Identifying Features that Shape Perceived Consciousness in Large Language Model-based AI: A Quantitative Study of Human Responses

TL;DR

The paper investigates which textual features of AI responses lead humans to attribute perceived consciousness to large language models, using 99 passages from Claude 3 Opus and a survey of 123 participants. It combines eight feature ratings with regression analyses and cosine-based clustering to reveal that metacognitive self-reflection and emotional expressiveness boost perceived consciousness, while emphasis on knowledge reduces it, with seven distinct respondent clusters showing heterogeneous weighting. The study demonstrates that consciousness attribution is multidimensional and person-specific, further linked to prior LLM familiarity and chatbot use. These insights have implications for designing human-AI interactions and assessing psychosocial impacts of AI systems.

Abstract

This study quantitively examines which features of AI-generated text lead humans to perceive subjective consciousness in large language model (LLM)-based AI systems. Drawing on 99 passages from conversations with Claude 3 Opus and focusing on eight features -- metacognitive self-reflection, logical reasoning, empathy, emotionality, knowledge, fluency, unexpectedness, and subjective expressiveness -- we conducted a survey with 123 participants. Using regression and clustering analyses, we investigated how these features influence participants' perceptions of AI consciousness. The results reveal that metacognitive self-reflection and the AI's expression of its own emotions significantly increased perceived consciousness, while a heavy emphasis on knowledge reduced it. Participants clustered into seven subgroups, each showing distinct feature-weighting patterns. Additionally, higher prior knowledge of LLMs and more frequent usage of LLM-based chatbots were associated with greater overall likelihood assessments of AI consciousness. This study underscores the multidimensional and individualized nature of perceived AI consciousness and provides a foundation for better understanding the psychosocial implications of human-AI interaction.

Figures (3)

• Figure 1: LLM familiarity of survey respondents. (a) shows respondents’ total scores of prior knowledge of Large Language Models (LLMs), ranging from 0 to 18, with the y-axis indicating the number of respondents. The score was calculated by summing six items that assess prior knowledge of LLMs, with each item rated on a scale from 0 to 3. (b) illustrates LLM-based chatbot usage frequency, with categories from 1 (Never used) to 5 (Use daily) on the x-axis and the number of respondents on the y-axis.
• Figure 2: Significance and Directionality of Regression Coefficients by Feature. The scatter plot (a) depicts individual features derived from 82 significant regression models. The x-axis represents the regression coefficients, indicating both the magnitude and directionality (positive or negative) of effects, while the y-axis displays the statistical significance expressed as the negative logarithm of the adjusted P values. The horizontal dashed line at $-\ln(0.05)$ shows the threshold for statistical significance ($p$ < 0.05), with points above the line representing features that significantly contribute to perceived consciousness. (b) illustrates the number of models in which each feature's regression coefficient reached statistical significance. The y-axis indicates the count of significant occurrences across the 82 significant models. Colors within each bar differentiate between positive (blue) and negative (green) coefficients, highlighting the directionality of each feature's effect.
• Figure 3: Respondent Clustering by Combined Scores across Features. This clustermap visualizes the hierarchical clustering of respondents based on their combined score vectors across features. The dendrogram on the left shows the hierarchical clustering structure, with the dashed line marking the cutoff point used to define clusters. The color labels correspond to the seven distinct respondent clusters. Z-score normalization was applied to each respondent's combined score vector only for the visual clarification of the patterns across clusters.