Fake Data Is a Real Problem in Research. Here Is How the Industry Is Fighting It

If you have ever run an online survey and had the nagging feeling that some of the responses did not quite feel human, you were probably right.

Research fraud is one of the least discussed quality problems in the industry, partly because it is embarrassing to acknowledge and partly because it is difficult to quantify. But the Greenbook GRIT 2024 Report found that data quality concerns are now among the top three challenges facing research buyers globally. And with AI making it trivially easy to generate plausible-looking survey responses at scale, the problem is not getting smaller.

What Research Fraud Actually Looks Like

Not all fraudulent data comes from sophisticated operations. Most of it falls into four categories:

• Speeders: respondents who complete a 20-minute survey in four minutes, clicking through without reading questions. The data looks complete but is essentially random.
• Straightliners: respondents who select the same response option for every matrix question. Quick to complete, statistically destructive.
• Duplicate respondents: the same person completing the survey multiple times under different identities to collect multiple incentive payments.
• AI-generated responses: increasingly common since 2023, where large language models are used to generate survey responses that are linguistically plausible but not grounded in genuine human experience or behavior.

The most dangerous research fraud is not the kind that looks obviously wrong. It is the kind that looks exactly right.

How Researchers Are Detecting It

The research industry has developed a layered approach to fraud detection, and the most rigorous studies apply multiple checks simultaneously:

• Response time monitoring: flagging completions that fall below a statistically implausible minimum based on question count and complexity.
• Open-text analysis: reviewing verbatim responses for generic, unusually fluent, or off-topic text that may indicate AI generation.
• Attention checks: embedding questions that have a clearly correct answer (or explicitly instruct respondents to select a specific option) to verify that respondents are reading.
• IP and device fingerprinting: detecting duplicate submissions from the same device or IP address range.
• GPS and metadata validation: for mobile surveys, verifying that the respondent's physical location matches the claimed location.
• Consistency checks: asking the same substantive question in two different formats and flagging respondents whose answers are logically incompatible.

What Research Firms Should Be Doing

Fraud detection should be built into the project design, not added as an afterthought during analysis. That means:

• Setting minimum response time thresholds before data collection begins and excluding completions that fall below them.
• Using panel providers that have implemented device fingerprinting and duplicate detection on their end.
• Building open-text questions into surveys where AI-generated responses are a concern, and reviewing them before finalizing the data.
• Reporting excluded responses in the methodology section of the final report, with the exclusion criteria documented.

The Verification Dimension

For face-to-face and phone research, fraud takes different forms but the same principle applies: respondents or interviewers fabricating data rather than collecting it. This is why interviewer monitoring, back-checks (re-contacting a sample of respondents to verify the interview took place), and GPS-logged field submissions are standard quality control practice in professional fieldwork operations. Platforms like ProjectBist support this through verified researcher profiles where credentials, past project experience, and client ratings are documented, reducing the risk of engaging field teams without a traceable quality track record.

Sources: Greenbook GRIT Report 2024; Qualtrics Data Quality Best Practices 2025; ESOMAR Quality in Research Guidelines