TL;DR
This document presents the proceedings of SICon 2025, a workshop dedicated to understanding and modeling social influence (SI) in human-AI conversations. It highlights the growing importance of SI for creating more natural, effective, and personalized AI interactions, and showcases research spanning NLP, AI, psychology, and communication.
What Problem Does This Paper Address?
The core problem addressed by the research presented in these proceedings is the need to imbue AI systems with the ability to understand and exert social influence (SI) in conversations. Social influence, defined as the change in an individual's thoughts, feelings, attitudes, or behaviors due to interaction, is a fundamental aspect of human communication. Current AI systems often lack this capability, leading to interactions that feel unnatural, unpersuasive, or unable to adapt to user needs. This gap limits the effectiveness of AI in various applications, from negotiation and persuasion to therapy and customer service. The workshop aims to bridge this gap by fostering interdisciplinary research that integrates insights from fields like psychology, communication, and game theory with NLP and AI.
Methods
The proceedings showcase a variety of methods applied to the study of social influence in conversations. These include:
- LLM Roleplay: Simulating human-chatbot interactions to study influence dynamics.
- Prompt Engineering and Fine-tuning: Investigating best practices for using Large Language Models (LLMs) in social science tasks, focusing on prompt refinement and fine-tuning strategies.
- Multimodal Deception Detection: Developing benchmarks for detecting deception, which is often intertwined with social influence tactics, using multimodal data.
- Persona-Grounded Dialogue Evaluation: Assessing the persuasiveness of LLMs when they adopt specific personas in dialogues.
- Ethical Compliance and Trust Frameworks: Developing taxonomies for evaluating ethical compliance and trust in AI systems, which are crucial for understanding the responsible application of social influence.
- Empathy Evaluation: Measuring LLM empathy in conversations, a key component of building rapport and influencing users positively.
- Gendered Discourse Analysis: Examining how gendered language in podcasts and LLMs reflects and potentially perpetuates social biases.
- Intent-Driven Multi-Agent Frameworks: Creating frameworks for analyzing manipulation in dialogues, particularly in complex settings like courtrooms.
- Conversational Steering: Developing methods for LLMs to guide long emotional support conversations effectively.
- Human-like Conversational Behaviors: Investigating LLM behaviors that mimic human conversational patterns, including text overlap.
- Advertising Response Modeling: Creating models of conversational engagement to understand social influence in consumer responses to advertising.
- Parameter-Efficient Fine-Tuning: Exploring techniques for fine-tuning LLMs to balance linguistic adaptation and safety in social influence systems.
Datasets
The paper does not explicitly list specific datasets used for training and evaluation across all presented works. However, the research implies the use of:
- Dialogue datasets for training and evaluating conversational AI systems.
- Datasets related to persuasion, negotiation, therapy, and argumentation.
- Multimodal datasets for deception detection.
- Podcast transcripts and LLM-generated dialogues for discourse analysis.
- Courtroom dialogue transcripts for manipulation analysis.
Key Insights and Ideas
- The Ubiquity of Social Influence: Social influence is a pervasive element in all human interactions, and its absence in AI makes them less effective and natural.
- Interdisciplinary Necessity: Effectively modeling social influence requires a multidisciplinary approach, drawing from NLP, AI, psychology, communication, and game theory.
- LLMs as Tools for SI Research: LLMs are proving to be powerful tools for both studying and implementing social influence in AI, enabling more sophisticated dialogue generation and analysis.
- Ethical Considerations are Paramount: As AI systems become more capable of influencing users, ethical considerations, trust, and safety become critical research areas.
- Nuance in Conversational Dynamics: Understanding subtle conversational cues, like text overlap and empathy, is key to building influential and engaging AI.
- Persona and Context Matter: The effectiveness of social influence is highly dependent on the persona adopted by the AI and the specific conversational context.
Key Results
The proceedings present a range of results from various research papers, including:
- LLM Roleplay: Demonstrated success in simulating human-chatbot interactions for studying influence.
- DecepBench: A benchmark for multimodal deception detection, indicating progress in identifying deceptive behaviors.
- Persona-Grounded Dialogues: Evaluation of LLM persuasiveness in persona-driven conversations.
- PROTECT Framework: A taxonomy for ethical compliance and trust, providing a structured approach to evaluating AI behavior.
- LLM Empathy: Evaluation of LLM empathy in Korean conversations, showing varying degrees of success.
- Masculine Defaults: Analysis revealing gendered discourse patterns in podcasts and LLMs.
- Courtroom Dialogue Analysis: Development of an intent-driven framework for analyzing manipulation.
- Conversational Steering: Methods for LLMs to manage long emotional support conversations.
- Text Overlap: Research on LLMs exhibiting human-like conversational behaviors, including text overlap.
- Advertising Response: A model for understanding social influence in consumer responses to advertising.
- Parameter-Efficient Fine-Tuning: Exploration of techniques to balance linguistic adaptation and safety in SI systems.
Specific quantitative results (e.g., accuracy scores, performance metrics) are not consolidated into a single summary but are detailed within individual papers.
Limitations and Future Work
- Data Scarcity: The development of robust SI models is often hindered by the lack of large-scale, diverse datasets specifically annotated for social influence phenomena.
- Ethical Challenges: The potential for misuse of social influence capabilities by AI systems raises significant ethical concerns that require ongoing research and mitigation strategies.
- Generalizability: Models developed for specific conversational tasks or domains may not generalize well to other contexts.
- Measuring Influence: Quantifying the precise impact of social influence in complex, multi-turn conversations remains a challenge.
- Future Work: The proceedings suggest future research directions in developing more sophisticated SI models, exploring novel applications, enhancing ethical safeguards, and fostering greater interdisciplinary collaboration. Specific areas include improving deception detection, understanding nuanced persuasive strategies, and creating AI that can ethically and effectively navigate social dynamics in conversations.