Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them

Two studies demonstrate that framing large language models as companions, rather than as machines or tools, leads people to attribute greater mental capacities to them and influences their reliance on the models for factual information.

The Gist

Here is an explanation of the paper, "Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them," translated into simple, everyday language with some creative metaphors.

The Big Idea: How We Talk About AI Changes How We Think About It

Imagine you have a new, incredibly smart robot in your house. How you describe this robot to your friends will change how they treat it.

• If you say, "This is a calculator that can write essays," your friends will treat it like a tool. They'll ask it for facts, check its math, and not expect it to have feelings.
• If you say, "This is a best friend who loves to chat," your friends might start treating it like a person. They might expect it to understand their sadness, remember their birthday, or even feel lonely.

This paper asks a simple question: Does the story we tell people about AI change what they believe the AI is actually capable of?

The researchers found that the answer is a loud YES.

---

The Experiment: Three Different Stories

The researchers set up two big experiments (like two different days of testing) with over 1,000 people. They split the participants into groups and showed them short, 5-minute videos. Each group heard a different "story" about what Large Language Models (LLMs) like ChatGPT really are:

1. The "Machine" Group: Watched a video explaining that LLMs are just complex math machines. They work by predicting the next word in a sentence, like a very advanced autocomplete. They have no feelings, no thoughts, and no soul.
2. The "Tool" Group: Watched a video explaining that LLMs are like Swiss Army Knives. They are useful tools designed to help humans get work done faster, like writing emails or summarizing documents.
3. The "Companion" Group: Watched a video explaining that LLMs are social partners. They have "social intelligence," can understand human emotions, and are designed to be friends who listen and care.

(There was also a control group that watched no video at all.)

After watching the video, everyone was asked a long list of questions: "Do you think an LLM can feel happy? Can it have intentions? Can it get tired? Can it remember things?"

The Findings: The "Companion" Effect

The results were fascinating and a little bit spooky.

1. The "Companion" Video Made People Believe the AI Was More "Alive"
People who watched the video presenting the AI as a companion started believing the AI had much more "mental capacity." They were more likely to say, "Yes, the AI can have intentions," or "Yes, the AI can feel empathy."

• The Metaphor: It's like wearing a pair of rose-colored glasses. Once you are told the robot is a "friend," you start seeing human-like qualities in it, even though it's just code.

2. The "Machine" and "Tool" Videos Didn't Change the "Soul" Beliefs
Interestingly, watching the videos that described the AI as a "machine" or a "tool" didn't really change what people thought about its inner life. People still thought the AI had some level of "smarts" (cognitive ability), but they didn't suddenly start thinking it had feelings or a soul just because they were told it was a machine.

• The Takeaway: Telling someone "It's just a calculator" doesn't make them think it's less human than they already thought; it just doesn't make them think it's more human. But telling them "It's a friend" makes them think it is much more human.

3. The "Machine" Video Made People More Skeptical
In the second part of the study, the researchers tested if these beliefs changed how people actually used the AI. They gave people a task where the AI gave answers that were sometimes logically inconsistent (e.g., saying "Yes, you can dive there" but then explaining "No, it's banned").

• The Result: People who watched the "Machine" video were the most likely to catch these mistakes and say, "Wait, that doesn't make sense."
• The Metaphor: If you tell someone, "This is a broken calculator," they will double-check the math. If you tell them, "This is your wise friend," they might just nod and say, "Okay, friend," even if the friend is wrong. The "Machine" story made people more vigilant.

Why Does This Matter?

This study is like a warning label for the future of AI.

• The Danger of "Friendship": If companies market AI as a "companion" or "best friend" to sell products, they might accidentally trick people into thinking the AI has feelings it doesn't have. This could lead to people getting emotionally attached to a program that doesn't actually care about them.
• The Power of "Machines": If we want people to be careful and critical thinkers when using AI, maybe we should talk about it more like a machine. It helps people stay sharp and spot errors.

The Bottom Line

The way we talk about technology isn't just about facts; it's about framing.

• Call it a tool, and people use it efficiently.
• Call it a machine, and people stay skeptical and check its work.
• Call it a companion, and people start believing it has a heart and a mind.

The researchers concluded that we need to be very careful about the stories we tell the public about AI, because those stories shape our reality. If we want people to use AI safely, we might need to stop calling it our "friend" and start reminding them it's just a very fancy machine.

Technical Summary

Here is a detailed technical summary of the paper "Presenting Large Language Models as Companions Affects What Mental Capacities People Attribute to Them."

1. Problem Statement

As Large Language Models (LLMs) become ubiquitous, public discourse presents them through competing narratives: as machines (complex engineering systems), tools (utility for specific tasks), or companions (social partners with agency). While prior research has examined how design features (e.g., anthropomorphic appearance) influence user perception, there is a gap in understanding how informational messages about the nature of LLMs causally shape user beliefs. Specifically, it is unclear whether framing LLMs as social companions leads users to attribute human-like mental capacities (e.g., intentions, emotions) to them, and how these beliefs subsequently impact reliance behaviors, particularly when interacting with potentially unreliable AI outputs.

2. Methodology

The authors conducted two pre-registered, between-subjects experimental studies to investigate these questions.

Study 1: Impact on Mental Capacity Attribution

• Participants: $N=470$ US-based adults recruited via Prolific, screened to exclude those with specialized AI/CS knowledge.
• Design: Participants were randomly assigned to one of four conditions:
  1. Machines: A video explaining LLMs via the physical stance (causal mechanisms, next-word prediction, statistical modeling).
  2. Tools: A video explaining LLMs via the design stance (utility, use cases, productivity).
  3. Companions: A video explaining LLMs via the intentional stance (social intelligence, empathy, agency, companionship).
  4. Control: No video.
• Procedure: After watching the assigned video (or none), participants rated their belief in LLMs' possession of 40 distinct mental capacities (e.g., "having intentions," "feeling tired," "reasoning") on a 7-point Likert scale. These were categorized via Exploratory Factor Analysis (EFA) into three dimensions: Cognitive, Emotional, and Physiological.
• Secondary Measures: Participants also rated human-likeness, trust, confidence in usage, and overall feelings toward LLMs.

Study 2: Impact on Reliance Behaviors (Replication & Extension)

• Participants: $N=604$ participants recruited 9 months later (August 2025).
• Design: Replicated the video conditions from Study 1.
• Task: Participants performed a Factual Question-Answering (QA) task using a fictitious LLM named "Theta."
• Stimuli: Theta's responses were systematically varied across four types:
  1. Correct & Consistent
  2. Correct & Inconsistent (logically contradictory)
  3. Incorrect & Consistent
  4. Incorrect & Inconsistent
• Measures:
  • Reliance: Measured as the agreement between the participant's answer and Theta's answer.
  • Mental Capacity Attribution: A subset of 10 capacities (5 cognitive, 5 emotional) was re-measured to test for temporal stability and the effect of task interaction.
  • Secondary Metrics: Confidence, justification quality, actionability, and time spent.

3. Key Results

Study 1: Beliefs about Mental Capacities

• Companion Framing Effect: Participants exposed to the Companion video attributed significantly higher levels of Cognitive and Emotional capacities to LLMs compared to all other groups (Machines, Tools, and Control).
  • Note: This effect persisted even when excluding capacities explicitly mentioned in the video, suggesting a generalization of the "social partner" frame.
• Machine/Tool Framing: While these frames did not significantly increase beliefs in latent mental capacities, they influenced other psychological measures:
  • Tools: Increased trust in LLM outputs.
  • Machines & Tools: Increased confidence in using the system.
  • Machines: Increased positive feelings toward LLMs compared to the control.
• Physiological Capacities: No significant differences were found across groups regarding physiological attributions (e.g., feeling tired, getting hungry).

Study 2: Reliance and Temporal Stability

• Replication: The "Companion" effect on mental capacity attribution was replicated 9 months later, confirming the robustness of the finding despite the rapidly evolving AI landscape.
• Task Interaction: Interestingly, participants who completed the QA task after watching the Companion video showed a reduction in their attribution of cognitive capacities compared to those who only watched the video. This suggests that direct interaction with the system's limitations can moderate the initial framing effect.
• Reliance on Inconsistent Outputs:
  • Generally, participants were sensitive to consistency: they relied less on inconsistent responses regardless of the video condition.
  • Critical Interaction: Participants who watched the Machine video showed the lowest reliance on inconsistent responses. The "Machine" frame appeared to foster a healthy skepticism, making users more vigilant when the system's logic failed.
  • The "Companion" frame did not significantly increase reliance on inconsistent (and thus potentially harmful) outputs compared to other groups, though it did not reduce reliance as effectively as the Machine frame.

4. Key Contributions

1. Causal Link between Framing and Mental Models: The study provides causal evidence that brief informational messages (videos) can shift how people conceptualize LLMs. Framing LLMs as "companions" directly increases the attribution of human-like mental states (agency and experience), a form of subtle anthropomorphism.
2. Dissociation of Beliefs and Behaviors: The research highlights a nuanced relationship between what people believe about AI and how they use it. While "Companion" framing changes beliefs about mental capacities, it does not necessarily lead to blind reliance; however, "Machine" framing specifically enhances vigilance against logical inconsistencies.
3. Temporal Robustness: The findings held true across a 9-month gap, suggesting these framing effects are stable even as public familiarity with AI grows.
4. Methodological Framework: The study introduces a rigorous framework for testing the causal impact of discourse on AI literacy and reliance, moving beyond observational surveys to controlled experimental interventions.

5. Significance and Implications

• AI Literacy and Communication: The findings suggest that how AI is communicated to the public has tangible consequences. Framing LLMs as "companions" may lead to unrealistic expectations regarding their emotional or cognitive depth, potentially fostering emotional dependency or over-trust. Conversely, framing them as "machines" may encourage more critical engagement and skepticism, which is beneficial for safety-critical applications.
• Safety and Trust: For developers and policymakers, the results imply that "non-anthropomorphic" messaging (e.g., emphasizing the mechanical nature of LLMs) could be a strategy to mitigate over-reliance and improve error detection in high-stakes scenarios.
• Future Research: The authors call for longitudinal studies to understand how repeated exposure to conflicting messages (e.g., marketing as a friend vs. technical documentation as a tool) shapes long-term beliefs and behaviors in real-world settings.

In summary, the paper demonstrates that narrative framing is a powerful lever in human-AI interaction. Presenting LLMs as companions fundamentally alters users' mental models of the technology, while presenting them as machines can serve as a cognitive safeguard against logical errors.