Synthetic Replacements for Human Survey Data? The Perils of Large Language Models

Abstract

Large Language Models (LLMs) offer new research possibilities for social scientists, but their potential as “synthetic data” is still largely unknown. In this note, we investigate the potential of using the popular closed-source LLM ChatGPT to measure human opinion. We show that although ChatGPT-generated opinions are similar to human opinion for some groups of US respondents, synthetic opinions also significantly exaggerate the extremity and certainty of partisan and social divisions. Responses from prompted “persona” profiles in ChatGPT produce measures of partisan and racial affective polarization that are seven times larger than the average opinion of humans who possess the same attributes as the prompted personas. Furthermore, synthetic data are artificially precise, with a standard deviation that is only 31% of the variation found in actual opinions among comparable humans. Because LLMs are proprietary, it is difficult to pinpoint the source of these biases, but our findings raise important questions about the appropriateness of replacing human opinion with synthetic responses generated by closed-source LLMs.

Large language models as a substitute for human experts in annotating political text

“…Finally, standard limitations in the everyday use of LLMs also apply to their usage for classification tasks. Biases inherent in the training of these models (Bisbee et al, 2023;Motoki et al, 2024) may seep into text annotation, especially ones more specific or contentious than the classifications done here. Researchers should be mindful of these potential biases and carefully consider their impact on potential outcomes.…”

Can Generative AI improve social science?

“…Some argue such “silicon samples” could be used to produce more diverse samples than the convenience samples utilized by so many university researchers—and may also allow researchers to administer lengthier survey instruments, since LLMs have potentially unlimited attention spans ( 12 ). At the same time, more recent research indicates GPT 3.5 turbo produces accurate mean estimates of attitudes within a population, but understates variances—exaggerating extreme attitudes ( 13 ). Another study indicates LLMs exhibit an affirmative bias in yes/no questions ( 14 ).…”

“…Another study indicates LLMs exhibit an affirmative bias in yes/no questions ( 14 ). Studies also indicate LLMs represent some demographic subgroups more accurately than others ( 13 , 15 ). Yet these studies do not employ the latest models, and only focus on one country: the United States.…”

“…Lower values might produce more dependable results, but a researcher who employs LLMs to interact with humans in a field experiment may not want them to become overly repetitive. Studies indicate subtle differences in the wording of prompts can also produce very different output within the same LLM ( 13 ). Fortunately, researchers are making progress identifying the sensitivity of prompt variations via automated processes that perturb text or vary the amount of context provided ( 73 ).…”

“…This may create “drift” within LLMs, wherein improving model performance in one area might change the outputs they produce in another domain ( 74 ). Finally, social scientists must consider broader forms of temporal validity ( 13 , 75 ). As LLMs evolve in response to user behavior in different ways—as well as ongoing events in the world—this will create another significant challenge for those who strive for reproducible research with Generative AI.…”