Up to 40% of online research survey responses are fraudulent. Even some major survey platforms have rates of suspected AI agent responses from 6% to 41%, one study found. Much of it comes from humans lying in order to qualify for a survey or interview, but new digital technologies have made it easier for AI research fraud and bot farms to create completely synthetic responses. Not only is research fraud serious, but it’s also a prevalent problem in the market research industry that requires deep expertise to mitigate.
What is research fraud, and how does it work?
When a research participant doesn’t tell the truth, misrepresents themselves, or unleashes bots (often out of a data center), or has AI agents programmed to fill out surveys on their own, someone is committing research fraud. Participants who turn to AI or a search engine for help finding or editing an answer also compromise the data, even if they don’t see it as cheating.
The most common form of research fraud is when humans lie about who they are in order to qualify for a survey, interview, focus group, or any other kind of primary research. This could mean saying they are a manager instead of a line worker or work in tech instead of retail. Pretending to be in a different location, such as being in the US when they are in another country, is a common tactic for fraudsters.
People can also be honest in their background but skip through questions without answering them diligently or honestly. They may also use AI to generate a response, especially for questions in which participants have to write sentences instead of just picking from a list of pre-populated options. One study found that about a third of online participants use AI at least sometimes on open-ended questions.
Bot farms deploy multiple bots that follow programmed scripts to scrape and take online surveys, while AI agents read and reason through a survey the way a person would. These are especially dangerous because they give completely made-up answers and can swarm a single survey with hundreds or thousands of responses. Some of what looks like a bot farm turns out to be rooms of real people taking surveys at scale.
Does it matter if I have fraudulent data?
Without taking steps to prevent and remove fraudulent data, data contamination occurs. This means that the results from your data can’t be trusted to make decisions. Fraudulent data introduces bias, which makes it unreliable. In one sexual risk survey, removing the 11% of responses that were fraudulent erased the statistical significance of the findings. People can make very expensive or important decisions based on data that isn’t real. Taking steps to guard against research fraud and AI research fraud raises the cost of conducting research and extends timelines, but it’s worth it to ensure real data that reflects what humans actually believe, want, and need.
RGC understands this challenge firsthand. For one large survey, over 2,000 fraudulent responses were filtered out throughout the research process. This meant that the fraud protections built into the study were working, but it also meant that a lot of time and resources needed to go to countering this problem. Even so, the data was delivered on time.
Ways to Prevent Research Fraud
Research should utilize a variety of fraud prevention measures that cover human error, human misrepresentation, human reliance on AI input, bots, and AI agents, as using only one or two measures can create a false sense of security in results. At the same time, it’s key to avoid flagging and removing non-fraudulent responses just because they look a bit different or don’t confirm pre-existing assumptions. Below are methods drawn from recent research, grouped by the stage of research where they apply.
| Mechanism | What it does | What it can’t do | Effectiveness | RGC Example |
|---|---|---|---|---|
| Recruitment and Screening | ||||
| Sampling | Probability sampling and limited mass public outreach decrease the chance of bad actors, especially bot farms and AI agents, finding the research call | Non-probabilistic sampling measures and large social media reach increase chances of fraud, with some studies reporting fraud rates up to 96% of responses | High | Targeted recruitment & probability sampling |
| Warning | Lay out clear instructions for what is and isn’t allowed for informed consent | Force bad actors to be honest | Medium | Clear and detailed terms of service that ties fraud or any kind of misrepresentation to disqualification and loss of incentive |
| Screeners | Targets specific groups of people that qualify based on a set of questions | It can’t prevent people from lying on the screener | High | Screeners are targeted for each research project without leading information that could signal the topic of the research to limit fraudulent responses to qualify |
| Knowledge and Verification Questions | Questions requiring industry, location, or some other particular knowledge provide a way to check for a real and appropriately targeted respondent | AI agents may still answer questions correctly | Medium | When appropriate for a specific survey or screener, questions are added that produce verifiable data |
| CAPTCHAs | Prevent non-AI bots from accessing online surveys | Reliably block AI agents, though newer versions like reCAPTCHA v3 catch some | Low | Standard with research platform |
| IP Checks | Match IPs to geolocation and identical IPs for multiple entries | Can’t catch VPNs or determine the difference between a bot farm and an apartment building | Medium | Use a third-party provider to check geolocation and batched data-center IPs |
| Survey Design | ||||
| Quality Design | Clear instructions and well-constructed questions reduce the reasons humans rush or turn to AI for help | Only researchers with design expertise can create high-quality instruments | High | Principal has a PhD and over a decade of research experience |
| Commitment Question | Asks participants to commit to data integrity, which discourages honest people from lying or using AI to answer | Doesn’t apply to bots, AI agents, or people set on committing fraud | Medium | In screeners and surveys, asks willingness to uphold data integrity and honesty and filters out ‘No’ responses |
| Attention Checks | Filter out humans who are not paying attention to a particular question | Filter out bots, AI agents, or organized human fraud | Low | Typically one attention check is included in each survey |
| Multiple Short Answer Questions | Gives a chance to check for human-written content | Protect against people lying or reliably catch AI-written answers, which can read better than human ones | Low | Use qualitative short-answer questions within screeners and surveys |
| Activities | Photo and number matching to trick bots and AI agents | Research shows AI agents can identify images like brand logos without difficulty | Low | State which photo describes a particular noun or verb |
| Inconsistent Responses | Asks for the same fact in different ways across screeners, surveys, and interviews, such as age and date of birth, which should produce matching answers | Catch AI agents given a persona, which answer consistently, or declare fraud definitively for all answers; needs expertise to understand the difference between this and a genuine finding | Medium | Includes consistency checks in research instruments |
| Interviews and Focus Groups | ||||
| Camera Requirement | Requires cameras on during interviews and focus groups so that human presence and behaviors can be confirmed | Respondents may still have lied in the screener or with answers in the interview, and AI avatars can now appear in video calls | High | Cameras are always a requirement in these scenarios |
| Interview Behavior | Interviewer discretion to examine for suspicious behavior, such as long pauses, typing, answers different from screener | Subjectivity in analysis of some behavior, especially for certain populations like people with disabilities | High | Every interview has a structure to check for suspicious behavior and an ability to stop the interview when issues arise |
| Data Review and Payment | ||||
| Response Times | Filters out responses with outlier completion times or batches submitted within a minute of each other | Catch AI agents, which often take longer than humans, or bots programmed with time delays per question | Low | Use in-platform response time data to filter out speedy respondents |
| Third-party Platforms | Research and payment platforms add layers of security and fraud detection measures, such as duplicate IP checks | The capabilities of each platform will differ | Variable | Survey platforms flag suspicious answers, patterns, response time, and other factors. Payment platform has strong security for a final check against bot farms, data centers, and other indicators to prevent payouts to fraudulent respondents and filter them out. |
| AI Literacy | Train researchers to understand how prompts, hallucinations, and outputs work to catch AI content | It can be subjective and difficult to tell the difference between sole AI output and AI-edited output | Medium | RGC’s principal has AI certifications to enable ethical use and deep understanding to catch AI-generated fraud |
| Payment Policies | Delayed, non-automated payment allows time for fraud checks, and harassment about payment flags respondents on a behavior basis, especially those with multiple accounts as they send multiple emails with identical text | It can’t separate those genuinely with payment concerns from others | Medium | RGC delays payment but always pays within one to two weeks. This allows for proper in-house and third party full fraud detection filtering. |
Overall, it’s important to consider the vulnerabilities of each research project, as variation in the effectiveness of a specific tool can be context-specific. An open-link survey with an incentive needs different protections than interviews recruited from a verified customer list. No single tool can catch everything, so the tools work best in combination, with a response flagged when several signals line up. This also protects real respondents from being removed just because one answer looks different. Researchers also have to stay up-to-date on the evolution of technological capabilities like AI so that the tools employed can stay at the forefront of combating bad data. Open-ended questions and basic CAPTCHAs were once reliable guards against bots, but AI agents can now get past them.
If you are commissioning research, ask your research partner how they prevent and catch fraud at each stage of the project and how they keep real respondents from being wrongly removed. A good answer should be specific to your study.
Works Referenced
Ali, S., Esposito Barbosa, W., Fassl, M., Ganapathi, A., Mink, J., & Aviv, A. J. (2026). SoK: Mapping threats to defenses in online survey fraud. In Proceedings of the Twenty-Second Symposium on Usable Privacy and Security (SOUPS 2026). USENIX Association. https://www.usenix.org/conference/soups2026/presentation/ali
Bonnamy, J., Carr, B., Lazarus, M. D., & Connell, C. (2025). Survey sabotage: Insights into reducing the risk of fraudulent responses in online surveys. Anatomical Sciences Education, 18(8), 767–773. https://doi.org/10.1002/ase.70015
Cho, V., & Dueñas, S. (2025). Surprises from the shadows: Confronting intrusions from bots and generative AI in survey research. Journal of Educational Research & Practice, 15, 1–21. https://doi.org/10.5590/JERAP.2025.15.2091
Davies, M. R., Monssen, D., Sharpe, H., Allen, K. L., Simms, B., Goldsmith, K. A., Byford, S., Lawrence, V., & Schmidt, U. (2024). Management of fraudulent participants in online research: Practical recommendations from a randomized controlled feasibility trial. International Journal of Eating Disorders, 57(6), 1311–1321. https://doi.org/10.1002/eat.24085
Gerber, S., & Cash, S. B. (2026). Identifying generative AI use among genuine responders in online survey research. Communications Psychology, 4, Article 111. https://doi.org/10.1038/s44271-026-00506-8
Jaffe, S. N., Moss, A. J., Hartman, R., Rosenzweig, C., Gautam, R., Robinson, J., & Litman, L. (2026). The bots ruining social science are not bots at all. Perspectives on Psychological Science, 21(2), 127–137. https://doi.org/10.1177/17456916251404872
Kaplan, G., Garg, S., Wan, M. W., & Smith, D. M. (2025). Scammer participants and AI-assisted interviews in qualitative health research: An example from a study of expectant parents with a rare genetic condition. International Journal of Qualitative Methods, 24, 1–13. https://doi.org/10.1177/16094069251390983
Lebrun, B., Temtsin, S., Vonasch, A., & Bartneck, C. (2024). Detecting the corruption of online questionnaires by artificial intelligence. Frontiers in Robotics and AI, 10, Article 1277635. https://doi.org/10.3389/frobt.2023.1277635
Martherus, J., Podkul, A., Cook, E., & Liebowitz, R. (2025). How to detect AI-assisted interviews in online surveys. Survey Practice, 18. https://doi.org/10.29115/SP-2025-0016
NORC at the University of Chicago. ([Year]). [Article title]. https://www.norc.org/research/library/fraud-problem-reshaping-survey-research.html
Pinzón, N., Koundinya, V., Galt, R. E., Dowling, W. O'R., Baukloh, M., Taku-Forchu, N. C., Schohr, T., Roche, L. M., Ikendi, S., Cooper, M., Parker, L. E., & Pathak, T. B. (2024). AI-powered fraud and the erosion of online survey integrity: An analysis of 31 fraud detection strategies. Frontiers in Research Metrics and Analytics, 9, Article 1432774. https://doi.org/10.3389/frma.2024.1432774
Singer Pressman, M. (2026). When artificial intelligence enters the research process: Implications for data integrity in human participants research. Journal of Empirical Research on Human Research Ethics, 21(4), 377–383. https://doi.org/10.1177/15562646261453181
Zhang, G., Walatka, R., Chen, S. Y., Moore, A. K., Urminsky, O., Fernandez, K., Low, A., Bogard, J. E., & Fox, C. R. (2026). Estimating the threat of AI-agent responding across online survey platforms. Proceedings of the National Academy of Sciences, 123(42), e2606460123. https://doi.org/10.1073/pnas.2606460123
Zhang, S., Xu, J., & Alvero, A. J. (2025). Generative AI meets open-ended survey responses: Research participant use of AI and homogenization. Sociological Methods & Research, 54(3), 1197–1242. https://doi.org/10.1177/00491241251327130