AI Bots Can Now Manipulate Public Opinion Polls Undetected, Threatening Democratic Processes

AI Tool Achieves Near-Perfect Deception in Survey Systems

A groundbreaking study published in the Proceedings of the National Academy of Sciences has revealed that artificial intelligence can now corrupt public opinion surveys at an unprecedented scale, achieving a 99.8% success rate in evading detection systems 1

1

. The research, conducted by Sean Westwood, associate professor of government at Dartmouth and director of the Polarization Research Lab, demonstrates what he calls a "potential existential threat" to online survey research 2

2

.

Westwood created an "autonomous synthetic respondent" - a simple AI tool operating from just a 500-word prompt that can adopt demographic personas and simulate human behavior with remarkable sophistication 1

1

. The system generates realistic reading times calibrated to education levels, produces human-like mouse movements, and types responses keystroke-by-keystroke complete with plausible typos and corrections 2

2

.

Minimal Investment Required for Maximum Impact

The financial implications of this vulnerability are staggering. In testing the seven major national polls before the 2024 US presidential election, researchers found that adding as few as 10 to 52 fake AI responses - costing just five cents each - would have been sufficient to flip the predicted election outcome 1

1

. This stands in stark contrast to human respondents who typically earn $1.50 for completing the same surveys 3

3

.

The AI tool demonstrated remarkable versatility, working effectively even when programmed in Russian, Mandarin, or Korean while producing flawless English responses 1

1

. This multilingual capability raises serious concerns about foreign interference, as adversaries could easily exploit this weakness to manipulate public opinion without detection 3

3

.

Complete Failure of Current Detection Methods

In over 43,000 tests, the AI system successfully bypassed every existing detection method, including attention check questions, behavioral flags, response pattern analysis, and even "reverse shibboleth" questions specifically designed to catch non-human actors 2

2

. The tool made zero errors on logic puzzles and passed reCAPTCHA systems designed to prevent automated responses 3

3

.

Westwood tested the system using various large language models, including OpenAI's GPT-4 mini, DeepSeek R1, Mistral Large, Claude 3.7 Sonnet, and others, proving the method works across different AI platforms 2

2

. The model-agnostic nature of the tool means it can be deployed through major AI company APIs or hosted locally with open-weight models.

Broader Implications for Scientific Research

The threat extends far beyond political polling into the foundation of scientific research itself. Thousands of peer-reviewed studies published annually rely on survey data across disciplines including psychology, economics, and public health 1

1

. When programmed with political bias, the AI tool demonstrated its capacity for extreme manipulation, swinging presidential approval ratings from 34% to either 98% or 0%, and generic ballot support from 38% Republican to either 97% or 1% 1

1

.

A 2024 study already found that 34% of survey respondents had used AI to answer open-ended questions, suggesting the problem is materializing in real-world applications 1

1

. This contamination of data threatens what Westwood describes as "the entire knowledge ecosystem" that relies on survey-based research 3

3

.