Artificial intelligence is often portrayed as a productivity machine built to write emails, summarise documents, and help developers code faster. While those use cases are real, fresh large-scale usage data suggests the reality of how people actually interact with AI in 2025 is far more complex, surprising, and human than popular narratives imply. A recent data-driven analysis of real-world AI interactions reveals that people are not just using AI as a work assistant but also as a creative partner, a problem-solving agent, and even a roleplay companion. These insights offer a clearer picture of where AI is heading and how it is becoming embedded into daily digital life.
The study draws from an enormous dataset analysing over 100 trillion tokens, representing billions of interactions across a wide range of large language models. Instead of examining conversation content, the analysis focused on metadata such as model choice, task category, language, and prompt length, ensuring privacy while still uncovering meaningful behavioural trends. Because the data spans hundreds of models from dozens of providers and reflects global usage rather than a single platform, it offers one of the most comprehensive views yet into how AI is actually being used at scale.
Open-source AI’s expanding footprint
One of the most notable trends is the steady rise of open-source AI models. By late 2025, open-source systems accounted for roughly one-third of all observed AI usage, with major jumps following high-profile releases. This growth highlights increasing trust in open ecosystems and shows that developers and users alike are willing to adopt alternatives beyond closed commercial models when performance and flexibility align with their needs. Open-source AI is no longer a niche experiment; it is a core part of the global AI landscape.
The unexpected dominance of roleplay and creative use
Perhaps the most surprising insight from the data is what people are doing with open-source AI models. More than half of all open-source model interactions are centred on roleplay and creative storytelling rather than traditional productivity tasks. Users are engaging in character-driven conversations, interactive fiction, and game-like scenarios at a scale that far exceeds expectations. This finding challenges the assumption that large language models are primarily tools for professional efficiency.
These interactions are not random or shallow. The data indicates that a large share of roleplay usage involves structured narratives and defined gaming contexts, showing that users treat AI as a flexible storytelling engine. For many, AI serves as a space for exploration, creativity, and even companionship. This invisible yet massive use case is already influencing how AI developers think about model alignment, safety, and long-term engagement.
Programming becomes AI’s fastest-growing application
While creative roleplay dominates certain segments, programming has emerged as the fastest-growing category across all AI models. At the beginning of 2025, coding-related interactions made up a relatively small portion of overall usage. By the end of the year, they represented more than half of all AI activity. This dramatic shift reflects how deeply AI has integrated into modern software development workflows.
Developers are no longer asking for simple snippets. Average prompt sizes for programming tasks have increased several times over, often spanning thousands of tokens and sometimes exceeding the length of entire codebases. AI is now being used for advanced debugging, architectural analysis, refactoring, and multi-step reasoning. Certain models have become particularly strong in this area, dominating developer preference for much of the year, although competition continues to intensify as new models close the gap.
China’s rapid rise in the global AI ecosystem
Another major transformation revealed by the data is the growing influence of Chinese AI models. By late 2025, models developed in China accounted for around 30 percent of global AI usage, a sharp increase from earlier in the year. Systems built by companies such as DeepSeek, Qwen, and Moonshot AI have seen explosive adoption, processing trillions of tokens within months.
This rise is also reflected linguistically. Simplified Chinese has become the second most common language used in AI interactions worldwide, and Asia’s overall share of AI usage has grown dramatically. The data points to a more multipolar AI world, where innovation and adoption are no longer dominated by a single region.
The shift toward agentic AI systems
Beyond what people ask AI to do, how they interact with it is also changing. The data highlights the rapid emergence of agentic AI, where models operate within multi-step workflows rather than responding to isolated prompts. These systems reason across longer contexts, invoke tools, and persist through extended interactions to complete complex tasks.
By the end of 2025, more than half of AI interactions fell into this reasoning-optimised category. This shift marks a fundamental evolution from AI as a text generator to AI as an autonomous digital agent capable of planning, execution, and iterative problem-solving. Users increasingly rely on AI to handle entire processes rather than single outputs.
Understanding long-term user loyalty
The study also sheds light on why users stick with certain AI models over time. Researchers observed what they describe as a “Glass Slipper Effect,” where a model that is first to solve a user’s critical problem creates unusually strong retention. When users discover a model that fits their needs perfectly at the right moment, they tend to integrate it deeply into their workflows and are far less likely to switch later.
This finding suggests that success in the AI market is not only about being first or cheapest but about delivering meaningful value at exactly the right time. Once embedded, switching costs become behavioural as well as technical, reinforcing long-term loyalty.
Why pricing is not the main driver of usage
Contrary to common assumptions, AI usage does not strongly respond to price changes. The data shows that even significant price reductions lead to only modest increases in usage. Expensive premium models and low-cost alternatives coexist successfully, indicating that users weigh factors such as reasoning quality, reliability, and capability more heavily than raw price.
This behaviour suggests that AI has not yet become a commodity market. Instead, differentiation through performance and trust still matters, allowing multiple pricing tiers to thrive simultaneously.
What these insights mean for the future of AI
Taken together, these findings reveal an AI ecosystem that is far more diverse and human-centric than surface-level narratives suggest. AI is powering serious professional workflows, especially in programming, while also enabling creative expression and new forms of interaction. Global adoption patterns are shifting, agentic systems are becoming the norm, and user loyalty depends on real problem-solving rather than hype.
Most importantly, the way people use AI varies widely by region, language, and purpose. Understanding these real-world patterns will be essential as AI continues to integrate into everyday life. The gap between how AI is imagined and how it is actually used remains significant, but studies like this help bring clarity. In 2025, AI is not just a tool for efficiency; it is a multifaceted technology reshaping how people create, think, and interact with digital systems.
