Presupposition: The Part of Your Question Participants Cannot Decline (2026)

A leading question pushes a participant toward an answer. A presupposing question does something stranger and more durable: it embeds a fact that the participant has to accept in order to answer at all. The difference matters because the standard fix for a leading question -- strip the loaded adjectives, keep it neutral -- does nothing to a presupposition. What stopped you from upgrading? contains no loaded words whatsoever, and it still hands the participant a premise they did not choose.

The practical test is one line long, and it is the only thing from this article you strictly need: negate the question and see what survives. If a fact is still standing after the negation, it is a presupposition, not a claim, and your participant cannot decline it without derailing the interview to argue with you.

Presupposition is the content that survives a no

Ask What stopped you from upgrading? and the participant can answer Nothing stopped me -- but that answer still concedes the frame that upgrading was a thing they considered and something intervened. The premise is not in the answer set. It is underneath it.

Run the negation test on three questions a product team would actually ship:

Each of those survives an explicit denial. That is the signature. Compare the repaired versions, which put the premise into the answer set where the participant can actually reject it:

• Did anything stop you from upgrading, and if so, what?
• How do you find the export flow, if you use it?
• Are you using the mobile app currently?

Notice that the repairs are longer and clumsier. That is the trade, and it is worth making, because the cost of the smooth version is a number you will later report as a finding.

The evidence: a question can move an answer three-fold

The cleanest demonstration is not from an eyewitness lab. It is from a study whose participants believed they were doing ordinary product research.

Elizabeth Loftus reports an unpublished study of 40 people interviewed about their headaches and about headache products, in her own words, "under the belief that they were participating in market research on these products." Two questions carried the experiment.

The first varied only the example numbers offered alongside the question: In terms of the total number of products, how many other products have you tried? 1? 2? 3? against the same question ending 1? 5? 10? The low-anchor group reported trying an average of 3.3 other products. The high-anchor group reported 5.2 -- a 58 percent increase, t(38) = 3.14, p < .01, from three digits that were never the question.

The second varied a single adverb: Do you get headaches frequently, and, if so, how often? against Do you get headaches occasionally, and, if so, how often? The frequently group reported an average of 2.2 headaches per week. The occasionally group reported 0.7 -- t(38) = 3.19, p < .01. One adverb moved a self-reported behavioural frequency by a factor of more than three.

Neither question is leading in the usual sense. Neither contains a loaded adjective, a recommendation, or an opinion. Both presuppose a scale, and the participant answered on the scale they were handed.

The pattern generalises beyond self-report. Loftus cites Harris (1973), whose participants were told the experiment concerned the accuracy of guessing measurements and were asked either How tall was the basketball player? or How short was the basketball player? The estimates averaged about 79 and 69 inches. Asked How long was the movie? they averaged 130 minutes; asked How short was the movie?, 100 minutes. Thirty minutes of runtime, created by one adjective, in a task where participants had been explicitly instructed to guess as accurately as they could.

The five presuppositions that get into research questions

Once you have the negation test, the categories are easy to spot.

Existential, carried by the definite article. The asserts that the thing exists and that you both know it. Loftus puts it precisely: an investigator who asks, "Did you see the broken headlight?" essentially says, "There was a broken headlight. Did you happen to see it?" In product research this arrives as the onboarding friction, the pricing concern, the workaround -- all of which you named, not the participant.

Factive, carried by verbs like know, realise, notice. Why do you think teams struggle with handoffs? presupposes that teams struggle with handoffs. The participant can dispute it, but disputing costs them a turn and a bit of social friction, and most will not.

Change-of-state, carried by stop, start, keep, still. Are you still exporting to spreadsheets? presupposes you were. Every answer, yes or no, confirms the history.

Scalar, carried by a marked adjective. This is the Harris effect. How difficult was setup? presupposes non-zero difficulty and sets the zero point at the difficult end. How was setup? does not.

Enumerative, carried by your answer options. A single_choice or multiple_choice list presupposes that the real answer is in the list. A scale presupposes the dimension is the right one and that the endpoints are where you put them. This one is structural rather than verbal, and it is the presupposition most often shipped without anyone noticing, because the wording review looks at the question stem and not the options.

Where this stops and the memory literature starts

Presupposition has a downstream cousin. If a question plants a premise, the premise can persist into what the participant later remembers, which is a different and better-documented failure.

Loftus demonstrated it in the same 1975 paper across four experiments and 490 participants. Participants asked how fast a car was going when it ran the stop sign later reported seeing a stop sign 53 percent of the time, against 35 percent for those asked about the car turning right. Asked a week later whether they saw a barn, 17.3 percent of those whose earlier question mentioned passing a barn said yes, against 2.7 percent of the rest -- a barn that was never in the film.

That contamination-of-memory half is covered in depth in the guide to the misinformation effect in user interviews, which is where to go for sequencing rules and first-mention audits. This article is about the narrower and more immediate problem: not what the participant will remember next week, but what they are permitted to say to you right now.

The distinction from leading questions matters too. Avoiding leading questions covers the rewrite framework for questions that signal a preferred answer. Presupposition is what is left after you apply that framework and the question still constrains the response.

A pre-launch presupposition audit

Run this on every question before fielding. It takes about ten minutes for a fifteen-question guide.

1. Negate each question. Write the denial a participant would have to produce. If a fact is still standing, flag it.
2. Circle every definite article. For each the, ask who established that the thing exists. If the answer is we did, in the kickoff, change it to a or cut it.
3. Circle every marked adjective and adverb. Difficult, slow, frustrating, frequently, easy -- each one sets a zero point. Replace with the unmarked form or an open stem.
4. Read your answer options as claims. For each single_choice and multiple_choice question, ask whether a reasonable participant could have an answer outside the list. If yes, either add it or pair the question with an open_ended follow-up.
5. Check your scale endpoints. A scale question running from somewhat useful to extremely useful presupposes usefulness. Anchor at a true zero.
6. Ask who benefits from the premise. If accepting the premise makes your roadmap look correct, you have found the one worth rewriting first.

If you are running the study in Koji, the natural place for this audit is the study brief, before the questions are locked and the interview link goes out -- the premises are easiest to see while the research goal and the question set are still on the same screen.

The sixth check is the one that earns its keep. Presuppositions do not enter guides at random. They enter because the team already believes something and phrased the question from inside that belief -- which means the flagged premises correlate almost perfectly with the assumptions you most need tested. For a structured way to turn those into questions that can actually come back negative, see falsifiable research questions.

How Koji handles this

Three properties of the platform bear directly on presupposition.

The AI interviewer asks the question you wrote. A human moderator under time pressure paraphrases, and paraphrase is where premises get added -- the moderator who has heard eleven people complain about pricing starts asking the twelfth what about the pricing put you off. Koji's AI moderator delivers the written stem as written to every participant, so an audited question stays audited for the whole fieldwork period. That consistency is worth quantifying on its own, which is the subject of interviewer variance and moderator drift.

Structured questions make the enumerative presupposition visible. Koji supports six question types -- open_ended, scale, single_choice, multiple_choice, ranking, and yes_no -- and the type you pick is itself a claim about the answer space. A ranking question presupposes the items are comparable on one dimension. A yes_no presupposes the question has two answers. Seeing the type declared next to the stem, as the structured questions guide lays out, makes the assumption reviewable instead of implicit.

Open-ended follow-ups do not inherit the premise. Koji's AI probes each answer on its own terms rather than running a fixed script of follow-ups, so a participant who rejects your frame gets pursued down their own path rather than dragged back to yours. Pairing a closed question with an open_ended probe is the cheapest insurance available against an answer list that was wrong. The mechanics are in probing questions in user interviews.

None of this makes a presupposition impossible -- a premise written into the stem will be delivered faithfully to all 500 participants, which is precisely the risk. Consistency is a multiplier on question quality in both directions. That is why the audit above happens before launch and not after.

Frequently asked questions

What is a presupposition in an interview question?

It is a fact the question treats as already established, which the participant must accept to answer at all. The test is negation: if the fact is still standing after the participant denies the question, it is a presupposition. What stopped you from upgrading? presupposes that something stopped you, and answering nothing still concedes the frame.

How is a presupposition different from a leading question?

A leading question signals which answer you want, usually through loaded wording, and it is fixed by neutral rewriting. A presupposition embeds a premise underneath the question, survives neutral rewriting, and survives the participant saying no. What stopped you from upgrading? has no loaded words and is still presupposing.

How much can a presupposition actually change an answer?

In a study of 40 people who believed they were doing market research on headache products, asking Do you get headaches frequently produced an average of 2.2 headaches per week while occasionally produced 0.7 -- more than a three-fold difference from one adverb. Changing the example numbers offered with a question moved reported product trials from 3.3 to 5.2.

Do answer options count as presuppositions?

Yes, and they are the most commonly missed kind. A single_choice list presupposes the real answer is in the list, and a scale presupposes both the dimension and where the zero sits. Pair closed questions with an open_ended follow-up so participants outside your answer space have somewhere to go.

Does an AI moderator reduce presupposition problems?

It removes the drift kind, not the design kind. Koji's AI interviewer asks the written question as written, so premises do not creep in through moderator paraphrase mid-fieldwork. A premise already in your stem is delivered consistently to everyone, which is why the negation audit belongs in your pre-launch checklist.

What is the fastest way to audit a guide for presuppositions?

Negate every question and write down what still stands. Then circle every the, every marked adjective such as difficult or frequently, and every change-of-state verb such as stop or still. Finally, read your answer options as claims and ask whether a reasonable participant could fall outside them.

Related Resources

• Avoiding leading questions -- the rewrite framework for questions that signal a preferred answer, and the layer this article sits underneath.
• Structured questions guide -- how the six question types work and why the type you choose is itself a claim about the answer space.
• The misinformation effect in user interviews -- what happens when a planted premise persists into what participants remember.
• Survey question wording guide -- phrasing, order, and context effects across a full questionnaire.
• Probing questions in user interviews -- how to follow an answer without importing a new premise into the follow-up.
• Falsifiable research questions -- turning flagged assumptions into questions that can come back negative.