Online research lives in a bad neighborhood. Its front door is on the same street as the “get paid to take surveys” sites, the text messages promising easy money for liking videos, and the quizzes that exist to harvest your email address. Honest panels live on that street too. From the sidewalk, they look like everyone else.
The neighborhood has shaped the residents. When a survey pays a dollar or two to whoever turns up, the people who turn up most often are the ones optimizing for dollars, and lately some of them aren’t people at all. Online survey fraud is what you get when you build a market that way.
Moving up means changing what we offer and whom we let in: respect for people’s time, surveys short and relevant enough to deserve an answer, something real in return, and participants we can actually verify. None of it is new technology, and all of it costs money.
What does the bad neighborhood look like?
Look at a survey invitation through the eyes of the person receiving it. An unexpected message offers money for a few minutes of clicking. According to the Federal Trade Commission, that is also how task scams begin: with an unexpected text or WhatsApp message offering online work but no specifics. About 20,000 Americans reported task scams in the first half of 2024, against about 5,000 in all of 2023, and reported losses to job scams reached $223 million in those six months.
The “get paid to take surveys” economy sits somewhere in between. Some of it is genuine panels paying small amounts. Some of it is affiliate marketing that signs you up for offers you didn’t want. The FTC’s advice for the worst end fits on a sticky note: never pay anyone to get paid.
The research industry chose this address. Opt-in samples are recruited, in Pew Research Center’s words, from online sources “like banner ads or social media.” In my experience a respondent can pass through several companies’ hands before reaching your questionnaire. Most of that chain is paid for completes that count, and very little of it for people who are real.
Who claims to be licensed to operate a nuclear submarine?
In February 2022, Pew asked an online opt-in sample a question with a known answer: are you licensed to operate a class SSGN nuclear submarine? The share of Americans with that qualification rounds to 0%. Among opt-in respondents under 30, 12% said yes. Among opt-in respondents who said they were Hispanic, 24% said yes, against 2% of non-Hispanics.
Pew wasn’t hunting for secret submariners. It was explaining why an opt-in poll published in December 2023 had found that 20% of Americans under 30 agreed that “the Holocaust is a myth.” When Pew asked the same question in January 2024 on its American Trends Panel, which recruits people by random sampling, 3% of adults under 30 agreed, the same share as in every other age group. Young Americans weren’t unusually gullible; the opt-in sample was carrying what Pew calls bogus respondents, people who “attempt to complete surveys with as little effort as possible to earn money or other rewards.”
Bogus respondents aren’t random noise, and that is what makes them dangerous. They lean toward “yes,” which inflates every estimate of an unusual belief or behavior. Pew suggests they may overreport to get past screening questions, and its later work found that insincere respondents tend to claim to be young or Hispanic, “likely falsely in many cases.” The damage concentrates in the subgroups clients most want to read about.
The standard defenses don’t help much. In an earlier Pew study, bogus respondents made up about 4% to 7% of opt-in samples, against 1% in panels recruited from random address lists, and 84% of them passed a standard attention-check question. Our samples, it turns out, are full of submariners in space. So when a headline says one in five young Americans believe something outlandish, the first act of critical thinking is to ask how the sample was recruited and how many of them drive submarines.
One in ten respondents has been to the space station
Can’t you just screen the frauds out? Pew tested exactly that in an opt-in survey of 11,114 adults fielded in November 2024, and published the answer in August 2026 under a title that saves you the reading: No Easy Fix for Bogus Respondents in Online Opt-In Polls.
The trap questions alone are worth the price of admission: one in ten respondents said they had visited the International Space Station. Asked about Polar-class icebreakers, a class of which only two ships were ever built, 8% said they had served on one. Some 6% used a social network called Fizzypress, which doesn’t exist, and 9% had received payments from the United States Railroad Administration, which ceased to exist in 1920. In all, 18% of the sample failed at least one trap.
The fixes disappointed. Trap questions and a commercial prescreening service improved data quality somewhat. Matching respondents to a voter file made the error slightly worse, because it threw out mostly honest people who simply declined to give their name and address.
Pew’s verdict is that purging bogus cases lowers error on most measures but that a surefire solution remains elusive. It also notes that polls recruited offline by random sampling are generally immune to this particular threat, though they have other problems.
Survey farms, professional respondents and “ants”
Some of the fraud is organized. In April 2025, federal prosecutors in New Hampshire unsealed an indictment against eight people connected to the panel companies Op4G and Slice MR, describing a conspiracy to bill $10 million in fraudulent market survey data. Prosecutors said senior leaders at Op4G decided in 2014 to boost revenue by fabricating survey data. According to the indictment, they recruited “ants” who posed as genuine respondents and coached them on screener questions, on how long to stay on a survey and on using VPNs to hide their real locations.
By October 2026, four defendants, including Op4G’s chief executive, had pleaded guilty to conspiracy to commit wire fraud; the others have not been convicted. The victims included hospitals and universities, and Retraction Watch reported in August 2026 that researchers at about half a dozen US institutions were reviewing or correcting studies. One journal had already retracted a paper.
The irony is hard to beat. Op4G stood for Opinions for Good, and its pitch was that members could send part of their survey pay to a nonprofit of their choice. A nobler currency didn’t help, because nobody was checking who was being paid.
Academics have their own version of the story. A 2020 audit of 38 studies on a crowdsourcing platform popular with social scientists found that fraud through virtual private servers, used to hide where respondents really were, was not new but had grown. Those respondents gave particularly poor data and could weaken experimental effects.
Professional respondents are a more ambiguous case. A 2025 study in Political Analysis tracked the web browsing of people in three online samples. In the one bought through a sample marketplace, survey sites made up 54.6% of participants’ web visits, and between a fifth and a third of them counted as professionals, depending on the definition.
The authors found that professionals did not, by and large, distort the conclusions. What worried them was repeat participation: in that same sample, 71.5% had attempted at least one questionnaire more than once. Someone who takes surveys for a living isn’t necessarily a fraud. A system that lets the same person in twice is broken.
Can AI bots pass survey attention checks?
Then there are respondents who were never people. In November 2025, Sean Westwood of Dartmouth published a study in PNAS describing an “autonomous synthetic respondent,” an AI agent driven by a prompt of about 500 words. According to the press release, across 43,000 tests it passed 99.8% of attention checks designed to catch bots, made no errors on logic puzzles, and evaded every AI detection method he tried. It costs about five cents per response, against the $1.50 or so a human typically earns.
The political demonstration should keep pollsters awake. In seven major national polls before the 2024 US election, adding between 10 and 52 fake responses would have flipped the predicted outcome. The paper went on to win the 2025 Cozzarelli Prize.
The fair objection is that the paper shows what an agent can do, not how many agents are already taking surveys, and I don’t know the answer either. But the economics point one way. When a fake complete costs five cents and pays more than a dollar, I wouldn’t bet a client’s segmentation on fraudsters being slow learners. Westwood’s own conclusion is that “the technology exists to verify real human participation; we just need the will to implement it.”
What would moving up take?
Pew’s 2026 report is the best evidence that filters only go so far. So moving up has to start earlier, with the deal we offer people and the way we let them in.
Start with time. A survey that is short and relevant to the person answering it gets better answers, and a 25-minute questionnaire sent to anyone who clicks selects for the people best at enduring it. I make that argument at length in You killed market research. Respecting time also means asking fewer people, less often, about things they actually know.
Then give something back. Cash works: the American Association for Public Opinion Research and the American Statistical Association say nothing works as well as cash for raising response rates, and I won’t argue with them. But a dollar or two, offered to anyone, mostly attracts the people for whom a dollar or two is the point.
What’s missing is everything else: telling people what the research found, what changed because of it, and treating a panel member as someone with a relationship to the researcher, not a row in a supplier’s inventory. A completed survey is a transaction. A person who cares enough to answer honestly is something else (participation is not engagement).
Why recruited and verified panels are worth the money
Recruit people instead of accepting volunteers. Probability-based panels pick members at random from a list of the whole population and don’t let anyone sign themselves up. Pew’s American Trends Panel samples households from the Postal Service’s list of residential addresses, and until 2024 it gave tablets and data plans to households without internet. The Dutch LISS panel draws its sample from the population register through Statistics Netherlands and gives a computer and internet access to households that would otherwise be left out.
Panels like these cost more per interview. They are also why Pew can describe its own panel as generally immune to bogus respondents.
Verify that participants are real. One identity per person, a history that follows them, and proof that they were invited rather than simply turned up. Westwood is right that this is a matter of will more than technology.
Build first-party relationships, and keep them honest. A company’s own customers, asked with consent, by a name they recognize, about something they actually experienced, are the best neighborhood online research can live in. The condition is that research stays research: the ICC/ESOMAR International Code says research depends on public confidence in its integrity and requires a clear distinction between research and non-research activities. The moment a “customer survey” turns into a sales follow-up, you’re back on the old street.
My take
Online research didn’t end up in the bad neighborhood by accident. It moved there because the rent was cheap, and it stayed because clients kept paying for completes without asking who completed them. The fraud rings, the bots and the submariners are what cheap rent buys.
Moving up costs money, and I won’t pretend otherwise. A verified respondent recruited at random costs more than a complete from an exchange. But the comparison is false, because the cheap complete isn’t the same product. One is an answer from a person you can find again; the other comes from someone who, with a probability nobody in the chain will tell you, has also been to the International Space Station.
My view is that the industry should stop treating sample as a commodity and start treating respondents as the only asset it really has. Ask fewer people fewer questions, pay them properly, tell them what you learned, check that they exist, and stop sending invitations that look like scams. The neighborhood improves when the neighbors do.
Questions people ask
Can you trust online opt-in polls?
Only with care. Pew Research Center has repeatedly found bogus respondents in opt-in samples who say yes to almost anything, which inflates estimates most for adults under 30 and Hispanic adults. Panels recruited by random sampling are far less exposed.
Can AI bots fake survey responses?
Yes. A 2025 study in PNAS built an AI agent that passed 99.8% of attention checks and evaded every detection method tested, at about five cents per response.
Are ‘get paid to take surveys’ sites legit?
Some are real panels that pay small amounts for your time. Walk away from anything that promises big earnings or asks you to pay, deposit money or buy a list before you start: never pay to get paid.
What is a probability-based panel?
A panel whose members are chosen at random from a list of the whole population, such as all residential addresses, instead of volunteering through ads. Nobody can sign themselves up, which keeps out most fraud.