In the previous post, I arrived at a direction that feels more specific than the topic cluster I carried through last semester: adaptive interaction design. I am not treating it as a finished thesis definition. It is a working direction that gives my reading and prototyping somewhere to go.
That shift changes how I look at the literature. Before, I was reading broadly in order to understand interruption, flow, cognitive load, memory, emotion and neuroadaptive interfaces. The breadth was necessary. It helped me see that interruptions are not isolated design failures but events that affect attention, task continuity and a person’s ability to return to what they were doing. Now I need to ask a different question. If an interface is going to adapt to the user, what should it pay attention to, what should it be allowed to change and how should the user remain involved?
The older work on adaptive and mixed-initiative systems has become especially useful here. It is easy to speak about adaptive interfaces as if they are a recent consequence of AI, but many of the central interaction questions are older. Horvitz (1999) wrote about mixed-initiative user interfaces at a time when the examples looked very different from current AI agents. Yet the core problem is familiar: a system may have enough information to make a suggestion or take an action, but it still has to decide whether this is the right moment to involve the user.
That matters because a useful system can also become a badly timed system. It may offer a recommendation while somebody is already working through a difficult task. It may ask for confirmation when the person has no real capacity to evaluate the decision. It may quietly act on the person’s behalf, then leave them to reconstruct what happened later. The design problem is not simply how to make the system more proactive. It is how to distribute initiative without turning the user into a passive observer of their own work.
This connects directly to the research I had already done on interruption. Adamczyk and Bailey (2004) showed that the timing of an interruption changes its effects on performance, emotion and the way people perceive the interrupting system. Their basic point remains important: “user attention is a scarce resource” (p. 271). An interface cannot assume that a message, recommendation or request for approval has the same cost at every point in a task.
Iqbal and Horvitz (2007) make the picture more realistic by looking at interruption and recovery in everyday computing work. Their work shows that disruption is not just the moment somebody looks away from a task. There is a longer process of reorientation afterwards. People need to recover their goal, remember the current state of the work and decide what comes next. This is why I am interested in adaptation not only as a way to prevent badly timed interruptions, but also as a way to support the return after they happen.
The return is where memory becomes a design concern. Altmann and Trafton (2002) describe pending goals in terms of activation and retrieval. In simple terms, an interrupted goal does not wait untouched until a person comes back to it. It has to compete with everything else that happened in the meantime. The task may still be open on the screen, but the reasoning behind it may have faded. A system that changes information, makes suggestions or completes work while the user is away has a responsibility to make the return understandable.
This becomes particularly relevant in AI-assisted work. Recent research on agentic systems is often framed around how much work an agent can perform autonomously. I find the moment of human re-entry more interesting. Zhou et al. (2026) examine when people should check multi-step agentic tasks. Their findings suggest that constant confirmation becomes costly, while waiting until the end can make correction much more expensive. The issue is not whether the human should supervise every action. It is how the system can involve the human at moments that protect both control and continuity.
The same question appears in workload-aware systems. The literature does not support a simple idea of a system knowing exactly when someone is busy. Workload overlaps with attention, task difficulty, stress, effort and fatigue, but it is not interchangeable with any of them. Kosch et al. (2023) show that HCI researchers use self-report, performance, behaviour and physiological measures, each with different strengths and limits. There is no single number that can tell an interface everything it needs to know about a person.
That uncertainty is not a reason to abandon adaptive design. It is a reason to design it carefully. Fairclough’s (2009) work on physiological computing describes a loop in which a system senses something about the user, infers a state, adapts and then observes the result. The word infers is important. Signals from gaze, EEG, typing or task performance are not direct access to a person’s internal life. They are partial evidence that may or may not be useful in context.
I find the interaction logic of this early experimental work more compelling than the promise of any individual sensor. The system does not need to claim that it has discovered a person’s true mental state. It can respond modestly to patterns that suggest a change in task pressure, then let the user inspect or reject that response. Treacy Solovey et al. (2015) argue that physiological input is often better suited to implicit and supportive adaptation than direct control. Their point is not that systems should become invisible. It is that they should not behave with more certainty than their input allows.
This is starting to give my research a clearer shape. I am not looking for the perfect way to measure attention. I am looking at what happens when an interface has incomplete information about the user’s situation and still needs to make decisions about timing, assistance and interruption. What would it mean for that interface to be helpful without becoming intrusive? How could it support recovery without taking ownership away from the person?
The answers will depend on the context. A work tool, a wearable device, a spatial interface and an AI agent all create different conditions. For now, I do not want to lock the research into one medium too early. I want to keep reading across these contexts while holding onto the same underlying question: how can systems adapt to human capacity without asking humans to adapt to the system first?
This post is a research checkpoint rather than a conclusion. The previous post helped me name a direction. This one is helping me understand the problems inside that direction.
References
Adamczyk, P. D., & Bailey, B. P. (2004). If not now, when?: The effects of interruption at different moments within task execution. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 271-278). ACM. https://doi.org/10.1145/985692.985727
Altmann, E. M., & Trafton, J. G. (2002). Memory for goals: An activation-based model. Cognitive Science, 26(1), 39-83. https://doi.org/10.1207/S15516709COG2601_2
Fairclough, S. H. (2009). Fundamentals of physiological computing. Interacting with Computers, 21(1-2), 133-145. https://doi.org/10.1016/j.intcom.2008.10.011
Horvitz, E. (1999). Principles of mixed-initiative user interfaces. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 159-166). ACM. https://doi.org/10.1145/302979.303030
Iqbal, S. T., & Horvitz, E. (2007). Disruption and recovery of computing tasks: Field study, analysis, and directions. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (pp. 677-686). ACM. https://doi.org/10.1145/1240624.1240730
Kosch, T., Karolus, J., Zagermann, J., Reiterer, H., & Schmidt, A. (2023). A survey on measuring cognitive workload in human-computer interaction. ACM Computing Surveys, 55(13s), Article 286, 1-39. https://doi.org/10.1145/3582272
Treacy Solovey, E., Afergan, D., Peck, E. M., Hincks, S. W., & Jacob, R. J. K. (2015). Designing implicit interfaces for physiological computing: Guidelines and lessons learned using fNIRS. ACM Transactions on Computer-Human Interaction, 21(6), Article 35, 1-27. https://doi.org/10.1145/2687926
Zhou, Y., et al. (2026). When should users check? Modeling confirmation frequency in multi-step agentic AI tasks. In Proceedings of the CHI Conference on Human Factors in Computing Systems. ACM.
