AI-Generated Responses Undermine Crowdsourced Research Studies

Online surveys are being inundated by responses generated through AI, potentially compromising the integrity of critical data for scientific research.

Platforms like Prolific compensate participants modestly for answering questions posed by researchers. These platforms have gained popularity among academics for their simplicity in attracting subjects for behavioral studies.

Anne Marie Nusberger and her team at the Max Planck Institute for Human Development in Berlin, Germany, set out to examine the frequency of AI usage among respondents, triggered by their observations in previous studies. “The rate we were witnessing was truly startling,” she remarks.

They suspect that 45% of participants who submitted a single open-ended question on Prolific utilized AI tools to streamline their responses.

Further analysis of these submissions indicated more overt references to AI usage, characterized by phrases like “excessively repetitive” and “clearly non-human” language. “From the data we gathered earlier this year, it’s clear that a notable fraction of research is tainted,” she explains.

In follow-up studies conducted via Prolific, researchers implemented traps to capture chatbot users. Two instances of Recaptcha — a small test designed to differentiate humans from bots — identified only 0.2% of users as bots. A more complex Recaptcha, using both past activity and current behavior, eliminated an additional 2.7%. Although hidden from view, bots that were prompted to include the word “hazelnut” in their responses accounted for another 1.6%, while an extra 4.7% were detected when copying and pasting was restricted.

“Our goal is to respond adequately to online surveys, rather than resorting to full distrust,” advises Nussberger. It’s the onus of researchers, in her view, to handle the answers with greater skepticism and take precautions against AI-induced input. “However, the platforms bear significant responsibility. They must treat this matter with utmost seriousness.”

Prolific did not respond to a request for comment from New Scientist.

“The validity of online behavioral research has already faced challenges from participants misrepresenting themselves or employing bots to obtain rewards,” says Matt Hodgkinson, a freelance consultant in research ethics. “Researchers must collectively explore remote validation of human involvement or return to traditional face-to-face methodologies.”