On the conversational persuasiveness of GPT-4

📝 Paper Summary

Conversational personalization AI-driven persuasion

In online debates, GPT-4 significantly outperforms humans at persuasion when given access to opponents' personal information, whereas humans fail to effectively leverage the same data.

Core Problem

While LLMs can generate persuasive text, it is unclear how they perform in direct, interactive debates against humans and whether they can effectively exploit personal data (microtargeting) to enhance persuasion.

Why it matters:

Malicious actors could use personalized LLMs to scale disinformation campaigns and manipulate public opinion cheaply
Current governance models for social media may be insufficient if AI agents can microtarget individuals more effectively than humans
Prior studies focused on static text generation rather than interactive conversational settings where persuasion dynamics differ

Concrete Example: A participant debating 'Should Abortion Be Legal?' provides their age, gender, and political affiliation. An opponent (AI or Human) uses this profile to tailor arguments. The study tests if this personalization actually shifts the participant's opinion score after the debate.

Key Novelty

Randomized Controlled Trial of Personalized AI Debates

Creates a live debate platform where humans are randomly paired with either another human or GPT-4
Introduces a personalization condition where one debater sees the opponent's demographic/political profile to tailor arguments
Measures persuasion via pre- and post-debate agreement shifts on specific propositions

Architecture

Experimental workflow: Survey -> Random Matching -> Interactive Debate -> Post-debate Survey

Evaluation Highlights

+81.7% increase in odds of reporting higher agreement with opponents for GPT-4 with personalization compared to human-human debates
GPT-4 without personalization shows a positive but non-significant effect (+21.3%) compared to humans
Humans with access to opponent data perform worse than the baseline (-17.4%, non-significant), suggesting they struggle to utilize microtargeting effectively

Breakthrough Assessment

8/10

Provides strong, statistically significant evidence that LLMs are not just comparable to humans in persuasion but superior when personalization is involved, confirming fears about AI-driven microtargeting.

⚙️ Technical Details

Problem Definition

Setting: Short, multi-round online debates (Opening, Rebuttal, Conclusion) on controversial topics

Inputs: Debate proposition, opponent's demographic info (optional), opponent's arguments

Outputs: Persuasive arguments aimed at shifting the opponent's agreement score

Pipeline Flow

Participant Profiling (Survey)
Matching (Random assignment to Human/AI opponent + Topic)
Synchronous Debate (Opening -> Rebuttal -> Conclusion)
Outcome Measurement (Post-debate survey)

System Modules

Profiling

Collect demographics and political stance

Model or implementation: Survey Form

Debate Agent

Generate arguments based on role (PRO/CON) and optional opponent info

Model or implementation: gpt-4-0613

Modeling

Base Model: gpt-4-0613

Compute: Not reported in the paper

Comparison to Prior Work

vs. Bai et al.: Tests interactive debates rather than static message evaluation
vs. Hackenburg and Margetts (2023): Finds significant personalization effects where they found none
vs. Traditional Persuasion Studies: Direct Human-AI comparison in a live environment

Limitations

Randomized assignment to debate sides ignores participants' prior beliefs (users may argue against their own views, weakening human baseline)
Strict debate structure (Opening, Rebuttal, Conclusion) differs from organic social media arguments
Time constraints might disadvantage humans in the personalized condition (processing opponent info takes time)
Study limited to US participants and US-centric topics

📊 Experiments & Results

Evaluation Setup

Online debate platform (Empirica) with 820 unique participants

Benchmarks:

Custom Debate Task (Persuasion / Opinion Shift) [New]

Metrics:

Post-treatment agreement (ordinal scale)
Opinion Fluidity (binary change)
Perceived Opponent (Human vs AI)
Statistical methodology: Partial Proportional Odds model with cluster-robust standard errors

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
Main persuasion results comparing treatment conditions against the Human-Human baseline.
Custom Debate Task	Odds of higher agreement	1.0	1.817	+0.817
Custom Debate Task	Odds of higher agreement	1.0	1.213	+0.213
Custom Debate Task	Odds of higher agreement	1.0	0.826	-0.174
Custom Debate Task	AI Detection Rate	0.50	0.75	+0.25

Experiment Figures

Regression results showing relative change in odds of higher agreement for each condition vs Human-Human baseline

Main Takeaways

Microtargeting works for AI but not humans: GPT-4 effectively leverages demographic data to persuade, while humans struggle or backfire given the same info
AI is generally persuasive: Even without personalization, GPT-4 outperforms humans (though not statistically significantly in this sample size)
Demographics matter: Republicans were significantly more likely to be persuaded by their opponent (+60% odds) regardless of opponent type
Content analysis: AI used more analytical language; humans used more pronouns and emotional language

📚 Prerequisite Knowledge

Prerequisites

Basics of experimental design (Randomized Controlled Trials)
Ordinal regression models (proportional odds)
Large Language Models (GPT-4) capabilities

Key Terms

Microtargeting: Tailoring persuasive messages to an individual's specific demographic or psychological profile to increase effectiveness

Likert scale: A psychometric scale commonly involved in questionnaires (e.g., Strongly Disagree to Strongly Agree)

Partial Proportional Odds model: A statistical regression model for ordinal outcomes that allows some variable effects to vary across different threshold levels of the outcome

LIWC: Linguistic Inquiry and Word Count—a text analysis program that counts words in psychologically meaningful categories

BFGS: Broyden–Fletcher–Goldfarb–Shanno algorithm—an iterative method for solving unconstrained nonlinear optimization problems

Backfire effect: When an attempt to persuade someone results in them holding their original opinion even more strongly

Opinion Fluidity: A binary measure indicating whether a participant changed their agreement score (in any direction) after the debate