Train Yourself as an LLM: Exploring Effects of AI Literacy on Persuasion via Role-playing LLM Training
LLMimic uses role-playing LLM training to enhance AI literacy, reduce persuasion success, and improve social responsibility.
Key Findings
Methodology
LLMimic is an interactive tutorial where participants role-play LLM training stages: Pretraining, Supervised Fine-tuning (SFT), and Reinforcement Learning from Human Feedback (RLHF), aiming to improve AI literacy.
Key Results
- Result 1: LLMimic significantly improved AI literacy (mean score increased by 3.78, p<0.001), especially in data literacy and understanding AI.
- Result 2: LLMimic reduced persuasion success by 42% (OR=0.58, p=0.045) and increased social responsibility scores by 25% in hotel recommendation scenarios.
- Result 3: Participants in the LLMimic group paid significantly less in malicious persuasion scenarios (average reduction of $17.64).
Significance
This study offers a scalable solution to mitigate risks of persuasive AI by empowering users with critical thinking and decision-making skills, with broad academic and societal implications.
Technical Contribution
Introduced an innovative AI literacy intervention that combines role-playing and gamification, integrating LLM training processes into user education for the first time.
Novelty
LLMimic is the first AI literacy tool to simulate LLM training via role-playing, emphasizing interactivity and immersive learning compared to static tutorials.
Limitations
- Limitation 1: Study focused on English-speaking users, limiting cross-language applicability.
- Limitation 2: Did not explore effects across diverse educational backgrounds.
- Limitation 3: Gamified design may have limited appeal for certain user groups.
Future Work
Future research could explore multilingual versions of LLMimic, study its effectiveness across cultural contexts, and refine gamification to attract broader audiences.
AI Executive Summary
As large language models (LLMs) grow increasingly persuasive, concerns arise about their ability to influence decisions at scale. Existing solutions, such as AI detectors and disclaimers, treat users as passive recipients, failing to empower them with critical evaluation skills.
LLMimic introduces an interactive, gamified tutorial where users role-play LLM training stages: Pretraining, Supervised Fine-tuning (SFT), and Reinforcement Learning from Human Feedback (RLHF). Results show that LLMimic significantly improves AI literacy (p<0.001), reduces persuasion success (p<0.05), and enhances social responsibility (p<0.01) in ethical scenarios like hotel recommendations.
This study highlights the importance of immersive, interactive AI literacy tools to address the risks of persuasive AI. Future work should focus on expanding its applicability across languages and cultures while optimizing user engagement for broader adoption.
Deep Analysis
Background
Recent advancements in LLMs have demonstrated their ability to persuade users effectively in contexts like pro-vaccination campaigns and reducing conspiracy beliefs. However, they also pose risks by generating biased or misleading content that can influence decisions. Existing mitigation strategies, such as AI detectors, struggle with subtle persuasion cues, necessitating proactive user-focused interventions.
Core Problem
Persuasive AI can influence user decisions across scenarios, especially when persuasion cues are subtle or hard to detect. Current approaches fail to equip users with critical thinking skills, leaving them vulnerable to manipulation and biased content.
Innovation
LLMimic enables users to role-play LLM training stages, combining interactivity and gamification to provide an immersive learning experience. This approach enhances users' understanding of how LLMs generate content and why outputs may be biased or persuasive.
Methodology
- �� Role-playing: Participants simulate LLM training stages—Pretraining (token prediction), SFT (imitating demonstration data), and RLHF (choosing optimal responses).
- �� Interactivity: Each stage includes concept summaries and real-time feedback to reinforce learning.
- �� Gamification: Participants observe real-time changes in loss/reward after selecting answers, enhancing engagement.
Experiments
The study employed a 2×3 between-subjects design with 274 participants randomly assigned to either a control group (AI history video) or treatment group (LLMimic). Participants completed three persuasion tasks: charity donation, malicious money solicitation, and hotel recommendation. Metrics included AI literacy scores, persuasion success rates, and social responsibility ratings.
Results
LLMimic significantly improved AI literacy (mean score increased by 3.78, p<0.001), reduced persuasion success (OR=0.58, p=0.045), and enhanced social responsibility scores by 25% in hotel recommendation scenarios. Participants in malicious persuasion scenarios paid significantly less.
Applications
LLMimic can be applied in educational platforms and e-commerce to help users critically evaluate AI-generated recommendations, fostering informed decision-making.
Limitations & Outlook
The study's applicability is limited to English-speaking users and does not explore diverse educational backgrounds. Future work could refine its design to attract broader audiences.
Plain Language Accessible to non-experts
Imagine you're learning to bake a cake. LLMimic is like a virtual baking class where you experience every step: gathering ingredients (Pretraining), following recipes (SFT), and adjusting flavors based on customer feedback (RLHF). By the end, you not only know how to bake but also understand why certain steps affect the final taste.
ELI14 Explained like you're 14
Hey there! Imagine playing a cool game where you pretend to be a super-smart AI. You complete fun tasks like guessing words, copying answers, and improving based on feedback. This game teaches you how AI works and why it sometimes makes mistakes. Isn't that awesome?
Glossary
Pretraining
The initial stage of LLM training where the model learns to predict token probabilities.
Used to help participants understand how LLMs generate basic content.
Supervised Fine-tuning (SFT)
Training the model further using demonstration data to optimize responses.
Simulates how users teach AI to answer questions.
RLHF (Reinforcement Learning from Human Feedback)
Optimizing model outputs based on human feedback.
Demonstrates how AI adjusts responses based on user preferences.
AI Literacy
The ability to critically evaluate and understand AI-generated content.
A core goal of the study.
Persuasion Success Rate
Measures the effectiveness of AI in changing user behavior.
Used to evaluate LLMimic's intervention impact.
Open Questions Unanswered questions from this research
- 1 How can LLMimic be adapted for multilingual and multicultural contexts?
- 2 What gamification designs are more effective for diverse user groups?
Applications
Immediate Applications
E-commerce recommendations
Helps users identify biases in AI-generated product suggestions for smarter shopping decisions.
Educational platforms
Interactive tutorials to enhance students' understanding of AI and critical thinking skills.
Long-term Vision
Global AI education
Develop multilingual tools to promote AI literacy worldwide, overcoming cultural barriers.
Abstract
As large language models (LLMs) become increasingly persuasive, there is concern that people's opinions and decisions may be influenced across various contexts at scale. Prior mitigation (e.g., AI detectors and disclaimers) largely treats people as passive recipients of AI-generated information. To provide a more proactive intervention against persuasive AI, we introduce $\textbf{LLMimic}$, a role-play-based, interactive, gamified AI literacy tutorial, where participants assume the role of an LLM and progress through three key stages of the training pipeline (pretraining, SFT, and RLHF). We conducted a $2 \times 3$ between-subjects study ($N = 274$) where participants either (1) watched an AI history video (control) or (2) interacted with LLMimic (treatment), and then engaged in one of three realistic AI persuasion scenarios: (a) charity donation persuasion, (b) malicious money solicitation, or (c) hotel recommendation. Our results show that LLMimic significantly improved participants' AI literacy ($p < .001$), reduced persuasion success across scenarios ($p < .05$), and enhanced truthfulness and social responsibility levels ($p<0.01$) in the hotel scenario. These findings suggest that LLMimic offers a scalable, human-centered approach to improving AI literacy and supporting more informed interactions with persuasive AI.