How to Detect Fake Survey Responses: A 10-Signal Checklist (2026)
No single check catches fake responses. Ten signals that do work when you combine them, how to check each one by hand, and what to do with the responses you catch.
Fake responses come in two kinds
When people say "fake responses" they usually mean bots. Bots are real, and an open survey link posted on social media can collect hundreds of scripted submissions in an hour. But bots are the easier half. They tend to fail simple traps, and a basic spam filter catches most of them.
The harder half is human. A person who wants the gift card, or who just wants the survey to end, clicks through without reading. They pass every CAPTCHA because they are a person. Their answers look like data, sit in your spreadsheet like data, and quietly pull your averages toward the middle of every scale.
The checklist below covers both, roughly in the order you can check them.
The 10 signals
1. Completion time below reading speed
If a 20-question survey was finished in 40 seconds, nobody read it. A simple floor works: around 1.5 seconds per question is faster than anyone can read and answer. A relative check catches more: compare each response to the median time for that survey and look hard at anything under a third of it. Measure time on the server where you can, because a script can send whatever duration it likes.
2. Straight-lining
The same option position picked across a whole grid or a run of rating questions: all 4s, all "Agree", the first option every time. Some straight-lining is honest (a person who really is neutral on everything), so treat it as one signal and not a verdict.
3. Agreeing with everything
A softer cousin of straight-lining. The respondent varies their answers but always picks the most positive one. If someone strongly agrees with both "The onboarding was too long" and "The onboarding was the right length", they are not reading.
4. Exact and near duplicates
Identical answers to a recent submission usually mean the same person submitting twice, or a script replaying one payload. Near duplicates matter too: 90% of answers identical across a long survey, or a long open-text answer that shares most of its phrasing with someone else's, is a copy with small edits.
5. Gibberish and filler text
"asdf", "good", "nothing", "n/a" on a question that asked for a reason. Keyboard mashing is easy to catch with simple rules. Plausible filler ("It was fine, I liked it" on every open question) needs something that reads the answer against the question.
6. Answers that contradict each other
Age 19 and 25 years of work experience. An NPS of 10 followed by "terrible, I'm cancelling". A fraudster can fake a single answer convincingly, but keeping a whole made-up story straight across 30 questions is much harder.
7. Failed attention checks
An instructed item ("Select 'Disagree' for this question") has one correct answer. Missing it is strong evidence of not reading. Keep the checks fair: they should measure attention, not memory or trivia. Our attention check guide has examples that work and ones that backfire.
8. Fake contact details
Temporary email domains, phone numbers like 1234567890 or 9999999999, names that are keyboard rows. If your survey asks for contact details for a follow-up or an incentive, these are the cheapest fraud signal you have.
9. Answers that read as AI-written
Open-ended answers with essay structure, generic stock phrasing and no personal detail, sometimes with leftovers like "As an AI" or "[Your name]". Be careful here: articulate, emotional or non-native writing is not evidence of anything. Treat this as a light signal that matters most alongside others. More in how to spot ChatGPT-written answers.
10. Submissions that never loaded the form
A script can post straight to your submit endpoint without ever rendering the page. If your form tool issues a session token when the page loads, a submission without one did not come from a person filling in the form.
Scoring beats deleting
The mistake most teams make is treating each signal as a delete button. A real person can trip one check: a fast reader on a short form, a respondent who really does rate everything a 4. Delete on one signal and you throw away real people, and usually not at random.
A better approach is to give each signal a weight, add them up, and sort responses into three groups: clean, worth a look, and likely fake. Then you only spend manual review time on the middle group. Researchers in a 2024 PMC review of 31 fraud detection strategies made the same point: layered checks, with the reasoning written down, beat any single tool.
Doing this by hand in a spreadsheet
- Add a column for completion time and sort ascending. Mark anything under a third of the median.
- For each grid, add a column that checks whether every answer is the same value.
- Use a COUNTIF across the answer columns to find exact duplicates.
- Read the shortest open-text answers first. Gibberish clusters there.
- Add a score column that counts how many marks each row has, then review anything with two or more.
This works for a few hundred responses. Past that, or on a survey that runs continuously, it becomes a weekly chore that nobody does.
How SiliForm does it automatically
SiliForm runs these checks on every response in four layers. While a person answers, AI validation rejects answers that don't address the question, and a hidden spam trap and rate limit stop basic bots. At submit, it checks speed (on the server's clock), straight-lining, always-agreeable patterns, duplicates and near duplicates, fake contact details, gibberish, attention checks you marked, consistency rules you set, missing form sessions and bot detection. About 20 seconds later, an AI audit reads the whole response for filler, contradictions and AI-written answers.
Each response gets a score from 0 to 100 and a label: Clean, Review or Flagged, with the reason shown (for example "Flagged · Speeder +2"). One switch excludes flagged responses from your charts and exports. Nobody is ever blocked by a score, and you can override any label by hand. See AI answer validation for the first layer, or SiliForm for researchers for how it fits a study.
Frequently asked questions
How many survey responses are fake?
It depends heavily on how the survey is distributed. A widely cited 2015 analysis of 1,008 surveys estimated that about 17% of responses contained fabricated data, ranging from 5% in developed countries to 26% in developing ones. Open links shared on social media or paid panels with incentives tend to sit at the high end.
Should I delete every flagged response?
No. Look at the weight of evidence. One weak signal, such as a fast completion time on a short form, is not enough on its own. Remove responses that carry several signals, and keep a record of what you removed and why so your cleaning is reproducible.
Can a CAPTCHA stop fake survey responses?
A CAPTCHA stops many bots, but it cannot stop a real person who rushes through without reading, and paid solving services get past it. Most bad data in modern surveys comes from people who pass the CAPTCHA, so you still need checks on the answers themselves.
Does SiliForm check responses automatically?
Yes. On the Pro plan (free during the 14-day trial), every response gets a 0 to 100 quality score with the reasons attached, and lands in Clean, Review or Flagged. You can exclude flagged responses from analytics and exports with one switch, and relabel any response by hand.