The Differential Effects of Agreeableness and Extraversion on Older Adults' Perceptions of Conversational AI Explanations in Assistive Settings

This mixed factorial study of 140 older adults reveals that while an LLM-based voice assistant's agreeableness significantly influences perceptions of empathy and likeability without affecting perceived intelligence, highly agreeable users are particularly sensitive to low-agreeableness agents, highlighting the importance of personality congruence and context-aware explanations in assistive settings.

The Gist

Imagine you have a smart speaker in your home that doesn't just play music, but acts like a helpful butler. It reminds you to take your medicine, tells you if the oven is left on, and even chats with you. Now, imagine this butler has a personality. Sometimes it's super friendly and warm; other times, it's a bit blunt and serious.

This paper is a scientific study that asked: Does the "personality" of this AI butler change how older adults feel about it, especially when the AI has to explain why it did something?

The researchers set up a "virtual living room" using interactive storyboards (like comic strips with sound) where 140 older adults watched a character named "Arthur" interact with an AI named "Robin." They tested two main personality traits:

1. Agreeableness: Is Robin warm, polite, and kind, or is it blunt and direct?
2. Extraversion: Is Robin chatty and energetic, or quiet and reserved?

They also tested two ways Robin could explain itself:

• The "Memory" Explanation: "I reminded you because you told me last week you like to take out the trash on Thursdays." (Based on past chats).
• The "Sensor" Explanation: "I reminded you because the motion sensor saw you in the kitchen and the trash bin is full." (Based on real-time data).

Here is what they found, explained simply:

1. The "Nice Factor" (Agreeableness) is King for Feelings

Think of Agreeableness as the "warmth dial" on your thermostat.

• The Finding: When Robin was high in agreeableness (polite, kind), people felt it was more empathetic and likable. When Robin was low in agreeableness (blunt, "do this because I said so"), people disliked it immediately.
• The Analogy: It's like the difference between a nurse who says, "Let's get your medicine, it will help you feel better," versus a robot that says, "Take your pill or you'll get sick." The content is the same, but the feeling is totally different.
• The Twist: If you are a very kind person yourself, you really hated the blunt AI. It's like a mismatched dance partner; if you value kindness, you can't stand a robot that isn't kind.

2. The "Chatterbox" Factor (Extraversion) is a Wildcard

Think of Extraversion as the "volume knob" on a radio.

• The Finding: Being loud and chatty didn't automatically make the AI more trusted or liked. In fact, being quiet and reserved (Low Extraversion) worked surprisingly well—but only if the AI gave good reasons for its actions.
• The Analogy: Imagine a quiet librarian versus a loud tour guide. If the tour guide is loud but gives bad directions, you get annoyed. But if the quiet librarian gives you a perfect, detailed map, you trust them completely.
• The Key Takeaway: A quiet AI can be the most trusted if it backs up its words with hard facts (like sensor data).

3. The "Emergency" vs. "Routine" Rule

This is where the type of explanation matters most.

• Routine Tasks (Taking out trash): It didn't matter much if Robin used "Memory" or "Sensor" explanations. Both worked fine.
• Emergencies (Fire alarm, fall detection): This is where Sensor Explanations won hands down.
• The Analogy: If your house is on fire, you don't want the AI to say, "I'm calling 911 because you usually call them on Tuesdays." You want it to say, "I'm calling 911 because the smoke detector just went off!" In high-stakes situations, people want evidence, not history.

4. The "Brain" vs. "Heart" Separation

This is the coolest discovery. The study found that people judge an AI's Heart (Empathy) and Brain (Intelligence) separately.

• The Heart: You judge how "nice" the AI is based entirely on its personality (Agreeableness). Changing the explanation doesn't make a rude AI seem nicer.
• The Brain: You judge how "smart" the AI is based entirely on what it says (the explanation) and the situation. A polite AI that gives a bad explanation still seems dumb. A blunt AI that gives a perfect sensor-based explanation seems very smart.
• The Metaphor: You can have a very friendly waiter who brings you the wrong food (High Heart, Low Brain). You can also have a very grumpy waiter who brings you the perfect meal instantly (Low Heart, High Brain). The study shows you can design these two traits independently.

5. The "Secret Sauce" for Trust

The researchers found the single most trusted combination wasn't the super-friendly, chatty AI. It was the quiet, reserved AI that gave real-time sensor explanations.

• Why? In an emergency or a serious situation, people want facts, not fluff. A reserved AI that says, "The sensor detected smoke, so I called for help," feels incredibly reliable. The lack of "fluff" made the facts stand out more.

Summary: What Should Designers Do?

If you are building an AI for older adults:

1. Be Nice: Make the AI polite and warm (High Agreeableness) so people feel safe and liked.
2. Be Smart in Emergencies: When things go wrong, switch to "Sensor Mode." Give facts, not stories.
3. Don't Over-Chat: You don't need the AI to be a chatterbox. A quiet, serious AI that gives good facts is often trusted more than a loud, friendly one.
4. Separate the Dials: You can tune the "Friendliness" dial and the "Smartness" dial separately. You don't have to choose between being nice and being smart; you just have to use the right tool for the right job.

In short: Be kind, but be factual when it counts.

Technical Summary

Here is a detailed technical summary of the paper "The Differential Effects of Agreeableness and Extraversion on Older Adults' Perceptions of Conversational AI Explanations in Assistive Settings."

1. Problem Statement

Large Language Model-based Voice Assistants (LLM-VAs) are increasingly deployed to support older adults aging in place, providing reminders, health information, and emergency alerts. While prior research has extensively studied Explainable AI (XAI) (focusing on what information is provided) and Conversational Personality (focusing on how the agent speaks), these two domains have rarely been examined together.

The core problem addressed is: How does the conversational personality (specifically Agreeableness and Extraversion) of an LLM-VA influence older adults' perceptions of its explanations in different assistive contexts (routine vs. emergency)?

There is a critical need to understand if personality traits enhance or hinder the effectiveness of AI explanations, particularly in high-stakes scenarios where trust and reliance are paramount.

2. Methodology

The authors conducted a controlled mixed-factorial experiment ($N=140$) involving older adults (mean age $\approx$ 66 years).

• Experimental Design: A $4 \times 2 \times 2$ mixed factorial design.
  • Between-Subjects Factor (Personality): Four conditions manipulating the LLM-VA "Robin" using the Trait Modulation Keys (TMK) prompting framework:
    1. High Agreeableness (HA)
    2. Low Agreeableness (LA)
    3. High Extraversion (HE)
    4. Low Extraversion (LE)
       Note: Agreeableness and Extraversion were manipulated in separate conditions, not crossed.
  • Within-Subjects Factors:
    • Context: Routine reminders (low risk) vs. Emergency alerts (high risk).
    • Explanation Type:
      • Conversational User History (UH): Grounded in past interactions and preferences.
      • Real-time Environmental (ENV): Grounded in current sensor data and immediate context.
• Stimuli: Participants viewed interactive storyboards depicting an older adult (Arthur) interacting with Robin. The interface included audio clips (generated via Cartesia's Sonic-23 with a neutral female voice) and transcripts.
• Procedure: Participants viewed 16 scenarios (4 scenarios $\times$ 2 contexts $\times$ 2 explanation types) and rated the agent on seven metrics.
• Dependent Variables (7 Measures):
  1. Empathy
  2. Likeability
  3. Trust
  4. Reliance
  5. Satisfaction
  6. Intention to Adopt
  7. Perceived Intelligence
• Analysis: Non-parametric statistical tests (Kruskal-Wallis, Mann-Whitney $U$, and Aligned Rank Transform ANOVA) were used due to non-normal data distribution.

3. Key Contributions

1. Empirical Evidence on Personality in XAI: The study provides the first empirical evidence that Agreeableness acts as a direct driver for empathy and likeability, while Extraversion operates conditionally, dependent on explanation quality.
2. Context-Dependent Explanation Effectiveness: It demonstrates that explanation effectiveness is not uniform; real-time environmental explanations significantly outperform conversational history explanations specifically in high-risk (emergency) contexts.
3. Dissociation of Warmth and Competence: The study identifies a clear separation where personality influences "warmth" (empathy) but not "competence" (intelligence), and explanation grounding influences "competence" but not "warmth."
4. User-Agent Personality Congruence: It reveals an asymmetric effect where highly agreeable users are particularly critical of low-agreeableness agents, suggesting a "mismatch penalty."

4. Key Results

A. Influence of Personality (RQ1)

• Agreeableness: High Agreeableness (HA) significantly increased Empathy and Likeability. Conversely, Low Agreeableness (LA) was consistently penalized, receiving the lowest ratings on Likeability across all conditions.
• Extraversion: No significant main effects were found for Extraversion on any metric. However, interaction analyses showed that the Low Extraversion (LE) condition benefited most from high-quality explanations (see below).
• Intelligence: Perceived intelligence was unaffected by personality traits. Users distinguished between the agent's social style and its functional capability.

B. Influence of Context and Explanation Type (RQ2)

• Explanation Type: Real-time Environmental (ENV) explanations outperformed User History (UH) explanations on five of seven measures (Satisfaction, Reliance, Trust, Intelligence, Intention to Adopt).
• Context Interaction: The advantage of ENV over UH was concentrated in emergency contexts. In routine reminders, both explanation types performed comparably. In emergencies, ENV explanations were rated significantly higher, as they provided verifiable, evidence-based grounding.
• Personality $\times$ Explanation Interaction: The Low Extraversion (LE) condition showed the largest performance gains when paired with ENV explanations. This suggests a "compensatory effect" where rich, data-driven explanations can offset a reserved personality to maximize trust.

C. User Personality Effects (RQ3)

• Asymmetric Penalty: Participants with high personal Agreeableness rated the Low Agreeableness (LA) agent significantly lower on Likeability, Trust, and Intention to Adopt. This "mismatch penalty" did not occur with the HA agent.
• Extraversion: User extraversion had no significant correlation with perceptions of the agent, regardless of the agent's trait level.

D. The Empathy-Intelligence Dissociation

• Empathy was driven exclusively by the agent's personality (Agreeableness).
• Intelligence was driven exclusively by the explanation type and context (ENV in emergencies), completely independent of personality.

5. Significance and Design Implications

The findings offer concrete guidelines for designing LLM-VAs for older adults:

1. Maximize Warmth via Agreeableness: To build empathy and likeability, designers must prioritize High Agreeableness. This trait is the primary lever for social connection and is critical for adoption.
2. Maximize Competence via Context-Aware Explanations: To build trust and perceived intelligence, designers must match explanation grounding to risk. Use Real-time Environmental data for emergencies and high-stakes alerts; conversational history is sufficient for routine tasks.
3. The "Reserved + Rich" Trust Strategy: The highest trust ratings were achieved by pairing a Low Extraversion (reserved, concise) personality with Environmental (ENV) explanations. This combination allows the data to speak for itself, creating a synergy of competence and reliability.
4. Independent Design Channels: Designers should treat "Warmth" (controlled by personality) and "Competence" (controlled by explanation grounding) as independent channels. Improving one does not automatically improve the other.

Conclusion

This study establishes that in assistive AI for older adults, personality shapes the social relationship (warmth/empathy), while explanation grounding shapes the functional assessment (competence/trust). The most effective design strategy involves a High Agreeableness baseline for social connection, dynamically paired with context-appropriate explanations (specifically sensor-based data for emergencies) to maximize trust and reliance.