AI and online panels

A fascinating paper about the potential threat AI can pose to online research panels.

The advancement of large language models poses a severe, potentially existential threat to online survey research, a fundamental tool for data collection across the sciences.This work demonstrates that the foundational assumption of survey research—that a coherent response is a human response—is no longer tenable.

I designed and tested an autonomous synthetic respondent capable of producing survey data that possesses the coherence and plausibility of human responses. This agent successfully evades a comprehensive suite of data quality checks, including instruction-following tasks, logic puzzles, and “reverse shibboleth” questions designed to detect nonhuman actors, achieving a 99.8% pass rate on 6,000 trials of standard attention checks.

The synthetic respondent generates internally consistent responses by maintaininga coherent demographic persona and a memory of its prior answers, producingplausible data on psychometric scales, vignette comprehension tasks, and complex socioeconomic trade-offs. Furthermore, its open-ended text responses are linguistically sophisticated and stylistically calibrated to the level of education of its assigned persona.

Critically, the agent can be instructed to maliciously alter polling outcomes, demonstrating an overt vector for information warfare. More subtly, it can also infer a researcher’s latent hypotheses and produce data that artificially confirms them.These findings reveal a critical vulnerability in our data infrastructure, rendering most current detection methods obsolete and posing a potential existential threat to unsupervised online research. The scientific community must urgently develop new data validation standards and reconsider its reliance on nonprobability, low-barrier online data collection methods.

This is quite stunning. Online panels do have numerous features to identify bots etc. One I use is asking if respondents have a licence to pilot a space shuttle. But this says that AI can blow past these features and fool them.

I suspect what will happen is a sort of AI arms race where panel providers will use AI more and more to detect bots, and the bad actors will use AI to get even smarter against being detected.

This is one reason why Curia has not gone online panel only for most polls, and does a mixture of phone and online. The phone samples, while costly, provide a useful reference against the online panels.