For decades, the concept of a "user model"—a computational representation of a user's knowledge, goals, or preferences—has been a central point of tension in human-computer interaction (HCI) and artificial intelligence. The core challenge has not just been technical, but conceptual: what exactly is a user model, and how deeply must a system simulate a human mind to be genuinely useful?
The Historical Friction of User Modeling
In the early days of intelligent interfaces, the field suffered from a fundamental split. As explored in User Models and User Interfaces: A Case for Domain Models, Task Models, and Tailorability (AAAI 1990), AI researchers viewed explicit user models as indispensable for sophisticated tasks like natural language understanding, whereas interface designers argued that clean, intuitive system design could bypass the need to model a user's internal cognitive state entirely. Because the definition of a user model remained ambiguous, progress was highly fragmented across isolated AI subfields, a problem highlighted in Vive la difference! Individualised interaction with users (ijcai 1995).
This fragmentation was exacerbated by what the authors of User Modeling and User Interfaces (AAAI 1990) termed the "Conversation Metaphor." Early systems tried to treat human-computer interaction like human-to-human dialogue. To make this work, systems attempted to model the user's internal cognitive processes and hidden states.
However, predicting and tracking a user's exact cognitive state is often computationally intractable. This led to a realization: trying to build a perfect psychological simulator of a user inside a single application was a dead end. Instead, early critics suggested focusing on rigid domain and task models, leaving the user model fragmented or entirely unaddressed.
The Modern Pivot: Cross-Contextual Generalization
The limitation of early user modeling wasn't just the intractability of cognitive simulation; it was also application isolation. A user model built for an email client knew nothing about the user's behavior in a web browser or a spreadsheet. This fragmentation prevented systems from achieving a holistic understanding of a user's overarching goals.
To resolve this, recent architectural shifts focus away from application-specific silos and intractable cognitive simulation toward unified, passive observation. In Creating General User Models from Computer Use (UIST 2025), Shan Rizvi et al. introduce General User Models (GUMs). Instead of trying to explicitly decode a user's internal psychology or hardcoding rules for a specific app, GUMs learn from the continuous stream of unstructured computer interactions across all applications.
[Old Paradigm: Application Silos]
Browser Model ---> Only Browser Data
Email Model ---> Only Email Data
[New Paradigm: General User Models (GUMs)]
Unstructured OS Stream (Apps, Web, Docs) ---> Unified GUM ---> Cross-Contextual Assistance
Why General User Models Work
GUMs succeed where historical attempts stalled by changing the underlying data representation and scope:
- From Structured Rules to Unstructured Synthesis: Instead of requiring hand-crafted domain expertise or rigid stereotypes (as discussed in early frameworks like Vive la difference! Individualised interaction with users (ijcai 1995)), modern GUMs ingest raw, unstructured interaction logs. This sidesteps the need to explicitly define or constrain what a user model "should" look like.
- Bypassing the Conversation Metaphor: Rather than attempting the intractable task of simulating a user's internal cognitive state, GUMs focus on observable context, preferences, and implicit beliefs demonstrated across diverse tasks.
- Cross-Contextual Reasoning: By aggregating data from general computer use rather than a single application, the system can perform holistic reasoning. For example, if a user researches a topic in a browser, a proactive AI assistant utilizing a GUM can anticipate their needs when they open a word processor, bridging the gap that previously isolated application-specific models.
By moving from explicit, application-bound cognitive simulators to flexible, cross-contextual aggregators of digital behavior, the field is realizing the long-held goal of truly individualized human-computer interaction.
Key Foundational Work
User Models and User Interfaces: A Case for Domain Models, Task Models, and Tailorability (1990) — contributed 6 concept elements to this direction.
User Modeling and User Interfaces (1990) — contributed 6 concept elements to this direction.
Vive la difference! Individualised interaction with users (1995) — contributed 3 concept elements to this direction.