How French People Use AI in Everyday Life

Using the Compar:IA dataset to analyze French AI conversations and reveal real-world usage patterns

Article written by Dr. Charles de Dampierre (CEO & Co-Founder at Bunka.ai)

In partnership between compar:IA (French Ministry of Culture) and Bunka.ai

---

The Challenge of Understanding AI's Real-World Impact

Despite the widespread adoption of generative AI systems, it remains surprisingly difficult to understand how people actually use these technologies in their everyday lives. Market reports typically focus on adoption rates and industry growth, while academic studies often examine controlled scenarios that may not reflect authentic usage.

The gap between what we assume about genAI conversational usage and the reality of how people incorporate these tools into their daily routines presents a significant blind spot in our understanding.

Why is it important?

Studying how people actually use conversational AI systems is key for three reasons. First, it reveals potential failures if the systems that only emerge in diverse real-world situations that are impossible to anticipate in controlled settings. Second, it helps identify when specialized interfaces are needed, for example, when coding assistance becomes common enough to justify dedicated developer tools. Finally, usage patterns serve as a valuable societal barometer, giving researchers and businesses insights into genuine public needs and priorities that traditional research methods might miss entirely.

While datasets such as WildChat, ShareGPT, and LMSYS provide valuable insights into human-AI interactions, none have specifically focused on French usage patterns. To solve those issues, we analyze data from the Compar:IA dataset, the biggest comprehensive collection of conversations from the French Ministry of Culture's AI comparison platform (comparia.beta.gouv.fr), launched in October 2024. What makes this dataset particularly valuable is that it represents genuine, unconstrained usage primarily in French, capturing natural interactions rather than controlled scenarios. This dataset contains 175,000+ authentic questions and responses from over 59,475 distinct users interacting with more than 30 different conversational AI models of various sizes, both open-source and proprietary.

---

The 4 key findings from the Bunka.ai's AI Usage Report are

• French users primarily engage with conversational AI across technical domains (Computing Technology, 11%), educational contexts (Knowledge Education, 10%), and professional fields (Business Workplace, 9%), with distinct interaction patterns (different types of language levels) emerging between professional and personal usage contexts.
• French users primarily employ AI to learn new things (27%), get advice (19%), creating content (16%) and retrieving information (16%), revealing a strong diversity of usage across the different domains of personal and professional life.
• Different topics show distinct task patterns: Health-related topics primarily involve advice-seeking (50%), Creative & Arts related topics emphasize content creation (45%), Social-related topics feature the most planning activities (22%) while Science-related topics focus on learning (55%).
• French users prefer using AI as a learning partner (68%) rather than just a task automator (32%), with collaborative learning approaches dominating most topics (especially technology-related interactions at 87%, humanities at 78%, and science at 79%), while command-based interactions only become common in work-related contexts (46%) and creative projects (53%).

---

Challenges in Analysing the Data

Classifying and analyzing AI conversations is complicated for two main reasons: first, these interactions have many dimensions (like user goals, topics, relationship styles, tasks, etc.) and traditional classifications (such as sentiment analysis) often fail to capture the nuanced ways people integrate AI into their workflows and daily lives, especially because this is still a very unknown behavior. Second, there's a massive volume of text data to analyze, making it hard to quickly understand thousands of weekly exchanges and take actions if needed.

We used visual topic modeling to intuitively understand the content, alongside unsupervised dimension analysis to discover natural patterns in real-world usage data without imposing predetermined classifications.

Our approach combined two complementary methods: visual topic modeling to create intuitive graphical representations of content patterns, and unsupervised dimension analysis to identify natural classifications that emerge from real-world usage data without imposing predetermined categories.

1. Visual Topic Modeling for Broad and Quick Understanding

We created a topographical map of AI usage patterns inspired by the BunkaTopics package, our visual topic modeling tool, to generate a bird's-eye view of conversation clusters. This approach, similar to Anthropic's framework, maintained user privacy while revealing natural groupings in the data.

2. Unsupervised Classification Analysis

A key challenge when studying how people use AI is avoiding the bias that comes from creating categories before examining the data. Instead of forcing conversations into predetermined boxes, we let our classification system emerge naturally from the actual conversations. This approach revealed authentic French usage patterns that standard classifications would miss, including unexpected relationships between topics and tasks. Our only input was defining which dimensions to analyze (tasks, topics, emotions, language), while the specific values for these dimensions emerged naturally from the data distribution rather than being predetermined.

• Tasks: The functional objective of user-AI exchanges (e.g., getting advice, creating something, analyse, etc)
• Topics: The domains represented in conversational content (e.g., Law, Health, Business, etc)
• Emotional Engagement: The affective valence and intensity manifested during interactions
• Language Complexity: The linguistic sophistication of the interactions (e.g., Technical language, every-day life language, etc)

To identify categories within each dimension, we first used Claude Sonnet to analyze a sample of 5000 conversations, asking it to classify content by finding the most common topics, tasks, emotions, and language patterns in every interaction. Claude initially generated hundreds of potential categories, which we then systematically consolidated through iterative refinement (for example, reducing roughly 100 topics to 15 broader topics) to create practical classification systems for our analysis.

We then added a last dimension, Human-AI Relationship inspired from the work of Anthropic about the relationship between users and AI. The values include Validation, Task Iteration and Learning (Augmentative usage) and Directive and Feedback Loop (Automation Usage).

For efficient large-scale analysis, we used a LLaMA 3.3-70B model to classify a bigger sample of the dataset across these dimensions, validating the approach through manual annotation of a representative subset which confirmed high classification accuracy (F1 scores ranging from 80% to 96%). The use of a LLama 70B makes it possible to do LLM-as-juge on 25,000 interactions at a lower cost than using Claude Sonnet. The full dimensions values found after the analysis can be found in the table below.

Dimensions and their values. We annotated the data on a sample of 50 interactions for each interaction.

---

Results: How French Users Engage with AI

Task Distribution & Human-AI Relationship

This distribution suggests French users primarily value conversational AI systems for educational purposes, decision support, content creation, and information retrieval. Learning-oriented tasks (knowledge acquisition, conceptual understanding, and educational exploration) constitute 27% of interactions. Advisory functions (guidance solicitation, recommendation, decision-making support) represent 19% of usage. Content generation (creative writing, original content production) and information retrieval (targeted information access about things or people) demonstrate equivalent use at 16% each. Analytical functions (data examination, content evaluation) show more limited representation at 6%.

Additionally, French users favor augmentative behaviors (Learning at 64%, Task Iteration at 2%, and Validation at 1%) over automated behaviors (Directive at 32%), indicating they primarily view AI as a collaborative learning partner rather than a tool for simple task automation. Augmentative behaviors are related to tasks that are collaborative between the user and the AI. Automotive tasks require just the user input and are often a "one-shot" questions require minimal interaction from the user.

Topic Distribution

Analysis of topic distribution across French AI interactions reveals a diversified engagement pattern with technological domains showing modest predominance. Computing & Technology (programming languages, software development, hardware configurations, cybersecurity protocols) leads at 11%. Knowledge Education (pedagogical methodologies, academic content development, skills acquisition frameworks, etc) follows at 10%. Three domains demonstrate equivalent representation at 9%: Social Humanities (sociological theory, psychological concepts, philosophical inquiry, historical analysis etc), Business Workplace (management methodologies, organizational dynamics, professional environment optimization, etc.), and Law Governance (regulatory frameworks, legal systems, political structures, etc.).

Key Correlations

Task-Topic Alignment

Different topics naturally trigger specific types of AI interactions. Health-related conversations primarily focus on advice-seeking, technology discussions center around problem-solving, natural science topics drive learning-oriented exchanges, social activities prompt planning assistance, and cultural domains are correlated with content generation.

Tasks vary across topic domains, with Natural Sciences dominated by Learning tasks (64%), Health Wellness showing the highest proportion of Advice (50%), Creative Arts featuring the most Creating activity (45%), Food Cuisine displaying the highest Retrieval usage (37%), and Society Recreation showing the greatest Planning activity (22%).

Examples (Topic x Task)

Health Wellness x Advice

The user wished to receive a simple fitness training program to get started with working out.

Computing & Technology x Problem Solving

The user wished to receive assistance with a website.

Natural Science x Learning

The user sought information about the time it takes for the Earth to complete one rotation and one revolution.

Health Wellness x Planning

The user wished to develop a detailed communication plan to raise awareness about specialized Alzheimer's care teams among medical professionals.

Complexity-Topic Connections

Language sophistication varies predictably by subject matter. Technical topics like AI, Economics, Natural Science, Law, and Technology feature specialized vocabulary and complex sentence structures, reflecting their professional, expertise-driven nature. In contrast, everyday topics such as Creative Arts, Media-Communication, Food-Cuisine, Society-Recreation, and Social-Humanities show more conversational language patterns, focusing on accessibility and practical application. This clear division highlights how users naturally adjust their communication style between professional contexts (prioritizing precision) and personal usage (emphasizing approachability).

Examples (Topic x Language-level)

Law Governance x Expert Language

The user sought information on how multilateral trade agreements address health and environmental standards.

Food Cuisine x Everyday Talk

The user wished to learn how to make a chocolate-filled pastry.

Automation vs Augmentation

Overall, our analysis revealed similar patterns with 65% augmentation and 35% automation usage. The balance between directive and collaborative AI use follows clear domain patterns. Business contexts, Creative Arts, and Food & Cuisine favor automation-oriented approaches, with users efficiently delegating specific tasks for direct results with minimal interaction (like drafting a mail, asking for a kid's bedtime story or asking for a recipe). In contrast, domains like Social Sciences & Humanities, Health & Wellness, and AI & Robotics show strong preference for augmentation, featuring extended learning dialogues where users work alongside AI to enhance their understanding rather than simply outsourcing tasks.

Examples (Topic x AI-Interaction Mode)

Creative Arts x Directive

The user wished to create a story where a character transforms into a fish to defeat an elephant.

Social Humanities x Learning

The user was told a 100-word story about a child discovering an ancient tradition in the Mayenne region.

---

Conclusion

Our analysis provides insights into how French users are integrating conversational generative AI into their daily lives, revealing nuanced patterns across multiple dimensions of interaction. The strong preference for collaborative augmentation over simple automation suggests users are finding value in chatbot systems as an interactive partner rather than merely a labor-saving tool. These findings align with previous Anthropic research showing that 60% of Claude users employ it for augmentation versus 40% for automation—our analysis revealed similar patterns with 65% augmentation and 35% automation usage.

These findings contribute to ongoing discussions among technologists, policymakers, and social scientists about AI's role in society. Further research will help us better understand the long-term implications of these human-AI relationships and how they might shape both technology development and social practices in the years ahead.

The Compar:IA platform had several technical limitations during the study period. The integrated models were restricted to text-only processing, excluding image, video and audio analysis, and users could not import external documents or access advanced features such as web search functionality or image generation tools. Those added capabilities could have led to new behaviors. Additionally, the platform's suggested prompts may have introduced selection bias by influencing users' choice of topics and tasks. The full list is available here.

We recognize that we are only at the beginning of understanding AI's impact on everyday life. As AI capabilities grow and become more integrated into daily routines, fundamental questions arise: How will these systems reshape professional practices? What social and cognitive skills might be enhanced or diminished through AI interaction? How might different cultural contexts shape distinct patterns of AI adoption and usage? Future research could employ additional methodologies from social sciences to better understand user behaviors, including cohort studies and anthropological approaches. The current study relies solely on interaction analysis without incorporating users' socio-economic (SES) characteristics (age, occupation, gender), which could provide valuable contextual insights into usage patterns. Additionally, concepts from Human-Machine Interaction (HMI) and psychology could help us better understand these new types of relationships that people are developing.

---

BibTeX

If you'd like to cite this report, you can use the following BibTeX key:

@online{bunka2025aiusage,
    author = {Charles de Dampierre and Louis-Marie Lorin and Malte Meng},
    title = {Bunka.ai AI Usage Report: How French People Use AI in Everyday Life},
    date = {2025-05-19},
    year = {2025},
    url = {https://bunka.ai/articles/french-ai-usage-study},
}

Acknowledgements

This research was conducted by the Bunka.ai Research Team (Charles de Dampierre, Louis-Marie Lorin and Malte Meng). Thanks for the comments of the Compar:IA Team, Lucie Termignon and Hadrien Pelissier.

Appendix

Dimensions Annotation Score