[Paper Review] How Well Can LLMs Echo Us? Evaluating AI Chatbots' Role-Play Ability with ECHO

Man Tik Ng, Hui Tung Tse+4 more|arXiv (Cornell University)|Apr 22, 2024

Artificial Intelligence in Healthcare and EducationMedicine3 citations

TL;DR

This paper introduces ECHO, a Turing test-inspired framework that evaluates large language models (LLMs) in role-playing as ordinary individuals using real-person profiles. It demonstrates that GPTs achieves a 48.3% success rate in deceiving acquaintances, outperforming GPT-3.5 and GPT-4, while LLMs like GPT-4 can detect machine-generated text but struggle to identify human-authored content, revealing model-specific biases.

ABSTRACT

The role-play ability of Large Language Models (LLMs) has emerged as a popular research direction. However, existing studies focus on imitating well-known public figures or fictional characters, overlooking the potential for simulating ordinary individuals. Such an oversight limits the potential for advancements in digital human clones and non-player characters in video games. To bridge this gap, we introduce ECHO, an evaluative framework inspired by the Turing test. This framework engages the acquaintances of the target individuals to distinguish between human and machine-generated responses. Notably, our framework focuses on emulating average individuals rather than historical or fictional figures, presenting a unique advantage to apply the Turing Test. We evaluated three role-playing LLMs using ECHO, with GPT-3.5 and GPT-4 serving as foundational models, alongside the online application GPTs from OpenAI. Our results demonstrate that GPT-4 more effectively deceives human evaluators, and GPTs achieves a leading success rate of 48.3%. Furthermore, we investigated whether LLMs could discern between human-generated and machine-generated texts. While GPT-4 can identify differences, it could not determine which texts were human-produced. Our code and results of reproducing the role-playing LLMs are made publicly available via https://github.com/CUHK-ARISE/ECHO.

Motivation & Objective

• To address the gap in evaluating LLMs' ability to emulate ordinary individuals rather than celebrities or fictional characters.
• To develop a framework that enables Turing test-based evaluation of LLMs in role-playing scenarios using real human profiles.
• To assess whether LLMs can distinguish between human- and machine-generated text, identifying potential biases in their judgment.
• To ensure ethical data handling by protecting participant privacy and obtaining informed consent.
• To provide a reproducible benchmark for evaluating role-playing LLMs using real personal data.

Proposed method

• ECHO constructs role-playing LLMs using personal background data from ten real individuals, including demographics, interests, and communication styles.
• Human evaluators—acquaintances of the target individuals—are tasked with identifying which responses were generated by the actual person versus the LLM.
• The framework compares four role-playing methods: RoleGPT, Juliet, Role-Play Prompting (RPP), and OpenAI’s GPTs, using GPT-3.5-Turbo and GPT-4-Turbo as foundational models.
• Evaluation includes ten question types across domains such as personal history, preferences, and current context to test authenticity and coherence.
• A control experiment assesses whether LLMs like GPT-4 and Gemini-Pro can distinguish between human and machine-generated responses, measuring bias and accuracy.
• All personal data is processed locally; no data is shared with third parties, and all responses are deleted after six months to ensure privacy.

Experimental results

Research questions

• RQ1 Can LLMs effectively impersonate ordinary individuals in role-playing scenarios, as measured by human evaluators’ ability to distinguish real from synthetic responses?
• RQ2 How does the performance of different role-playing methods (e.g., GPTs, RoleGPT, RPP) compare in deceiving human acquaintances of the target individual?
• RQ3 Can LLMs accurately identify whether a given text was written by a human or another LLM, and do they exhibit biases in their judgments?
• RQ4 What are the inherent biases in LLMs when evaluating human vs. machine-generated text, particularly in terms of model similarity or text length?
• RQ5 To what extent do LLMs fail to replicate subtle human nuances such as emotional depth, cultural context, and idiosyncratic communication patterns?

Key findings

• GPTs achieved the highest success rate of 48.3% in deceiving human evaluators, outperforming GPT-3.5 and GPT-4 across all role-playing methods.
• GPT-4 demonstrated superior role-playing capabilities compared to GPT-3.5, with a higher deception rate among the tested models.
• GPT-4 and GPT-4-Turbo both achieved over 90% success in distinguishing machine-generated from human-generated text, indicating strong detection ability.
• Despite strong detection performance, GPT-4 could not reliably determine which texts were human-produced, suggesting a bias toward identifying its own generation patterns.
• GPT-4 showed a significant instruction bias, with a 63.5% disparity in accuracy between settings, indicating it is more confident in identifying machine-generated content than human content.
• No significant length bias was observed in LLMs’ ability to differentiate between human and machine-generated text, as confirmed by comparison with a control model.