Adaptation moves the interface, and moving it costs the user

Defenders of adaptivity usually make one argument: a system that changes to fit the person ought to beat one that stays still. The studies that test it test two propositions: changing the level of automation applied to the work, and moving the surface where the controls sit and the hands have learned to reach.

A person who uses an interface repeatedly stops reading it and starts reaching for it, so its controls join the task. Every automatic change invalidates part of that memory, and the cost appears as a slower selection and a longer search. The eighth of ten units, this one prices what adaptation costs a fluent user. The ninth unit asks whether a conversation uses its capacity to ask in order to repair.

Adapting form pays when the target is ability rather than taste

Gajos, Wobbrock and Weld (2007) generated an interface optimised for each user's motor capability and evaluated it with motor-impaired users. One participant completed tasks the standard interface had made impossible. The rest averaged a 20 per cent time saving, from a 3 per cent slowdown to a 43 per cent speedup. The study is small and preliminary, and it adapts once, into a stable arrangement. Motor capability also changes with age, so a model fitted once captures an older user's hand only on the day it was built.

SUPPLE (Gajos and Weld, 2004), the system behind the ability-based interfaces, shows what adapting form actually is. It takes a functional description of an interface, a model of the device and a model of the person's typical activity, then searches for the rendering that meets the device's constraints at the least estimated effort, treating interface generation as an optimisation problem. It is one research system, and its behaviour when the capability model is wrong is not established.

Moving the surface under a learned control costs the user time

Findlater and McGrenere (2004) compared static, adaptable and adaptive split menus with 27 participants. The static menu was significantly faster than the adaptive one, which reordered itself around recent use, and the adaptable menu, rearranged by the participant, was significantly faster under certain conditions, with most participants preferring it.

Gajos, Czerwinski, Tan and Weld (2006) show why. The team compared three adaptive interfaces on Microsoft Word with a non-adaptive baseline in two experiments, with 26 and eight participants. The Split Interface copied hard-to-reach functions into a separate toolbar and beat the baseline on performance and satisfaction, while the Moving Interface moved the same functions and carried the higher theoretical benefit yet did not significantly improve either.

Findlater and McGrenere (2007) separate feature findability, whether a person can locate a feature they want, from feature awareness, whether they know it exists. Findability was better in a minimal layered interface, but awareness of advanced features was better for participants who started with the full one.

Gajos and colleagues (2006) drew the dimensions from their own and earlier results: spatial stability, locality, predictability, the frequency of adaptation and the accuracy of the mechanism. Adaptation that happened once per session had produced strongly positive results; the same interface adapting at every interaction did worse than a non-adaptive baseline.

What survives is the form the person triggers or approves

McGrenere, Baecker and Booth (2007) ran a field study comparing automatic adaptive menus with a prototype whose simple interface the user customised. Most users personalised effectively, and user-controlled adaptation gave better navigation and learnability.

Bunt, Conati and McGrenere (2009) name the middle position: a mixed-initiative system recommends a customisation and leaves the decision to the user, which resolves the dilemma between an adaptable interface that burdens the user and an adaptive one that surprises them.

Timing an interruption is cheap, and relocating a control is not

If moving the surface is expensive, changing the timing of an interruption is cheap. Bailey and Konstan (2006) ran a controlled experiment with 50 participants in which identical peripheral tasks appeared either between primary tasks or during them. Interruption during a task cost between 3 and 27 per cent more time, about twice the errors, between 31 and 106 per cent more annoyance and twice the rise in anxiety.

Iqbal and Bailey (2008) built the system that follows: statistical models of task state detect breakpoints and deliver messages at them. In two studies, breakpoint scheduling reduced frustration and reaction time relative to immediate delivery, and nothing on the screen moved.

Horvitz (1999) framed that decision as a computation. His mixed-initiative position decides when to act by comparing the expected value of acting with the expected cost of being wrong, illustrated with the LookOut scheduling assistant.

The other proposition raises the level of automation, not the surface

Parasuraman, Sheridan and Wickens (2000) classify it by the information-processing stage it acts on, from acquisition through analysis and decision to action implementation. The level should be chosen per stage, because automation changes human activity and adds coordination demands.

Kaber and Endsley (2004) tested adaptive allocation directly, over low, intermediate and high levels allocated at cycle times covering 20, 40 and 60 per cent of task time. The level, not the schedule, drove performance and situation awareness. Intermediate levels raised awareness without improving performance or reducing workload, and the effects were not additive.

Shneiderman (2020) makes the separation explicit: his framework for human-centred artificial intelligence asks for high levels of human control and high levels of computer automation at once.

The contested question: design flaw or property of the thing?

The first position holds that the cost is a design choice: automatic adaptation pays when it is stable, infrequent and accurate. Gajos and colleagues (2006) locate an adaptive interface's success in spatial stability rather than in adaptation itself, and Gajos, Wobbrock and Weld (2007) measure a benefit for ability-based adaptation, while McGrenere and colleagues (2007) and Bunt and colleagues (2009) show that the advisory and user-triggered forms avoid the costs the automatic form incurs.

The second holds that the cost is inherent to automatic change, and that a stable interface with user control wins wherever work is repeated. Findlater and McGrenere (2004) found the static menu faster than the adaptive one, and Gajos and colleagues (2006) found that their Moving Interface did not beat the baseline, while reliance on imperfect automation must still be calibrated against predictable behaviour (Lee and See, 2004).

The studies reporting success mostly adapt in a designated region, once, from a measurement, or in time rather than space. The studies reporting failure move the parts the person uses by habit. What would settle it is a design neither side has run: a field deployment rather than a laboratory session, in which an interface adapts only within a stable region and the study measures expertise, so that it controls motor memory rather than assuming it.

The design-flaw position rests on laboratory experiments and a field study (Gajos et al., 2006; Gajos, Wobbrock and Weld, 2007; McGrenere et al., 2007); the inherent-cost position rests on one controlled comparison (Findlater and McGrenere, 2004) and a review of trust (Lee and See, 2004). Only one postdates 2010, the 2023 replication of the measured cost (Gaspar-Figueiredo et al., 2023). While the dispute is open, adapt the advice and the timing of a change, and move the surface only where no habit exists. The strongest objection is that the central comparison has no modern replication.

Where this is heading

The direction is towards designs that automate heavily and keep control visibly with the person. Shneiderman's (2020) framework states the goal. The open questions are the predictability and frequency of an adaptive mechanism (Gajos et al., 2006) and appropriate reliance when automation is imperfect (Lee and See, 2004). The strongest interface-adaptation studies are laboratory experiments, and most predate 2010. The finding that automatic relayout loses to a stable surface still has no modern replication, but the measured cost of moving a menu now has one: Gaspar-Figueiredo and colleagues (2023) replicated, with 40 participants and brain-activity measures, that different adaptive menus produce measurably different cognitive load.

Sources

• Horvitz, E. (1999). Principles of mixed-initiative user interfaces. Proceedings of the SIGCHI conference on Human factors in computing systems. doi:10.1145/302979.303030 Sets out a decision-theoretic framework in which the system acts when the expected value of acting exceeds the expected cost of being wrong, illustrated with the LookOut scheduling assistant. A position statement on one prototype rather than a controlled evaluation, it supplies the rule for timing adaptation rather than form.
• Gajos, K. Z., Wobbrock, J. O. and Weld, D. S. (2007). Automatically generating user interfaces adapted to users' motor and vision capabilities. Proceedings of the 20th annual ACM symposium on User interface software and technology. doi:10.1145/1294211.1294253 Generated an interface optimised for each user's motor capability and evaluated it with motor-impaired users. One participant did tasks the standard interface had made impossible, and the rest averaged a 20 per cent time saving, from a 3 per cent slowdown to a 43 per cent speedup. The unit's measured case for adapting form.
• Gajos, K. and Weld, D. S. (2004). SUPPLE: automatically generating user interfaces. Proceedings of the 9th international conference on Intelligent user interfaces. doi:10.1145/964442.964461 Generated an interface by optimising a functional specification against a device model and a model of the user's activity, treating rendering as an optimisation problem rather than a rule set. The machinery behind the later ability-based interfaces.
• Gajos, K. Z., Czerwinski, M., Tan, D. S. and Weld, D. S. (2006). Exploring the design space for adaptive graphical user interfaces. Proceedings of the working conference on Advanced visual interfaces. doi:10.1145/1133265.1133306 Built three adaptive interfaces on Microsoft Word and compared them with a non-adaptive baseline in two experiments, with 26 and eight participants. The spatially stable Split Interface beat the baseline on performance and satisfaction, while the Moving Interface, which carried the higher theoretical benefit, did not improve either.
• Findlater, L. and McGrenere, J. (2004). A comparison of static, adaptive, and adaptable menus. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/985692.985704 Compared static, adaptable and adaptive split menus with 27 participants. The static menu was significantly faster than the adaptive menu, the adaptable menu was significantly faster than the adaptive one under certain conditions, and most participants preferred the adaptable menu.
• Findlater, L. and McGrenere, J. (2007). Evaluating reduced-functionality interfaces according to feature findability and awareness. INTERACT 2007, Lecture Notes in Computer Science. doi:10.1007/978-3-540-74796-3_59 Separated feature findability from feature awareness and compared layered reduced-functionality interfaces with a full one. Findability was better in the minimal layered interface, but awareness of advanced features was better for participants who started with the full one.
• McGrenere, J., Baecker, R. M. and Booth, K. S. (2007). A field evaluation of an adaptable two-interface design for feature-rich software. ACM Transactions on Computer-Human Interaction. doi:10.1145/1229855.1229858 Ran a field experiment with a commercial word processor, comparing automatic adaptive menus with a prototype whose simple interface the user customised. Most users personalised effectively, and user-controlled adaptation gave better navigation and learnability than the system-controlled menus.
• Bunt, A., Conati, C. and McGrenere, J. (2009). Mixed-initiative interface personalization as a case study in usable AI. AI Magazine. doi:10.1609/aimag.v30i4.2264 Presents a mixed-initiative system that recommends a personalisation and leaves the decision to the user, addressing obtrusiveness and control. It is offered as the middle ground between an adaptable interface that burdens the user and an adaptive one that surprises them.
• Shneiderman, B. (2020). Human-centred artificial intelligence: reliable, safe and trustworthy. International Journal of Human-Computer Interaction. doi:10.1080/10447318.2020.1741118 Argues for designs that hold high levels of human control and high levels of computer automation at once, treating excessive control and excessive automation as twin dangers. A framework rather than a study, it is cited for the separation, not for an effect.
• Parasuraman, R., Sheridan, T. B. and Wickens, C. D. (2000). A model for types and levels of human interaction with automation. IEEE Transactions on Systems, Man, and Cybernetics, Part A. doi:10.1109/3468.844354 Classifies automation by the information-processing stage it acts on, from acquisition through analysis and decision to action, and by the level at each stage. The level should be chosen per stage, because automation changes human activity and imposes new coordination demands.
• Kaber, D. B. and Endsley, M. R. (2004). The effects of level of automation and adaptive automation on human performance, situation awareness and workload in a dynamic control task. Theoretical Issues in Ergonomics Science. doi:10.1080/1463922021000054335 Ran a dual-task experiment over low, intermediate and high levels of automation allocated at cycle times covering 20, 40 and 60 per cent of task time. The level, not the schedule, drove performance and situation awareness, and the effects were not additive.
• Bailey, B. P. and Konstan, J. A. (2006). On the need for attention-aware systems: measuring effects of interruption on task performance, error rate, and affective state. Computers in Human Behaviour. doi:10.1016/j.chb.2005.12.009 Ran a controlled experiment with 50 participants in which identical peripheral tasks appeared either between primary tasks or during them. Interruption during a task cost between 3 and 27 per cent more time, about twice the errors, between 31 and 106 per cent more annoyance and twice the rise in anxiety.
• Iqbal, S. T. and Bailey, B. P. (2008). Effects of intelligent notification management on users and their tasks. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/1357054.1357070 Built a notification manager that uses statistical models of task state to deliver messages at breakpoints, in two studies. Breakpoint scheduling reduced frustration and reaction time relative to immediate delivery, and content relevance determined the right breakpoint.
• Lee, J. D. and See, K. A. (2004). Trust in automation: designing for appropriate reliance. Human Factors. doi:10.1518/hfes.46.1.50.30392 Reviews the evidence on trust and models how context, the automation's characteristics and the user's cognition shape reliance. It argues for designing for appropriate reliance rather than maximum trust, since reliance too low and reliance too high both degrade performance.
• Gaspar-Figueiredo, D., Abrahao, S., Insfran, E. and Vanderdonckt, J. (2023). Measuring User Experience of Adaptive User Interfaces using EEG: A Replication Study. Proceedings of the 27th International Conference on Evaluation and Assessment in Software Engineering. doi:10.1145/3593434.3593452 Replicated an earlier experiment with 40 participants across twenty graphical adaptive menus, measuring cognitive load, engagement, attraction and memorisation by electroencephalogram. Confirmed that different adaptive menus produce measurably different brain responses, with the users' ratings tracking those signals, so it tests the physiological cost of menu design rather than the static-versus-adaptive comparison.