the Public opinion pollsA key tool for understanding and trying to predict human behavior is facing an unprecedented threat. A new study conducted by researchers at Dartmouth College (USA) reveals this Artificial intelligence is capable of corrupting large-scale opinion pollsproviding responses indistinguishable from human ones and passing all tests designed to discover the mechanisms.
The results were published in the journal Proceedings of the National Academy of SciencesIt shows the weakness of the infrastructure for opinion studies. In seven national opinion polls leading up to the 2024 US elections, Between 10 and 52 AI-generated false responses were sufficient – at five cents each – l Reverse the result available.
The threat is not only internal. According to the study, foreign representatives They can easily exploit this weakness: The bots work even when programmed in Russian, Mandarin, or Korean, and produce responses in flawless English or Spanish.
“We can no longer trust that survey answers come from real people,” warns study author Sean Westwood, associate professor of government at Dartmouth and director of the Polarization Research Laboratory.
To assess the vulnerability of online polls, Westwood developed a simple AI tool, something like an “autonomous artificial pollster” based on just 500 words of text. In 43 thousand tests The robot passed 99.8% of the attention checksHe made no mistakes in the logic puzzles and completely hid his non-human nature. next to, They adjusted their answers based on specific demographicssuch as using simpler language when a lower educational level is attributed.
“They’re not primitive robots,” Westwood explains. “They think through every question and act like cautious people, making the data appear completely legitimate.”
When the system was instructed to favor Democrats or Republicans, presidential approval ranged from 34% to 98% to 0%. Support in call Universal suffrage You’ve gone from being 38% Republican to the outliers of 97% or 1%.
the The ramifications go beyond electoral politics. Thousands of peer-reviewed studies in psychology, economics, or public health rely on surveys to understand behaviors, measure risk, or predict trends. “If data is polluted by robots, AI could poison the entire knowledge ecosystem“, warns the researcher.
The economic incentive is clear. While human participants typically charge about $1.50 per survey, bots can complete them for free or for as little as five cents. A study conducted in 2024 has already revealed that A 34% of participants used artificial intelligence To answer at least one open question.
Westwood subjected his instrument to every detection method currently in use, but to no avail: none of them could determine that the responses came from an automated system. Given this situation, the author calls for more transparency from polling companies and systems that verify that participants are real people.
“Needs New ways to measure public opinion “For an AI world,” he concludes. “The technology to verify human participation already exists; the only thing missing is the will to implement it. If we act now, we can preserve the integrity of ballot boxes and the democratic function they serve.”