{"site":{"name":"Koji","description":"AI-native customer research platform that helps teams conduct, analyze, and synthesize customer interviews at scale.","url":"https://www.koji.so","contentTypes":["blog","documentation"],"lastUpdated":"2026-07-16T11:16:28.301Z"},"content":[{"type":"documentation","id":"95a22727-f348-48a9-9e9d-10fdd475e700","slug":"attention-check-questions","title":"Attention Check Questions: How to Catch Low-Effort Survey Responses Without Annoying Real Participants","url":"https://www.koji.so/docs/attention-check-questions","summary":"Attention check questions are items planted in a survey to verify respondents are reading and processing questions rather than answering on autopilot, catching inattentive/low-effort respondents, straightliners, speeders, bots, and click-farm fraud. Types include the instructional manipulation check (IMC), instructed-response item, bogus/implausible item, logic/consistency check, known-answer item, reverse-coded item, and timing/honeypot metadata. Evidence: ~30% of panel respondents fail simple checks and an estimated 30-40% of online responses are fraudulent or unusable (CloudResearch); removing inattentive respondents increases statistical power (Oppenheimer et al. 2009). Pitfalls: place early, require multiple failures, avoid over-policing, and note checks are reactive (Clifford & Jerit 2015 found they can raise social-desirability bias) and increasingly beaten by AI agents. Conversational AI interviews (Koji) reduce the source of inattention and add a 1-5 quality score, but still need fraud safeguards.","content":"**Attention check questions are items deliberately planted in a survey to verify that respondents are actually reading and processing the questions rather than clicking through on autopilot.** They exist to catch inattentive, low-effort (\"satisficing\") respondents — along with straightliners, speeders, bots, and click-farm fraud that quietly poison your data. Used well, they raise data quality and statistical power. Used badly, they annoy your best participants, introduce new bias, and — increasingly — get beaten by AI agents anyway.\n\nFor anyone making decisions from survey data, attention checks are worth understanding precisely because they are a double-edged tool: the same trap question that removes a bot can also remove a thoughtful respondent who read the item too generously.\n\n## What an attention check is (and is not)\n\nAn attention check assesses whether an *already-qualified* respondent is engaging in good faith. That makes it different from a **screening or qualification question**, which decides whether someone belongs in your sample at all (\"Do you own a car?\"). A screener controls *who* is in; an attention check controls *how carefully* they are answering.\n\n## The main types of attention check\n\n- **Instructional manipulation check (IMC).** A normal-looking question whose text buries an instruction to ignore the obvious answer — for example, a \"which sports do you play?\" item whose paragraph quietly says to skip the checkboxes and click the title instead.\n- **Instructed-response item (IRI).** An explicit directive inside a scale: \"To show you are paying attention, please select 'Strongly Agree' for this item.\"\n- **Bogus or implausible item.** A statement with an objectively obvious answer, such as \"I have never used a computer\" — a real, attentive participant should disagree.\n- **Logic or consistency check.** Two items that should not both be endorsed, like pairing \"I always vote in elections\" with \"I never vote in elections.\"\n- **Known-answer / factual item.** A fact that needs no special knowledge: \"A dolphin is a type of: animal / mineral / vehicle.\"\n- **Reverse-coded item.** The same construct asked in flipped polarity (\"I feel calm\" vs. \"I feel anxious\") to expose straightlining.\n- **Timing and honeypot detection (metadata).** Flags for implausibly fast completion or hidden form fields only a bot would fill — a respondent finishing a 10-minute survey in 90 seconds is a red flag.\n\n## The evidence: how much bad data is out there\n\nThe scale of the problem is larger than most teams assume.\n\n- **Around 30% of panel respondents routinely fail simple attention checks**, according to data-quality analyses from CloudResearch — a reminder that inattention is the norm, not the exception, in commodity panels.\n- **An estimated 30–40% of online survey responses are fraudulent or unusable**, with the majority driven by human click farms rather than \"bots,\" per CloudResearch's data-quality research.\n- **Fraud is often absurdly detectable in aggregate.** In one CloudResearch analysis, among respondents who failed basic screening, more than half claimed to be petroleum engineers — a rare occupation — illustrating just how much low-quality traffic reaches unguarded surveys.\n- **Removing inattentive respondents increases statistical power.** That is the entire thesis of Oppenheimer, Meyvis, and Davidenko's 2009 paper that introduced the IMC: screening out participants who ignore instructions reduces noise and lets real effects show through.\n\nThere is also a warning on the horizon. CloudResearch notes that recent research shows AI agents \"can pass thousands of attention checks with near-perfect accuracy\" — so the traditional trap-question battery is losing the arms race even as it degrades the experience for genuine humans.\n\n## Best practices — and the pitfalls\n\nAttention checks help only when they are designed with restraint:\n\n- **Place them early, not late.** Prolific advises putting checks near the start so participants are not screened out after investing significant effort; late checks confuse genuine fatigue with inattention.\n- **Require multiple failures before excluding.** A single misread is not proof of bad faith. Prolific, for instance, requires failing at least two checks in longer studies before removing a respondent — a guard against false positives.\n- **Do not over-police.** Too many checks, or overly tricky ones, increase drop-off and annoy conscientious participants. Diligent people sometimes \"agree\" with an absurd item for a defensible reason, so aggressive traps can wrongly exclude valid respondents and even threaten scale validity.\n- **Remember that checks are reactive.** Clifford and Jerit (2015) found that prompting attentiveness can *increase* social-desirability bias, because respondents who feel monitored change how they answer later questions. The measurement tool alters the thing it measures.\n- **Keep instructed items easy to read.** No tiny fonts or memory-recall tasks — you want to measure attention, not eyesight or working memory.\n- **Pair content checks with metadata.** Combine trap questions with timing, straightlining detection, and fingerprinting rather than trusting any single indicator.\n\nAs one survey methodologist (Joe Hopper of Versta Research) puts it, if you find a huge number of inattentive respondents, \"you may need to think more about how you are designing surveys in the first place.\"\n\n## The deeper problem attention checks reveal\n\nThat quote points at the real issue. Attention checks are a patch, not a cure. Long grid-and-matrix batteries and repetitive Likert walls *manufacture* the very disengagement, straightlining, and satisficing that the traps then try to police. So teams end up bolting trap questions onto a format that is itself the root cause of low effort. Worse, the checks are reactive (they can bias later answers) and increasingly beatable — meaning the patch is losing effectiveness even as it worsens the experience for real participants. If your format produces inattention, the durable fix is to change the format, not to add more traps.\n\n## The modern approach: make the format engaging instead\n\nThis is where AI-native research changes the equation. Instead of policing inattention after the fact, platforms like [Koji](/docs/how-ai-interviewers-work) reduce the *source* of it.\n\n- **A conversational format is inherently more engaging.** A back-and-forth AI interview holds attention far better than a 40-question grid, so it produces much less straightlining and satisficing — shrinking the need for heavy trap-question batteries in the first place.\n- **Real-time detection instead of silent discard.** Because the AI moderates live, it can spot a low-effort, contradictory, or gibberish answer *as it happens* and ask a clarifying follow-up — recovering a distracted-but-genuine respondent rather than throwing the interview away afterward.\n- **Automated quality gating.** Koji scores each interview for quality on a 1–5 scale, so low-quality sessions can be down-weighted or excluded programmatically — playing the same statistical-power role Oppenheimer's IMC did, but continuously and per-respondent instead of via one pass/fail trap.\n- **Structure without the grid.** Koji's six [structured question types](/docs/structured-questions-guide) — open_ended, scale, single_choice, multiple_choice, ranking, and yes_no — preserve the comparable, analyzable data researchers need, so conversational depth does not cost you rigor.\n\n**The honest caveat: conversation is not immunity.** Incentivized human fraud and increasingly capable AI agents still require bot detection, behavioral signatures, and metadata checks. AI-moderated interviewing *reduces* your reliance on annoying attention-check batteries; it does not remove the need for validity control entirely. The goal is fewer traps and better data — not zero safeguards.\n\n## A quick decision guide\n\nRunning a one-off survey on a commodity panel? Use a small number of well-placed attention checks (early, easy to read, and require two failures to exclude), paired with timing and straightlining flags. Running ongoing discovery or feedback at scale? Move the work upstream: a conversational, [structured](/docs/structured-questions-guide) AI interview with an automated quality score will beat a longer trap-laden grid on both data quality and participant experience — while still layering in fraud detection where incentives are high.\n\n## Common mistakes to avoid\n\nEven teams that know they need attention checks tend to make the same errors:\n\n- **Using one clever trap and trusting it.** A single tricky IMC produces both false negatives (bots that pass) and false positives (thoughtful people who fail). Layer a couple of simple checks with timing and straightlining metadata instead of betting everything on one gotcha.\n- **Excluding on a single failure.** One misread is noise, not evidence of bad faith. Set your exclusion rule at two or more failures for anything beyond a very short survey.\n- **Placing checks at the very end.** Late checks conflate genuine fatigue with inattention and waste the effort of respondents who answered the first 90% carefully. Put them early.\n- **Making the trap a reading test.** Tiny fonts, dense paragraphs, or memory-recall items measure eyesight and working memory, not attention — and they punish careful readers.\n- **Ignoring what a high failure rate is telling you.** If a large share of respondents fail, the honest diagnosis is often that the survey is too long, too repetitive, or too grid-heavy. That is a design signal, not just a respondent problem — and it is the strongest argument for moving to a more engaging format.\n\nTreat attention checks as one layer in a data-quality system — content checks, behavioral metadata, and fraud detection together — rather than a single switch that separates good respondents from bad ones.\n\n## Related Resources\n\n- [Survey Data Quality: Detect and Prevent Bad Responses](/docs/survey-data-quality-guide)\n- [Survey Fraud & Respondent Quality](/docs/survey-fraud-respondent-quality)\n- [Survey Response Bias](/docs/survey-response-bias)\n- [Screener Questions Guide](/docs/screener-questions-guide)\n- [Survey Design Best Practices](/docs/survey-design-best-practices)\n- [Statistical Significance in Survey Research](/docs/statistical-significance-survey-research)\n- [Structured Questions Guide](/docs/structured-questions-guide)","category":"Research Methods","lastModified":"2026-07-15T03:23:07.016501+00:00","metaTitle":"Attention Check Questions: Types, Examples & Best Practices (2026)","metaDescription":"Attention check questions catch inattentive, low-effort, and fraudulent survey responses. Learn the main types, how many to use, the pitfalls, and why a conversational AI interview reduces the need for them.","keywords":["attention check questions","attention check survey","instructional manipulation check","trap questions survey","low-effort survey responses","survey data quality","attention check examples"],"aiSummary":"Attention check questions are items planted in a survey to verify respondents are reading and processing questions rather than answering on autopilot, catching inattentive/low-effort respondents, straightliners, speeders, bots, and click-farm fraud. Types include the instructional manipulation check (IMC), instructed-response item, bogus/implausible item, logic/consistency check, known-answer item, reverse-coded item, and timing/honeypot metadata. Evidence: ~30% of panel respondents fail simple checks and an estimated 30-40% of online responses are fraudulent or unusable (CloudResearch); removing inattentive respondents increases statistical power (Oppenheimer et al. 2009). Pitfalls: place early, require multiple failures, avoid over-policing, and note checks are reactive (Clifford & Jerit 2015 found they can raise social-desirability bias) and increasingly beaten by AI agents. Conversational AI interviews (Koji) reduce the source of inattention and add a 1-5 quality score, but still need fraud safeguards.","aiPrerequisites":["Basic familiarity with survey design","Understanding of online panels and respondent recruitment"],"aiLearningOutcomes":["Define attention checks and distinguish them from screening questions","Choose among the main types of attention check for a given study","Place and calibrate checks to avoid false positives and participant annoyance","Understand why a conversational AI format reduces the need for attention-check batteries"],"aiDifficulty":"intermediate","aiEstimatedTime":"9 min read"}],"pagination":{"total":1,"returned":1,"offset":0}}