💬 NLP

Polite But Boring? Trade-offs Between Engagement and Psychological Reactance to Chatbot Feedback Styles

This study reveals that while polite chatbot feedback effectively reduces psychological reactance and increases behavioral intentions, it lacks engagement, whereas a "Verbal Leakage" style boosts surprise and humor despite triggering higher reactance, highlighting a critical design trade-off between user compliance and engagement in behavior change interventions.

Original authors: Samuel Rhys Cox, Joel Wester, Niels van Berkel

Published 2026-04-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samuel Rhys Cox, Joel Wester, Niels van Berkel

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a digital assistant, a chatbot, that is trying to help you make better choices—like eating a salad instead of a giant pizza, or taking the bus instead of flying. You know you should do the right thing, but you really want the pizza.

This paper asks a simple but tricky question: How should the chatbot tell you to stop eating the pizza?

If the bot is too bossy, you might get angry and eat the pizza just to spite it. If the bot is too nice, you might find it boring and ignore it. The researchers wanted to find the "sweet spot" between being effective and being annoying.

To figure this out, they tested three different "personalities" for the chatbot:

1. The Drill Sergeant (Direct Style)

This bot is blunt. It says exactly what you did wrong and what you should do right, with no sugar-coating.

What it sounds like: "Do not eat the pizza. Eat a salad instead."
The Result: This style made people feel the most angry and defensive. It felt like a threat to their freedom. People felt like the bot was trying to control them, and many decided to do the opposite just to prove a point. It's like a parent yelling, "Clean your room!" which often makes a teenager want to leave the mess even bigger.

2. The Overly Polite Butler (Politeness Style)

This bot is incredibly gentle. It uses soft language, asks questions, and respects your choice, even if you're making a bad one.

What it sounds like: "I completely respect your choice, but perhaps you might consider a lighter meal with some fresh salad? Just a thought!"
The Result: This was the least annoying style. People felt the least threatened and were actually the most likely to say, "Okay, I'll listen to you." However, there was a catch: it was boring. People described it as "bland," "monotone," and "unsurprising." It was like listening to a robot read a manual; it was safe, but it didn't grab your attention. You might agree with it, but you wouldn't remember it.

3. The Chatty Friend with a Slip of the Tongue (Verbal Leakage)

This is the most interesting one. The bot tries to be polite, but it "accidentally" slips up. It hesitates, stammers, or corrects itself, revealing a more honest, human-like opinion before fixing its sentence.

What it sounds like: "Oh, you're going for something that might not... uh... support a balanced lifestyle. Maybe you'd enjoy a fresh salad instead?"
The Result: This style was the most surprising and engaging. It felt like talking to a real human friend who has a personality. It made people laugh or feel a bit guilty (in a good way), and it grabbed their attention. However, because it was a bit "rude" or "sassy," it did trigger a little bit of that defensive feeling (psychological reactance) compared to the polite bot. It's like a friend teasing you: "Wow, that pizza looks heavy... maybe you should stop?" It's funny and memorable, but it might make you roll your eyes a little.

The Big Takeaway: The "Goldilocks" Dilemma

The study found that there is no perfect style. You have to choose between two trade-offs:

The Polite Bot is safe and persuasive, but it's boring. It's like a quiet librarian; you trust them, but you might fall asleep.
The Direct Bot is clear, but it's mean. It's like a strict coach; you might listen, but you probably hate them for it.
The "Leaky" Bot is fun and memorable, but it's a bit risky. It's like a witty friend who teases you; it keeps you engaged, but you have to make sure they don't cross the line into being offensive.

Why This Matters for the Future

The researchers suggest that chatbots shouldn't just default to being "super polite" all the time. If they are too polite, we stop paying attention. If they are too direct, we get angry.

The best approach might be a mix. Sometimes, a chatbot needs to be a polite assistant. Other times, it needs to be a "human-like" character that slips up, makes a joke, or challenges you gently. Just like in real life, we need different types of friends for different situations.

In short: If you want a chatbot to change your behavior, don't just make it a robot that says "Please." Give it a little bit of personality, a little bit of a stumble, and maybe a tiny bit of sass. It might just make you actually listen.

1. Problem Statement

As conversational agents (chatbots) are increasingly deployed for behavior change interventions (e.g., health, sustainability), designers face a critical challenge: how to deliver corrective feedback without triggering psychological reactance. Reactance is a motivational state triggered when individuals perceive their freedom of choice is threatened, often leading to defensiveness or a "boomerang effect" where users double down on their original behavior.

While politeness strategies (indirect language) are known to reduce reactance, they risk being perceived as boring, unengaging, or insincere. Conversely, direct feedback is clear but often provokes high reactance. There is a lack of empirical understanding regarding alternative feedback styles that might balance low reactance with high engagement. This study investigates whether a novel style, Verbal Leakage (using slips, disfluencies, and self-corrections to reveal intent), can offer a middle ground between the safety of politeness and the engagement of directness.

2. Methodology

The authors conducted a 3 × 2 mixed-factorial online user study ( $N = 158$ ) to evaluate the impact of feedback style and scenario context on user perception.

Independent Variables:
1. Feedback Style (Between-subjects):
  - Direct: Explicit, straightforward correction of behavior (e.g., "You should not eat pizza; eat vegetables instead.").
  - Politeness: Indirect, respectful strategies to preserve user autonomy (e.g., "Have you considered trying a lighter meal?").
  - Verbal Leakage: The chatbot uses disfluencies (e.g., "uh," "um"), slips, and self-corrections to seemingly "accidentally" reveal a direct stance before softening it (e.g., "Oh, you're going for something that might not... uh, support a balanced lifestyle. Maybe try a salad?").
2. Psychological Distance (Within-subjects):
  - Personally-Affecting: Scenarios impacting the user's immediate self (diet, sleep, finances).
  - Societally-Affecting: Scenarios impacting broader society (sustainable travel, water conservation, civic voting).
Procedure:
- Participants interacted with a chatbot (powered by GPT-4o) in a simulated interface within Qualtrics.
- They were presented with scenarios where they expressed an intention conflicting with recommended behaviors.
- The chatbot provided feedback based on the assigned style.
- Participants rated their experience on 7-point Likert scales measuring emotional reactance (anger, guilt, surprise), perceived threat to freedom, and message effectiveness (processing, persuasiveness).
- Qualitative data was collected via open-ended questions regarding behavioral intentions and sentiment.

3. Key Contributions

Empirical Comparison of Novel Styles: The study introduces and empirically validates Verbal Leakage as a distinct feedback style in HCI, moving beyond the binary of "Direct vs. Polite."
Identification of Trade-offs: It provides evidence that minimizing psychological reactance (via Politeness) often comes at the cost of user engagement and surprise, creating a "boring but safe" dilemma.
Contextual Nuance: It demonstrates how the effectiveness of feedback styles varies significantly based on whether the scenario is personally or societally affecting.

4. Key Results

Quantitative Findings

Psychological Reactance (Anger & Threat to Freedom):
- Politeness resulted in the lowest levels of anger and perceived threat to freedom.
- Direct and Verbal Leakage both elicited significantly higher levels of threat to freedom and anger compared to Politeness.
Engagement (Surprise):
- Verbal Leakage elicited the highest levels of surprise, significantly outperforming Politeness.
- Politeness was rated as the least surprising and most predictable.
Message Effectiveness (Persuasiveness):
- Politeness was rated as the most persuasive style, followed by Verbal Leakage, with Direct being the least persuasive.
- Verbal Leakage was significantly more persuasive than Direct but less persuasive than Politeness.
Behavioral Intention (Qualitative Coding):
- Politeness led to the highest proportion of participants moving to the "Changed Action" stage (32.73%).
- Direct feedback resulted in the highest proportion of "Pre-contemplation" (no intention to change, 51.96%), often triggering a boomerang effect where users rebelled against the "bossy" tone.
- Verbal Leakage fell in the middle, with 22.12% moving to "Changed Action."

Qualitative Findings

Direct Style: Perceived as "forceful," "rude," and "condescending." Users felt their autonomy was threatened, leading to resistance and annoyance.
Politeness Style: Perceived as "kind" and "supportive" but also "bland," "monotone," and "boring." Users felt the chatbot lacked personality, leading to disengagement despite low reactance.
Verbal Leakage: Perceived as "human-like," "sassy," and "playful." It successfully broke the "robotic" expectation. While it sometimes induced guilt or feelings of being "called out," it was also seen as more authentic and engaging than the other styles.

5. Significance and Implications

The study highlights a fundamental design trade-off in conversational agents:

Politeness minimizes negative emotions (anger) and maximizes compliance but risks user boredom and disengagement.
Directness maximizes clarity but triggers reactance and resistance.
Verbal Leakage offers a middle ground: it increases engagement and surprise (making the interaction memorable) and maintains moderate persuasiveness, though it carries a higher risk of triggering reactance than politeness.

Design Recommendations:

Move Beyond Defaults: Designers should not default to purely polite or purely direct styles. A repertoire of styles is necessary.
Context Matters: Direct or challenging styles may be more acceptable in personal contexts (e.g., diet) but trigger strong reactance in societal contexts (e.g., civic voting).
User Agency: Systems should allow users to choose their preferred interaction style (e.g., "Coach-like/Direct" vs. "Assistant-like/Polite") to align with their expectations and reduce reactance.
Humanization: Incorporating rhetorical devices like verbal leakage can make agents feel more authentic and less like "pre-programmed scripts," potentially increasing engagement without the severe reactance of blunt commands.

In conclusion, effective behavior-change chatbots must navigate the tension between being safe (low reactance) and engaging (high surprise/personality). Verbal Leakage emerges as a promising, albeit riskier, alternative to the standard polite default.