Attention Check Questions: Examples That Work and Ones That Backfire

What makes a check fair

CloudResearch, which runs a large research participant platform, lists a few tests for a good attention check. Paraphrased:

• It measures attention, not memory or general knowledge.
• It has exactly one correct answer.
• It isn't so easy that a skimmer passes, or so strict that a careful person fails.
• It doesn't ask respondents to do anything strange.

Every bad check below breaks one of these.

12 attention checks that work

Instructed items

1. "To show you are reading, please select 'Disagree'." (as a row inside a rating grid)
2. "For this question, choose the third option."
3. "Please leave this rating at 2 stars."
4. "Select 'Blue' from the list below." (with a list of colours)

Self-report items

5. "I am paying attention to this survey." Correct: Agree.
6. "I have answered every question in this survey without reading it." Correct: Disagree.
7. "I was born on February 30th." Correct: No.
8. "I have never used a computer or phone." Correct: No (they're filling in an online survey).

Consistency pairs

9. Ask age early and birth year late. They should agree within one year.
10. Ask "How many employees work at your company?" and later "Are you a solo founder?"
11. Ask whether they used a product, then ask about features of it. A "never used" plus detailed ratings is a mismatch.
12. Ask satisfaction on a scale, then "Would you recommend us?". A 1 and a 10 together deserves a look.

Checks that backfire

Morning Consult tested several kinds of checks on real respondents. Simple maths or word-association checks were failed by 3 to 6% of people. Trick questions about made-up products were failed by 11 to 49%. One directed-choice item with confusing multi-step instructions was failed by more than half. Those aren't all inattentive people. Many are people the question tricked.

• Memory checks. "What colour was the logo on the first page?" This tests memory, not attention.
• Trivia. "What is the capital of Australia?" Plenty of attentive people don't know it's Canberra.
• Fake brands. "Have you used Zentrova in the past month?" People guess, or confuse it with a real brand.
• Buried instructions. A long paragraph with "ignore the question below and type 'purple'" hidden in sentence four. You'll catch skimmers, but also people who read every word of the question and answered it.
• Insulting checks. "Are you a robot?" Fine once. Three times and honest respondents start to feel accused.

How many, and where

Use two for a survey of more than about ten questions, and space them out: one in the first third, one in the last third. Fatigue sets in late, so the second check is often the more useful one. Tell respondents up front that the survey includes a few checks. It nudges people to read, and it's fairer.

Attention checks are a human filter

Attention checks were designed to catch inattentive people, and they still do that well. They were never strong against automated fraud, and they're getting weaker: Westwood (2025) showed AI agents passing embedded checks. Pair them with checks that don't depend on the respondent cooperating, like timing, duplicates and bot detection.

Attention checks in SiliForm

In SiliForm you can mark any single-choice, dropdown or yes/no question as an attention check and pick its correct answer. On the Pro plan, a response that picks a different answer gets a "Failed check" flag. Unanswered checks are skipped, and a check you edited after the response came in is ignored rather than scored against the new answer.

For consistency pairs, set rules like "if Q3 is 'Solo founder', Q7 can't be '50+ employees'" under Settings, and a broken rule adds a "Rule broken" flag. The AI audit also looks for contradictions you didn't write rules for. All of these feed one quality score per response; none of them block the respondent.

To draft the rest of the survey, start with the AI survey generator; SiliForm for researchers covers the other quality checks.

Frequently asked questions