Enterprise AI Analysis

This section details how participants attributed various mental capacities (cognitive, emotional, physiological) to LLMs under different messaging conditions.

Through exploratory factor analysis, 40 mental capacities were grouped into 'emotional', 'cognitive', and 'physiological' categories. Regardless of the video, participants generally attributed more 'cognitive' capacities than 'emotional' or 'physiological' ones to LLMs, aligning with prior research.

The 'companion' message uniquely boosted attributions of both emotional and cognitive capacities, but not physiological ones. This nuanced effect suggests that companion framing encourages a more holistic, human-like perception of AI's internal states, beyond mere physical presence.

Examines how different messages about LLMs affect user reliance on their outputs, particularly in factual information retrieval tasks, considering correctness and consistency.

Participants' reliance on LLM responses was primarily sensitive to the logical consistency of the output, rather than its correctness. This finding, consistent with previous work, suggests users prioritize internal coherence over external factual accuracy when assessing AI outputs.

Crucially, the 'LLMs as machines' message significantly reduced reliance on inconsistent LLM responses. This indicates that framing AI as a machine encourages a more critical evaluation of its outputs, leading to increased vigilance when the system contradicts itself.

These findings suggest that messaging can be a powerful tool to shape appropriate reliance behaviors, encouraging users to be more discerning when interacting with LLMs, especially concerning internal consistency. This is vital for developing safer and more trustworthy AI applications.

Discusses the broader implications of anthropomorphism, communication strategies for AI literacy, and outlines areas for future research.

Presenting LLMs as companions can be seen as a subtle form of anthropomorphism, attributing human-like emotional and cognitive qualities. While this can make AI more approachable, it raises concerns about miscalibrated expectations, over-reliance, and emotional dependency.

The study highlights the importance of thoughtful AI communication. Short, educational messages, like those framing LLMs as machines or tools, can promote cautious behavior and increase AI literacy by emphasizing mechanistic understanding rather than human-like attributes.

Future work should explore these effects in real-world settings with diverse messages and tasks, potentially using formal mediation analyses to uncover the underlying psychological mechanisms. This will help guide the development of safe and trustworthy LLM-based applications.