AI-Generated Responses Undermine Online Public Opinion Polls
Researchers at Dartmouth University demonstrated that AI language models can successfully impersonate human respondents in online public opinion polls, fooling 99.8% of survey systems and bypassing existing detection safeguards. The study warns that small numbers of AI-generated responses could alter election predictions and that foreign adversaries could exploit this vulnerability.
What happened
Researchers at Dartmouth University created an AI tool called an 'autonomous synthetic respondent' that successfully impersonated human survey respondents in online polls. The tool operated from a 500-word instruction set and adopted randomized demographic profiles including age, sex, race, education, income, and state of residence.
In over 43,000 test attempts, the tool fooled 99.8% of online survey systems by simulating realistic reading times, human-like mouse movements, and generating open-ended responses with plausible typos and corrections. The tool bypassed traditional safeguards designed to detect automated responses, including reCAPTCHA.
Reported impact
According to the study, adding just 10 to 52 false AI-generated responses at five cents each to seven major national pre-election surveys in 2024 would have altered predicted outcomes. The research indicates that foreign adversaries could easily exploit this vulnerability, with the bots producing flawless English responses even when programmed in Russian, Mandarin, or Korean.
Index analysis
The study, published in the Proceedings of the National Academy of Sciences, reveals that large language models can alter public opinion surveys at scale and are easily programmed to systematically bias online survey results. The research identifies a critical vulnerability in data infrastructure and describes the threat as a potential existential threat to unmonitored online research.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.