"Who wants to be nagged by AI?": Investigating the Effects of Agreeableness on Older Adults' Perception of LLM-Based Voice Assistants' Explanations

This study of 70 older adults reveals that while high-agreeableness in LLM-based voice assistants generally enhances trust and likability, the preference for warmth over clarity shifts depending on the context (routine vs. emergency) and the user's own personality, highlighting the need for adaptive, personalized AI explanations.

The Gist

Imagine you have a helpful robot butler named Robin living in your home. Robin's job is to remind you to take your medicine or warn you if a smoke detector goes off. But here's the twist: Robin can talk to you in two very different "personalities."

• The "Warm & Fuzzy" Robin: This version is super polite, kind, and empathetic. It says things like, "Oh, hello! Just a gentle reminder that it's time for your vitamins. You're doing great!"
• The "Blunt & Bossy" Robin: This version is direct, efficient, and a bit cold. It says things like, "It is 8:00 AM. Take your vitamins now. Do not forget."

This paper is a study about older adults and how they feel when these two different Robins explain why they are giving a reminder or an alert. The researchers wanted to know: Does being nice actually make the robot seem smarter and more trustworthy?

Here is the breakdown of their findings, using some simple analogies:

1. The "Nice Guy" Effect (Routine Tasks)

When the task was something low-stakes, like a daily reminder to drink water, the Warm Robin won hands down.

• The Analogy: Think of it like a friendly neighbor vs. a strict traffic cop. If your neighbor gently reminds you to take out the trash, you like them, you trust them, and you're happy to listen.
• The Result: Older adults liked the Warm Robin more, trusted it more, and felt it was more empathetic. They were also more willing to "adopt" (keep using) this version of the robot.

2. The "Emergency" Exception (High-Stakes Tasks)

However, the study found a catch. When the situation was an emergency (like a fire alarm or a fall detection), the "Warm and Fuzzy" personality didn't help as much.

• The Analogy: If your house is on fire, you don't want a therapist who says, "Oh no, I'm so sorry you're having a scary experience, let's talk about your feelings." You want a drill sergeant who yells, "RUN! NOW!"
• The Result: In emergencies, people cared more about clarity than kindness. The "Warm" Robin didn't lose points for being nice, but the "Bossy" Robin didn't get penalized for being blunt either. In a crisis, "Get to the point" is the best personality.

3. The "Smart vs. Nice" Myth

A really interesting finding was that being nice did not make the robot seem smarter.

• The Analogy: Imagine a teacher who is very kind and hugs you, but gives you a C on your test. You like the teacher, but you don't suddenly think they are a genius. Conversely, a strict teacher who gives you an A is seen as smart, even if they are grumpy.
• The Result: The study showed that people separated "personality" from "intelligence." The Warm Robin was seen as likable and empathetic, but not necessarily smarter than the Blunt Robin. You can design a robot to be warm without tricking people into thinking it's a genius.

4. The "Real-Time" Truth

The study also looked at how the robot explained its actions.

• The Analogy: Imagine you ask, "Why did you turn on the sprinklers?"
  • Type A (History): "Well, you asked me to water the garden last Tuesday, so I'm doing it now." (This feels a bit like a memory lapse).
  • Type B (Real-Time): "I turned on the sprinklers because the soil sensor just told me the dirt is bone dry." (This feels like a live camera feed).
• The Result: People trusted the Real-Time explanation much more. It felt more grounded in the "now" and less like the robot was just guessing based on old conversations.

5. The "Mirror Effect" (Your Own Personality)

Finally, the study looked at the users' own personalities.

• The Analogy: If you are a very polite and agreeable person yourself, you hate it when someone is rude to you. You might think, "I'm a nice person, so why is this robot being so mean?"
• The Result: Older adults who were naturally very "agreeable" (kind/polite) were the harshest critics of the Blunt Robin. They penalized the rude robot much more than the other users did.

The Big Takeaway

The main lesson from this paper is that there is no "one-size-fits-all" robot personality.

If you are building an AI assistant for older adults:

1. Be warm and polite for daily reminders (it builds trust).
2. Be clear and direct for emergencies (safety comes first).
3. Use real-time data to explain your actions (it builds credibility).
4. Know your audience: If your user is a very nice person, don't be rude to them, or they will turn you off.

In short: Don't just be a robot; be the right kind of robot for the moment.

Technical Summary

Here is a detailed technical summary of the paper "Who wants to be nagged by AI?": Investigating the Effects of Agreeableness on Older Adults' Perception of LLM-Based Voice Assistants' Explanations.

1. Problem Statement

As Large Language Model (LLM)-based Voice Assistants (VAs) become integral to "aging in place" (supporting older adults in their homes), the way these systems explain their actions is critical for building trust. However, current research in Explainable AI (XAI) focuses heavily on technical model logic, often neglecting the conversational personality of the AI.

• The Gap: While LLM-VAs are often designed with high "agreeableness" (politeness, warmth) to increase engagement, it is unclear how this personality trait affects older adults' perception of explanations in different contexts (routine vs. emergency).
• The Risk: High agreeableness might reduce critical scrutiny or lead to overreliance, while low agreeableness might be perceived as "nagging" or judgmental.
• Research Questions (RQs):
  • RQ1: How does an LLM-VA's agreeableness influence older adults' perception of its explanations?
  • RQ2: How do situational context (routine vs. emergency) and explanation type (conversational history vs. real-time environmental data) moderate these perceptions?
  • RQ3: How does the older adult's own level of agreeableness influence their perception of the VA?

2. Methodology

The authors conducted a controlled experimental study with 70 older adults (ages 60–83, $M=65.97$) recruited via Prolific.

Experimental Design

The study utilized a 2 $\times$ 2 $\times$ 2 mixed factorial design:

1. Between-Subjects Factor: LLM-VA Personality (High Agreeableness vs. Low Agreeableness).
   • Manipulation: The fictional VA "Robin" was prompted using the Trait Modulation Key (TMK) framework to generate responses that were either warm/cooperative (High) or blunt/direct (Low).
2. Within-Subjects Factors:
   • Interaction Context: Routine (low-risk reminders) vs. Emergency (high-risk safety alerts).
   • Explanation Type:
     • Conversational History (UH): Based on prior user interactions.
     • Real-time Environmental (ENV): Based on immediate sensor data.

Procedure & Interface

• Stimuli: Participants interacted with interactive storyboards (web-based UI) featuring visual depictions of the VA and audio clips of the VA's voice.
• Flow: Participants heard a reminder/alert, followed by an explanation. They then rated the VA on seven validated metrics.
• Measures:
  • Dependent Variables: Intention to adopt, Perceived Empathy, Likeability, Trustworthiness, Reliance, Explanation Satisfaction, and Perceived Intelligence.
  • Independent Variable (User): Participants' own agreeableness was measured using the Mini-IPIP scale.

3. Key Results

RQ1: Effect of VA Agreeableness

• Positive Impact: High-agreeableness (HA) VAs were rated significantly higher in intention to adopt, empathy, likeability, trust, reliance, and satisfaction compared to Low-agreeableness (LA) VAs.
• No Impact on Competence: Crucially, agreeableness did not affect perceived intelligence. This suggests that social tone (warmth) and competence are separable dimensions in the user's mind.
• Qualitative Insight: Users described LA explanations as "abrupt," "judgmental," or "nagging," whereas HA explanations were seen as "friendly" and "thoughtful."

RQ2: Context and Explanation Type

• Context Sensitivity: The benefits of high agreeableness were stronger in routine contexts than in emergency contexts. In emergencies, clarity appeared to outweigh warmth; the "HA advantage" diminished significantly.
• Explanation Type: Real-time Environmental (ENV) explanations were generally preferred over History-based (UH) explanations, resulting in higher reliance, satisfaction, and perceived intelligence.
• Interaction: The HA VA's advantage over the LA VA was most pronounced in routine scenarios using history-based explanations.

RQ3: User Agreeableness

• Personality Matching: There was a significant negative correlation between the user's agreeableness and their ratings of the LA VA.
• Finding: More agreeable older adults penalized the low-agreeableness assistant more severely than less agreeable users did. This indicates that personality mismatch is particularly detrimental for agreeable users.

4. Key Contributions

1. Decoupling Warmth and Competence: The study provides empirical evidence that an AI can be perceived as "likable" and "trustworthy" without being perceived as "smarter," allowing designers to modulate personality independently of capability signaling.
2. Context-Aware Explainability: It challenges the "one-size-fits-all" approach to AI explanations. High agreeableness is beneficial for routine tasks but may be less critical (or even distracting) in high-stakes emergency scenarios where clarity is paramount.
3. User-System Personality Matching: It highlights that the user's own personality traits (specifically agreeableness) act as a filter; agreeable users are less tolerant of blunt AI behavior.
4. Grounding Preference: Real-time environmental data (sensor-based) was found to be more credible and trustworthy to older adults than explanations based solely on conversational history.

5. Significance and Design Implications

The paper proposes five Design Considerations (DCs) for LLM-based VAs in aging-in-place settings:

• DC1 (Contextual Sensitivity): Explanations must adapt to the situation (routine vs. emergency) rather than using a fixed conversational style.
• DC2 (Task-Dependent Personality): In emergencies, prioritize clarity over warmth.
• DC3 (Separable Dimensions): Designers can increase likeability via agreeableness without compromising perceived intelligence.
• DC4 (Data Grounding): Prioritize real-time environmental data over conversational history for explanations to enhance trust.
• DC5 (Audience Awareness): Recognize that agreeable users are more sensitive to tone; a blunt AI may alienate this specific demographic.

Conclusion

The study concludes that effective AI explanation design for older adults is a joint problem of content, delivery, and audience. While high agreeableness generally improves trust and adoption, it must be balanced against the urgency of the task and the personality of the user. Future work should explore longitudinal effects and additional personality traits beyond agreeableness.