Skip to main content

← All articles

Data quality · 3 min read

How to detect bots and AI agents in online surveys

Online surveys have always attracted people trying to get paid without doing the work. Now AI agents can fill in a survey too, writing fluent answers to open-ended questions. No single check catches everything, but several together catch most of it, and good design keeps honest participants from being caught by mistake.

Updated November 02, 2026

Signs of a bad response

Look for patterns, not single oddities. Each of these has innocent explanations on its own; together they tell a story.

  • Finishing far faster than the median completion time.
  • The same answer down a whole grid of questions (straight-lining).
  • Failing a clear attention check.
  • Open-ended answers that are generic, off-topic, or suspiciously polished and similar to other respondents' answers.
  • Pasted text in fields people normally type into.
  • Answers that contradict each other, such as an age that doesn't match a stated birth year.
  • Several accounts sharing one device or network.

Stop them at the door

The cheapest fraud to deal with is the fraud that never reaches your survey. Recruit from a panel that verifies people before they see a study: device and network checks, one verified payment account per person, and identity checks when risk or your study calls for them.

Check from inside the survey

In-survey signals catch what verification can't.

  • Timing on each page, and how often people leave the tab.
  • Paste and typing-speed counts on open-ended questions, without recording what anyone wrote.
  • A hidden instruction that only automated agents read, placed so screen readers skip it, then checking whether an answer follows it.
  • Open-ended questions that ask about the person's own experience, which are hard to answer generically.
  • Your survey tool's own bot detection, if it has one.

Be fair to real people

Fast readers, people using assistive technology and people on older devices can trip a single check. Don't reject on one signal. Use attention checks that are clear and fair, combine several signals, give a reason for every rejection, and let participants appeal. Pre-register your exclusion rules so you aren't deciding after seeing the results.

Report what you excluded

Reviewers increasingly ask how a sample was checked. Report your exclusion rules, how many responses each rule removed, and how participants were recruited and verified. A signed provenance report makes this a citation instead of a paragraph.

Populiva verifies participants before they see a study and shows these signals beside every submission.

How Populiva checks data quality →