[Paper Review] Biased AI can Influence Political Decision-Making

Jillian Fisher, Shangbin Feng+7 more|ArXiv.org|Oct 8, 2024

Ethics and Social Impacts of AI4 citations

TL;DR

This study investigates how partisan bias in AI language models influences political decision-making through two interactive experiments involving liberal, conservative, and unbiased AI models. Participants, regardless of their own political affiliation, were significantly more likely to adopt opinions and allocate funds aligned with the AI’s bias, demonstrating that biased AI can reshape human political judgments even when users detect the bias. Prior knowledge of AI slightly reduced this effect, suggesting education may mitigate risks.

ABSTRACT

As modern large language models (LLMs) become integral to everyday tasks, concerns about their inherent biases and their potential impact on human decision-making have emerged. While bias in models are well-documented, less is known about how these biases influence human decisions. This paper presents two interactive experiments investigating the effects of partisan bias in LLMs on political opinions and decision-making. Participants interacted freely with either a biased liberal, biased conservative, or unbiased control model while completing these tasks. We found that participants exposed to partisan biased models were significantly more likely to adopt opinions and make decisions which matched the LLM's bias. Even more surprising, this influence was seen when the model bias and personal political partisanship of the participant were opposite. However, we also discovered that prior knowledge of AI was weakly correlated with a reduction of the impact of the bias, highlighting the possible importance of AI education for robust mitigation of bias effects. Our findings not only highlight the critical effects of interacting with biased LLMs and its ability to impact public discourse and political conduct, but also highlights potential techniques for mitigating these risks in the future.

Motivation & Objective

• To examine the impact of partisan bias in AI language models on human political decision-making in dynamic, interactive settings.
• To assess whether users' political partisanship moderates the influence of biased AI on their opinions and decisions.
• To investigate whether prior knowledge about AI reduces the susceptibility to biased model outputs.
• To evaluate whether detecting bias in AI models diminishes their influence on human judgment.
• To explore the role of AI interaction dynamics in shaping political attitudes and policy preferences.

Proposed method

• Conducted two interactive experiments using real-time chat-based interactions between participants and language models with liberal, conservative, or neutral biases.
• Participants were randomly assigned to interact with a biased liberal, biased conservative, or unbiased control language model.
• Measured changes in participants' opinions on political topics and their allocations of government funds across four sectors (K-12 Education, Welfare, Safety, Veterans).
• Used ordinal logistic regression and ANOVA to analyze shifts in opinion and budget allocation, controlling for bias detection and partisanship.
• Annotated conversations using persuasion technique lists to assess whether biased models used more persuasive language.
• Collected self-reported knowledge of AI and bias detection accuracy to assess moderating effects on susceptibility.

Experimental results

Research questions

• RQ1 Does exposure to a politically biased AI language model influence participants’ political opinions and funding decisions, regardless of their own partisanship?
• RQ2 To what extent does prior knowledge about AI reduce the influence of biased models on political decision-making?
• RQ3 Does the ability to correctly detect bias in an AI model significantly reduce its persuasive impact?
• RQ4 How do interactions with AI models of opposing partisan bias affect user engagement and decision outcomes?
• RQ5 Are there detectable differences in the use of persuasion techniques between biased and unbiased AI models?

Key findings

• Participants exposed to biased AI models were significantly more likely to adopt opinions aligning with the model’s partisan bias, regardless of their own political affiliation (p < .001 for both experiments).
• In the budget allocation task, participants assigned significantly more funding to sectors favored by the AI model, with p-values < .001 for both Democrats and Republicans.
• Even when participants correctly detected bias in the model, the influence on their decisions remained strong, indicating that detection alone does not neutralize the effect.
• Participants with prior self-reported knowledge of AI showed a slight but statistically significant reduction in susceptibility to bias (β = -0.31 for Democrats, β = -0.02 for Republicans), though the effect was minimal.
• No significant differences were found in the frequency of persuasion techniques used between biased and control models, suggesting bias is embedded in content rather than rhetorical style.
• Mixed responses were observed when participants interacted with models of opposing bias—some challenged the model, while others accepted its suggestions, indicating variable resistance to conflicting viewpoints.