[Paper Review] A Linguistic Comparison between Human and ChatGPT-Generated Conversations

Morgan Sandler, Hyesun Choung+2 more|arXiv (Cornell University)|Jan 29, 2024

Artificial Intelligence in Healthcare and Education10 citations

TL;DR

The paper analyzes linguistic differences between human and ChatGPT-3.5 conversations using LIWC across 118 categories on 19.5K ChatGPT dialogues versus EmpathicDialogues, finding humans more authentic while ChatGPT shows stronger social, cognitive, and positive tone features.

ABSTRACT

This study explores linguistic differences between human and LLM-generated dialogues, using 19.5K dialogues generated by ChatGPT-3.5 as a companion to the EmpathicDialogues dataset. The research employs Linguistic Inquiry and Word Count (LIWC) analysis, comparing ChatGPT-generated conversations with human conversations across 118 linguistic categories. Results show greater variability and authenticity in human dialogues, but ChatGPT excels in categories such as social processes, analytical style, cognition, attentional focus, and positive emotional tone, reinforcing recent findings of LLMs being "more human than human." However, no significant difference was found in positive or negative affect between ChatGPT and human dialogues. Classifier analysis of dialogue embeddings indicates implicit coding of the valence of affect despite no explicit mention of affect in the conversations. The research also contributes a novel, companion ChatGPT-generated dataset of conversations between two independent chatbots, which were designed to replicate a corpus of human conversations available for open access and used widely in AI research on language modeling. Our findings enhance understanding of ChatGPT's linguistic capabilities and inform ongoing efforts to distinguish between human and LLM-generated text, which is critical in detecting AI-generated fakes, misinformation, and disinformation.

Motivation & Objective

• Motivate understanding differences between human and LLM-generated dialogues to address authenticity and detection of AI-generated text.
• Leverage LIWC to profile linguistic features and compare variability and authenticity between humans and ChatGPT conversations.
• Provide a new ChatGPT-generated companion dataset (2GPTEmpathicDialogues) to aid NLP research.
• Investigate latent affect cues in embeddings despite no explicit affect mentions in dialogues.

Proposed method

• Use LIWC-22 to code 118 linguistic categories from 19.5K dialogues (humans versus two ChatGPT instances).
• Create 2GPTEmpathicDialogues by coordinating two ChatGPT-3.5-Turbo instances to imitate EmpathicDialogues scenarios.
• Apply independent samples t-tests with Bonferroni correction (p<.001) to compare category means and Levene’s test for variance differences.
• Train and evaluate valence classifiers (Random Forest, SVM, MLP) on OpenAI text-embedding-ada-002 embeddings with 5-fold cross-validation.
• Use UMAP to visualize embedding distributions by valence and compute Dunn Index for cluster separation.

Figure 1 : Framework for generation and prompts used in creating the 2GPTEmpathicDialogues dataset. In this setup, two instances of the ChatGPT-3.5-Turbo API engage in conversation with each other through a coordinating program.

Experimental results

Research questions

• RQ1 Do ChatGPT-generated dialogues differ from human dialogues across LIWC categories in terms of variability, authenticity, social behavior, cognition, and emotion?
• RQ2 Can embedding-based valence classification detect latent affect cues in ChatGPT versus human conversations, despite no explicit affect mentions?
• RQ3 Does a ChatGPT-generated companion dataset (2GPTEmpathicDialogues) closely mirror the EmpathicDialogues human corpus for linguistic analysis?
• RQ4 What are the implications of linguistic differences for AI-text detection and misinformation risk?

Key findings

• Humans show greater variability and authenticity than ChatGPT in LIWC categories.
• ChatGPT shows higher levels in social processes, prosocial behavior, politeness, communication, attentional focus, analytical thinking, cognition, and positive emotional tone.
• No significant difference between ChatGPT and humans in overall positive or negative affect.
• ChatGPT embeddings exhibit latent valence cues, with classifiers achieving high F1-scores (SVM 90.0% on humans, 95.3% on ChatGPT).
• UMAP reveals clearer valence clustering in ChatGPT embeddings (Dunn Index 0.222) than in humans (0.153).
• Top misclassified emotions in valence classification include anxious, surprised, trusting, caring, sentimental, hopeful, and others across both datasets.

(a) Human Dialogues