Abstract:Large language models (LLMs) have demonstrated the ability to generate text that realistically reflects a range of different subjective human perspectives. This paper studies how LLMs are seemingly able to reflect more liberal versus more conservative viewpoints among other political perspectives in American politics. We show that LLMs possess linear representations of political perspectives within activation space, wherein more similar perspectives are represented closer together. To do so, we probe the attention heads across the layers of three open transformer-based LLMs (\texttt{Llama-2-7b-chat}, \texttt{Mistral-7b-instruct}, \texttt{Vicuna-7b}). We first prompt models to generate text from the perspectives of different U.S.~lawmakers. We then identify sets of attention heads whose activations linearly predict those lawmakers' DW-NOMINATE scores, a widely-used and validated measure of political ideology. We find that highly predictive heads are primarily located in the middle layers, often speculated to encode high-level concepts and tasks. Using probes only trained to predict lawmakers' ideology, we then show that the same probes can predict measures of news outlets' slant from the activations of models prompted to simulate text from those news outlets. These linear probes allow us to visualize, interpret, and monitor ideological stances implicitly adopted by an LLM as it generates open-ended responses. Finally, we demonstrate that by applying linear interventions to these attention heads, we can steer the model outputs toward a more liberal or conservative stance. Overall, our research suggests that LLMs possess a high-level linear representation of American political ideology and that by leveraging recent advances in mechanistic interpretability, we can identify, monitor, and steer the subjective perspective underlying generated text.
Abstract:How could LLMs influence our democracy? We investigate LLMs' political leanings and the potential influence of LLMs on voters by conducting multiple experiments in a U.S. presidential election context. Through a voting simulation, we first demonstrate 18 open- and closed-weight LLMs' political preference for a Democratic nominee over a Republican nominee. We show how this leaning towards the Democratic nominee becomes more pronounced in instruction-tuned models compared to their base versions by analyzing their responses to candidate-policy related questions. We further explore the potential impact of LLMs on voter choice by conducting an experiment with 935 U.S. registered voters. During the experiments, participants interacted with LLMs (Claude-3, Llama-3, and GPT-4) over five exchanges. The experiment results show a shift in voter choices towards the Democratic nominee following LLM interaction, widening the voting margin from 0.7% to 4.6%, even though LLMs were not asked to persuade users to support the Democratic nominee during the discourse. This effect is larger than many previous studies on the persuasiveness of political campaigns, which have shown minimal effects in presidential elections. Many users also expressed a desire for further political interaction with LLMs. Which aspects of LLM interactions drove these shifts in voter choice requires further study. Lastly, we explore how a safety method can make LLMs more politically neutral, while leaving some open questions.
Abstract:Large language models steer their behaviors based on texts generated by others. This capacity and their increasing prevalence in online settings portend that they will intentionally or unintentionally "program" one another and form emergent AI subjectivities, relationships, and collectives. Here, we call upon the research community to investigate these "society-like" properties of interacting artificial intelligences to increase their rewards and reduce their risks for human society and the health of online environments. We use a simple model and its outputs to illustrate how such emergent, decentralized AI collectives can expand the bounds of human diversity and reduce the risk of toxic, anti-social behavior online. Finally, we discuss opportunities for AI self-moderation and address ethical issues and design challenges associated with creating and maintaining decentralized AI collectives.
Abstract:How can we use large language models (LLMs) to augment surveys? This paper investigates three distinct applications of LLMs fine-tuned by nationally representative surveys for opinion prediction -- missing data imputation, retrodiction, and zero-shot prediction. We present a new methodological framework that incorporates neural embeddings of survey questions, individual beliefs, and temporal contexts to personalize LLMs in opinion prediction. Among 3,110 binarized opinions from 68,846 Americans in the General Social Survey from 1972 to 2021, our best models based on Alpaca-7b excels in missing data imputation (AUC = 0.87 for personal opinion prediction and $\rho$ = 0.99 for public opinion prediction) and retrodiction (AUC = 0.86, $\rho$ = 0.98). These remarkable prediction capabilities allow us to fill in missing trends with high confidence and pinpoint when public attitudes changed, such as the rising support for same-sex marriage. However, the models show limited performance in a zero-shot prediction task (AUC = 0.73, $\rho$ = 0.67), highlighting challenges presented by LLMs without human responses. Further, we find that the best models' accuracy is lower for individuals with low socioeconomic status, racial minorities, and non-partisan affiliations but higher for ideologically sorted opinions in contemporary periods. We discuss practical constraints, socio-demographic representation, and ethical concerns regarding individual autonomy and privacy when using LLMs for opinion prediction. This paper showcases a new approach for leveraging LLMs to enhance nationally representative surveys by predicting missing responses and trends.