
• Avoiding leading questions -- keeping your own wording out of the answer you are about to read.
• The complete guide to thematic analysis -- coding meaning across a corpus once you have decided what each answer meant.
• The graded hedge -- how to write a finding you can only partly support.