C3NLP @ ACL 2026 · San Diego

One Style Fits All?
Cultural Values Embedded in Conversational AI

How supportive vs. challenging chatbot styles shape user experience across Taiwan and Korea, and what that reveals about cultural scripts in LLMs.

Yi-Jun Chen1, I-Tsen Hsieh2, Li-Wun Chang2
1Yonsei University, 2Independent Researcher

↗ GitHub

Overview

How do NLP systems reflect and reinforce cultural values?

LLMs trained on web-scale corpora absorb dominant cultural values. When deployed across diverse populations, an interaction style that feels natural to one group may feel intrusive, or even threatening, to another. We put real users from underrepresented cultures in front of these systems and measured how they actually react.

Study framework diagram
LLMs embed cultural values that shape how chatbots interact. We tested whether people-pleasing tendency shapes how Taiwanese and Korean users experience these chatbot styles, through continuance intention, satisfaction, trust, and perceived threat.

The Gap We Address

Most existing work examines model outputs. Far fewer studies put real users from underrepresented cultures in front of deployed LLMs and measure how culturally embedded interaction styles are actually received. We focus on Taiwanese and Korean users, whose cultures and languages are underrepresented in mainstream training data, and ask whether a personality trait, people-pleasing tendency, changes the picture.

RQ1

Does people-pleasing tendency (PPT) moderate the effect of chatbot interaction style on users' intention to collaborate?

RQ2

Does PPT moderate the relationship between chatbot style and users' collaboration experience (satisfaction, trust, threat)?

RQ3

How do cross-cultural differences in style perception reflect different interactional expectations across cultural contexts?

Experimental Design

Two interaction styles, four design dimensions

We operationalized supportive and challenging interaction styles through GPT-4.1 system prompts (temperature 0.2) along four dimensions, each mapping onto contrasting cultural interactional scripts. Both styles were then deployed in a controlled interaction task with real users from Taiwan and Korea.

Supportive chatbot

Supportive Style

Harmony-oriented · Emotional support
  • Linguistic Positive emotion words, function-word mimicry, nonverbal cues (emojis)
  • Values Harmony-oriented, other-focused
  • Interaction Strategy Emotional support, encouragement
  • Tone Warm, affiliative
Challenging chatbot

Challenging Style

Autonomy-oriented · Reflective guidance
  • Linguistic Neutral language, no mimicry, minimal affective cues
  • Values Autonomy-oriented, self-directed
  • Interaction Strategy Reflective guidance, critical questioning
  • Tone Neutral, evaluative

Study Procedure

Three phases, one collaborative task

101 participants (Taiwan n=49, Korea n=52) from the MZ generation (born 1981–2007) completed a collaborative tourism-planning task in their native language, interacting with either the supportive or challenging chatbot.

1

Pre-Interaction Survey

Informed consent, Basic Information, People-Pleasing Scale

→
2

Human–Chatbot Interaction Task

3-stage tourism-planning task,
(recall → itinerary → promo copy)
min. 8 turns per stage

→
3

Post-Interaction Survey

Perceived Threat, Satisfaction, Trust, Continuance Intention, Open-ended questions

MeasureItemsScaleSource
People-Pleasing Scale (PPS)105-point LikertBlötner (2025)
Perceived Threat (PT)67-point LikertYanan (2023)
Satisfaction (SA)35-point LikertChin et al. (1988)
Trust (TR)87-point LikertDolinek and Wintersberger (2022); Scharowski et al. (2024)
Continuance Intention (CI)37-point LikertBhattacherjee (2001)

All instruments administered in the participant's native language (Taiwan Mandarin / Korean). All scales α > .70 in both samples.

Results

One style works; threat reflects the culture

1

Supportive style outperformed across both samples

Supportive style increased CI, satisfaction, and trust in both samples.

CI: β = 2.48 (TW) · 2.17 (KR), p < .001
SA: β = 1.83 (TW) · 1.44 (KR), p < .001
TR: β = 1.88 (TW) · 1.31 (KR), p < .001
2

Perceived threat is the only cultural divergence

The challenging style raised threat in Taiwan but NOT in Korea.

PT: β = −2.65 (TW, p<.001) · −0.73 (KR, n.s.)
3

PPT did not moderate style effects, but it wasn't absent

All Style × PPT interactions were n.s. However, PPT showed a Taiwan-only main effect on perceived threat (β = .08, p < .05), reflecting higher baseline sensitivity, not stronger responsiveness to style.

Style × PPT (all n.s.) · PT main effect TW only

Full regression coefficients for all outcome variables across both samples.

CISATRPT
PredictorSampleβpβpβpβp
Chatbot StyleTW2.48***<.0011.83***<.0011.88***<.001−2.65***<.001
KR2.17***<.0011.44***<.0011.31***<.001−0.73n.s.
People-Pleasing Tendency (PPT)TW0.07n.s.0.03n.s.0.02n.s.0.08*<.05
KR0.02n.s.0.02n.s.0.04n.s.0.08n.s.
Style × PPTTW−0.07n.s.−0.03n.s.−0.03n.s.−0.05n.s.
KR0.02n.s.0.02n.s.0.02n.s.−0.06n.s.
R²TW0.530.630.590.66
KR0.350.340.450.16

Discussion

Four findings, four design implications

The same conversational style improved collaboration across cultures, but did not carry the same social meaning.

1

Warm Cues Travel

Supportive style consistently increased continuance intention, satisfaction, and trust in both samples, regardless of cultural background.

2

Style Outweighs Personality

Supportive style benefits are robust across interpersonal orientations. PPT reflects users' baseline emotional sensitivity in human-AI interaction, not their responsiveness to specific chatbot styles.

3

Same Output, Different Scripts

Challenging style increased perceived threat in Taiwan but not Korea, even across two broadly collectivist settings. The same output activated divergent social scripts: harmony-sensitive in Taiwan, task-oriented in Korea.

4

Design for Culture, Not Stereotypes

Use supportive cues as a cautious default; audit directness; and adapt to linguistic signals, not static personality labels.

Conclusion

Supportive cues improved collaboration across Taiwan and Korea, but culture-specific threat responses show that conversational AI carries social norms—not neutral information alone.

Limitations

Future directions 🔜

These results suggest a cultural pattern, not a validated cultural mechanism.

Style Fidelity & User Drift

Off-task requests sometimes shifted the intended chatbot style.

Potential Language-Model Bias

English-dominant training may introduce Western defaults into Mandarin and Korean outputs.

Interpretation Remains Inferential

Quantitative patterns and open-ended clues do not directly validate participants' cultural interpretations.

Limited Cultures & Low-Stakes Task

Two East Asian samples and a tourism task limit generalization to other cultures and high-stakes domains.