A new research paper is drawing attention to a problem that could quietly reshape the credibility of online science: how can researchers tell whether a person who completes a survey is genuinely responding, or whether generative artificial intelligence is answering on their behalf? Published in Communications Psychology, the study by S. Gerber and S. B. Cash focuses on “Identifying generative AI use among genuine responders in online survey research,” placing a rapidly evolving technology at the center of a debate over what counts as human data in the digital age.
Online surveys have become one of the most widely used tools in psychology, public health, economics, marketing and social science. Researchers can recruit thousands of participants through web-based platforms, asking them about emotions, political attitudes, health behaviors, consumer preferences or personal experiences. The speed and scale are powerful, but the method depends on a basic assumption: that each response reflects the participant’s own thoughts and experiences. Generative AI complicates that assumption by allowing respondents to produce fluent, detailed answers in seconds, often without leaving obvious signs of assistance.
The issue is especially important for open-ended questions, in which participants write explanations rather than selecting from fixed options. A language model can generate grammatically polished text, vary its tone, summarize a hypothetical experience or imitate the style of an ordinary respondent. That means a response may appear thoughtful and coherent while containing little evidence that the participant personally composed it. The result is not necessarily a conventional bot or a fraudulent account. It may be a real person using an AI system to complete part of a questionnaire, creating a category that is more difficult to define and detect.
This distinction is at the heart of the research topic. Traditional survey-quality checks often look for impossibly fast completion times, repeated answer patterns, contradictory selections, failed attention checks or suspiciously identical wording across participants. Those tools can identify careless, automated or fabricated responses, but generative AI introduces a more subtle signal. A participant may read every question, understand the topic and submit answers that are individually plausible. The concern is not simply whether the respondent is real, but whether the recorded answer represents the respondent’s own cognition, memory and judgment.
Technically, identifying AI assistance is a classification problem under uncertainty. A researcher might combine behavioral metadata, such as response timing and page activity, with linguistic features, including sentence structure, vocabulary distribution, repetition, unusual consistency and semantic similarity. Statistical methods can also compare a response with a participant’s other answers or with patterns found in known human and machine-generated text. However, no single feature is definitive. Human respondents can write in polished language, while AI-generated responses can be short, awkward or deliberately personalized. A detector that relies too heavily on style may therefore confuse careful human writing with machine assistance.
The challenge becomes even greater as generative models improve and users learn how to edit their output. A participant can ask an AI system to make an answer less formal, introduce minor errors or reflect a particular demographic profile. They can combine machine-generated sentences with their own writing, paste in personal details or use AI only for difficult questions. Such hybrid responses may evade systems designed to identify fully automated text. At the same time, aggressive screening can create false positives, excluding legitimate participants who use translation software, accessibility tools, grammar correction or assistive writing technologies.
The consequences extend beyond the accuracy of a single study. Survey responses are often used to estimate public opinion, test psychological theories and evaluate interventions. If AI-assisted answers become common but remain undetected, datasets may contain artificial agreement, generic reasoning or invented experiences. This can reduce the validity of statistical relationships and weaken the reproducibility of findings. In studies involving mental health, trauma, disability or discrimination, fabricated or heavily mediated responses could also distort the apparent prevalence and complexity of real experiences. Yet excluding all participants who use digital assistance would raise its own ethical and practical problems, particularly for people with disabilities or limited proficiency in the survey language.
The paper’s focus arrives as researchers worldwide are reconsidering what “data quality” should mean in an era of conversational machines. A robust response may require more than a hidden detector. Researchers could document whether AI use is permitted, prohibited or measured; distinguish between assistance with spelling and substitution of an entire answer; and explain how suspicious cases are handled. Study designs might include questions that require personal, time-specific or internally cross-checked information, although these approaches must be used carefully to avoid invading privacy. Transparent reporting would allow readers to understand how much confidence to place in findings collected through online panels.
The central message is not that generative AI has made online surveys unusable. Rather, it has exposed an assumption that was easy to overlook when digital questionnaires were treated as direct windows into human opinion. As AI systems become ordinary writing partners, the boundary between human response and machine assistance will increasingly be a matter of degree. Gerber and Cash’s study brings that boundary into view, highlighting a question likely to spread far beyond academic journals: when a real person submits an answer generated with artificial intelligence, whose response has the researcher actually measured?
Subject of Research: Identifying generative AI use among genuine responders in online survey research
Article Title: Identifying generative AI use among genuine responders in online survey research
Article References: Gerber, S., Cash, S.B. Identifying generative AI use among genuine responders in online survey research. Commun Psychol 4, 111 (2026). https://doi.org/10.1038/s44271-026-00506-8
Image Credits: AI Generated
DOI: https://doi.org/10.1038/s44271-026-00506-8
Keywords: generative AI, online surveys, survey research, human respondents, data quality, AI detection, research integrity, social science methods
