💬 NLP

From Chatbots to Confidants: A Cross-Cultural Study of LLM Adoption for Emotional Support

This study presents the first large-scale cross-cultural analysis of LLM adoption for emotional support across seven countries, revealing significant variations in usage rates and perceptions driven by demographic factors and cultural context, while highlighting the diverse emotional needs users seek to address through these systems.

Original authors: Natalia Amat-Lefort, Mert Yazan, Amanda Cercas Curry, Flor Miriam Plaza-del-Arco

Published 2026-04-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Natalia Amat-Lefort, Mert Yazan, Amanda Cercas Curry, Flor Miriam Plaza-del-Arco

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

The Big Picture: The Digital "Ear"

Imagine you have a very smart, always-available friend who never judges you, never gets tired, and is free to talk to 24/7. That is essentially what this study is about: people using AI chatbots (like ChatGPT) not just to write emails or solve math problems, but to vent, feel less lonely, and process their emotions.

The researchers wanted to know: Who is talking to these digital friends, and why do some countries love them while others are skeptical?

To find out, they asked 4,641 people from seven different countries (the US, UK, Germany, France, Spain, Italy, and the Netherlands) about their experiences. They also looked at 731 real messages people sent to these bots.

1. The "Who": Who is using these bots?

The study found that not everyone is equally likely to turn to an AI for a heart-to-heart. It's like a club where the membership depends on your life situation.

The "Sweet Spot" Demographic: The people most likely to use these bots are aged 25–44, married, religious, and wealthy (or feel they have a high social status).
- Analogy: Think of it like a high-end coffee shop. The people who go there most often aren't the teenagers hanging out on the street corner or the elderly folks who prefer tea at home; they are the busy, established adults who have the money and the time to sit down and chat.
The "Non-Users": People who are older (65+), single, have less education, or lower income are much less likely to use these bots for emotional support.
The Gender Gap: Interestingly, gender didn't make a huge difference in who used them, but it did slightly affect how much they trusted the bot's privacy.

2. The "Where": The Great Cultural Divide

This is the most surprising part of the study. The researchers discovered a massive split in how people feel about AI, depending on where they live.

The "Anglosphere" (UK & USA): People here are the enthusiasts. They trust the bots, see huge benefits in them, and use them frequently.
- Analogy: In the US and UK, people treat AI like a new, exciting gym membership. They are eager to try it out, thinking it will help them get fit (mentally).
Continental Europe (Germany, France, Spain, Italy, Netherlands): People here are the skeptics. They are much more cautious, trust the bots less, and use them far less often.
- Analogy: In these countries, people treat AI like a strange, unlicensed mechanic. They worry, "Is this thing going to fix my car, or is it going to steal my data and break the engine?"
The Netherlands: The Dutch were the most skeptical of all. They were the least likely to trust the bots with their secrets.

Why the difference?
The study used a special statistical "lens" to separate culture from demographics. They found that even if you took a wealthy, married American and a wealthy, married German, the American would still be more likely to trust the bot. Culture matters. It seems that in English-speaking countries, there is a more open attitude toward using technology for mental health, whereas in parts of Europe, there is more concern about privacy and a stronger preference for human interaction.

3. The "What": What are people actually saying?

The researchers looked at 731 real messages people sent to the bots. They found that people aren't just asking for jokes; they are in pain.

The Main Topics:
1. Stress and Anxiety: "I'm overwhelmed."
2. Relationship Conflicts: "My partner isn't texting me back."
3. Loneliness: "I feel like I don't belong."
4. Family Issues: "How do I deal with my difficult sibling?"
5. Mental Health Struggles: Dealing with depression or grief.
Why they use them:
- 24/7 Availability: The bot is there at 3 AM when a human therapist is asleep.
- No Judgment: People feel safe saying things to a machine that they are too scared to say to a friend or family member.
- Cost: It's free (or cheaper than therapy).

4. The "Catch": Is it safe?

The paper highlights a tension. People feel safe talking to these bots because they think the bot is a "non-judgmental confidant." However, the authors point out a few risks:

The "Fake Friend" Problem: The bots are designed to be nice. They often agree with you or try to comfort you, but they don't truly understand you. They might encourage you too much without offering a reality check.
Privacy Illusion: Even though users feel safe, the data they share isn't protected by "doctor-patient confidentiality" laws. The companies own the data.
The "English Bias": The study notes that the bots work best in English. If you speak Dutch, French, or German, the bot might sound a bit unnatural, which could explain why non-English speakers are less enthusiastic.

Summary

This paper is a snapshot of a new world where people are turning to robots for comfort.

Who? Mostly younger, wealthier, married adults.
Where? Much more popular in the US and UK than in Europe.
Why? Because they are lonely, stressed, and need someone to listen without judging.
The Warning: While these bots are helpful, they are not perfect doctors. They are like a very polite, always-available neighbor who listens to your problems but doesn't actually know the full truth about your life or the legal rules protecting your secrets.

The researchers conclude that we need to be careful about how we build and regulate these tools to make sure they are safe for everyone, not just the people who are already comfortable with technology.

1. Problem Statement

While Large Language Models (LLMs) are increasingly utilized for instrumental tasks, their adoption as "always-available, non-judgmental confidants" for emotional support is growing rapidly. However, significant gaps exist in the empirical understanding of this phenomenon:

Lack of Population-Level Data: It is unclear how common LLM use for emotional support is among the general public.
Sociocultural Blind Spots: Existing research is predominantly confined to single-country (often US/UK) studies, leaving a gap in understanding how cultural backgrounds and sociodemographic factors shape perceptions and adoption.
Unknown Disclosure Patterns: There is limited evidence on what users actually disclose to these systems when seeking emotional support.
Conflation of Factors: Previous studies often fail to disentangle whether observed differences in adoption are due to genuine cultural effects or merely artifacts of varying demographic compositions (e.g., age, education, socioeconomic status) across regions.

2. Methodology

The authors conducted a large-scale, cross-cultural quantitative study using a mixed-methods approach.

Data Collection:
- Sample: 4,641 participants recruited via Qualtrics panels across seven countries: USA, UK, Germany, France, Spain, Italy, and The Netherlands.
- Sampling Strategy: Quota sampling ensured even distribution across countries ( $N=200$ per country) and gender balance.
- Survey Design: A four-step survey covering demographics, usage patterns, perception constructs, and an optional open-ended prompt collection.
- Quality Control: Strict filters excluded bots, duplicates, straight-lining, and failed attention checks. Completion times were validated against the median.
- Translation: The survey was translated into six languages (English, Spanish, French, Italian, German, Dutch) and reviewed by native-speaking experts for cultural and linguistic accuracy.
Analytical Framework:
- Demographic Analysis: Chi-square tests compared users vs. non-users.
- Perception Constructs: Three composite variables were created using 5-point Likert scales:
  1. Trust & Privacy (TRU): Credibility, transparency, and data safety.
  2. Perceived Benefits (BENF): Accessibility, cost-effectiveness, and psychological safety.
  3. Usage Intention & Behavior (USE): Attitude toward use and actual frequency.
- Statistical Modeling: Cumulative Link Mixed Models (CLMM) were employed to predict the three constructs.
  - Fixed Effects: Demographic variables (Age, Gender, Education, SES, Religion, Marital Status).
  - Random Effects: Country (to isolate cultural variance from demographic variance).
- Prompt Analysis: 731 real user prompts were collected. They were embedded using M3-Embedding (multilingual) and Sentence-Transformer, then clustered using HDBSCAN and UMAP. Topics were categorized via LLM (GPT-5.2) and manual evaluation.

3. Key Contributions

First Cross-Cultural Empirical Study: Provides the first population-level evidence of LLM adoption for emotional support across seven diverse Western nations.
Disentangling Culture vs. Demographics: Uses advanced mixed modeling to separate the effects of cultural background from demographic composition, revealing that both are significant but distinct drivers.
Real-World Prompt Corpus: Curates and analyzes 731 multilingual prompts, identifying the specific emotional topics users seek help with (e.g., loneliness, trauma, relationship conflicts).
Sociotechnical Landscape Mapping: Highlights the complex interplay between mental health stigma, digital literacy, privacy concerns, and access to traditional care in shaping AI adoption.

4. Key Results

A. Adoption Rates and Demographics

Adoption Variance: Adoption rates vary dramatically by country, ranging from 20% (France, Netherlands) to 59% (UK). The average across all seven countries is 28.9%.
User Profile: Users are significantly more likely to be:
- Aged 25–44 (vs. non-users who skew older).
- Higher Socioeconomic Status (SES) (The strongest predictor).
- Married/Partnered.
- Religious.
- Highly Educated (MSc/PhD).
Usage Patterns: ChatGPT is the dominant agent (79.7%). Primary use cases include stress/anxiety management (49.4%), emotional processing (41.3%), and relationship support (30.2%).

B. Cross-Cultural Perceptions (The "Anglosphere Divide")

Cultural Effect: Even after controlling for demographics, a stark divide exists.
- Anglosphere (UK, USA): Show consistently positive perceptions across Trust, Benefits, and Usage.
- Continental Europe: Show more skeptical or negative perceptions.
- The Netherlands: Stands out as the most skeptical country, with the lowest scores across all constructs, likely due to high privacy concerns and strict data protection norms.
Statistical Significance: The UK and USA form a distinct cluster, statistically different from all other nations. Mainland European countries (Germany, France, Spain, Italy) are largely homogeneous, though the Netherlands diverges significantly from them.

C. Mixed Model Findings

Demographic Drivers:
- SES: Being in the highest SES bracket (>7) is the strongest positive predictor of trust, benefits, and usage.
- Age: The 25–44 age group is most positive; being 65+ is the strongest negative predictor.
- Religion: Believers (regardless of specific faith) hold more positive views than non-believers.
- Marital Status: Married/partnered individuals are more positive than singles.
Cultural Drivers: The country-level random effects confirm that cultural context independently influences perception. The UK and USA have large positive intercepts, while the Netherlands has a large negative intercept.

D. Prompt Analysis (User Disclosures)

Content: Of the 731 prompts, 36.5% related to mental wellbeing.
Key Topics:
1. General stress and anxiety.
2. Relationship conflicts.
3. Family and caregiving issues.
4. Loneliness.
5. Trauma and mental health struggles.
Cross-Cultural Consistency: Topic distributions were consistent across countries, suggesting universal human needs for emotional support regardless of cultural context.

5. Significance and Implications

Ethical & Safety Concerns: The study reveals a "perceived safety" paradox. Users feel safe disclosing sensitive data due to the lack of human judgment, yet they often lack legal protection (no doctor-patient confidentiality) and may receive "sycophantic" or non-critical responses from LLMs.
Inequality in Access: High adoption among high-SES, educated, and younger groups suggests a "digital divide" where vulnerable populations (older, lower SES) may be excluded or lack the literacy to use these tools effectively.
Design & Governance: The findings call for:
- Multilingual Benchmarks: Current models may favor English speakers (Anglosphere enthusiasm), potentially offering a superior experience to English users compared to non-English speakers.
- Governance Frameworks: Urgent need for regulations regarding data privacy, especially given the sensitivity of mental health disclosures.
- AI Literacy: Programs to help users critically assess the limitations of AI as a therapeutic tool.
Future Research: Highlights the need to expand studies beyond Western contexts and to investigate the long-term psychological impacts of human-AI emotional bonding.

In conclusion, the paper establishes that LLM adoption for emotional support is not uniform but is shaped by a complex matrix of socioeconomic privilege and cultural attitudes toward privacy and mental health, with a distinct divide between English-speaking and Continental European populations.