Research on human judgement and decision making has documented a widerange of psychological biases in how people perceive, evaluate, and ultimately decide between alternatives. These biases are understood in contrast to normative economic principles of utility and risk, which reduce alternatives to quantitative variables and define axiomatic frameworks to find the optimal choice. Understanding and modelling this incongruity has been the focus of behavioural economics research, which aims to align economic models with human behaviour. Such research has been productive in understanding human decision-making behaviour: from investment decisions, to consumer purchases, job prospects, and academic performance – but there has been little investigation in its application to human–computer interaction. Users of interactive systems encounter decisions that share many properties with those studied in the behavioural economics literature. In particular, users often choose to invest actions and effort into using an interface, anticipating that they will receive a return of at least commensurate value in productivity. Models from behavioural economics have the potential to describe how users value such interfaces, and how they decide between them. For example, when faced with a text-entry interface that automatically replaces incorrect words, such models could predict the utility of instances where a correct replacement happens, where an incorrect replacement happens, and the influence of these events on the preferences held by the user. These are important issues for the designers of such interfaces, but are currently poorly understood and under-investigated. This thesis adapts a model of reference-dependent preferences from the behavioural economics literature to interaction. The model predicts the relationship between the outcomes that a user experiences and their evaluation of those outcomes via measures of utility. In particular, the model emphasises the importance of salient positive progress towards a task goal. Some of the model’s predictions were initially tested in two experiments that involved simple text-selection tasks using either a conventional letterby- letter selection technique, or a technique that attempted to assist the user by snapping their selection to word boundaries. The first experiment found a negativity bias: small components of negative progress (when the attempted assistance failed) overwhelmed subjects’ assessment of overall utility, and substantial objective performance (time) gains were required to overcome this assessment. The second experiment found that this effect was neutralised by manipulating the interface’s behaviour to appear more helpful – even though it contained the same performance disadvantages. A new task (drag-anddrop) was developed for a third experiment that extended this manipulation, and found a positivity bias: subjects preferred small elements of positive progress, despite substantial objective performance losses. The methodology developed for these experiments is based on subjects making binary choices between an experimental interface and a neutral reference condition. From a series of these binary choices, a utility/preference scale can be constructed to identify biases in subject preferences against a manipulated objective variable. The flexibility of this method for analysing interactive choices is demonstrated in a fourth experiment, which examined two psychological biases that are not predicted by the model – the peak-end rule and duration neglect – and found some support for their presence during interactive experiences. The main contribution of this thesis is the demonstration of a connection between human–computer interaction research and the behavioural economics literature: that the psychological biases and economic principles in prior work are present and applicable to interactive tasks. The model and experimental methodology provides a robust foundation for improving practice in human–computer interaction work on understanding user behaviour and experience.