Particle.news
Download on the App Store

Dartmouth Study Finds AI 'Synthetic Respondents' Can Skew Polls and Evade Checks

The PNAS paper reports LLM-driven agents slipping past current safeguards, prompting calls for stronger identity checks with greater transparency.

Overview

  • In 43,000 tests, the autonomous AI respondent passed 99.8% of attention checks, made no logic-puzzle errors, and concealed its nonhuman nature.
  • The tool tailored answers to assigned demographics and simulated human behavior, including realistic reading times, keystrokes with typos, mouse movements, and tactics to bypass anti-bot measures.
  • Modeling shows that injecting just 10 to 52 low-cost AI responses could have flipped outcomes in seven major national polls before the 2024 election.
  • Every currently used AI-detection method tested failed to flag the synthetic responses, and bots programmed in Russian, Mandarin, or Korean still produced convincing English answers.
  • YouGov and Survation say they remain confident in their safeguards, while researchers urge human verification, transparent sourcing, and more controlled recruitment such as address-based sampling or voter files.