What theory is for: the third wave, experience, and the contemporary theory debate

Year 3 / Module 17 Theories of interaction / Unit 10 / What theory is for: the third wave, experience, and the contemporary theory debate

What theory is for: the third wave, experience, and the contemporary theory debate

HCI3101 · Unit 10 of 10

The third wave broadened what a theory of interaction must explain to include meaning, emotion and culture without refuting the usability question, and the live disagreement in the field is not which school is right but whether a theory's function is to predict, to explain, or to generate design.

What theory is for: the third wave, experience, and the contemporary theory debate

A theory of interaction is a claim about where the explanation of an interaction lives, and about what the claim is for. The field disagrees about that second question, and it is more useful than the succession of schools it is usually told as. This tenth of ten units, the last, argues that the third wave widened what a theory of interaction must explain, bringing meaning, emotion and culture inside the subject, and asks what the nine before it were for. It did not refute the older usability question. The live dispute is not which school is right but whether a theory's function is to predict, to explain or to generate designs.

Three questions have to be kept apart, because the literature runs them together. The first is about subject matter: what an account of interaction is an account of. The second is about function: what such an account is for. The third is about warrant: what would count as evidence for it. The third wave settled part of the first and left the second open.

The third wave widened the subject without retiring the question

Bodker (2006) surveys second-wave theory as it met the third wave. In the third wave, she writes, use contexts and application types broaden relative to the second wave's focus on work. Technology spreads from the workplace into homes, everyday life and culture. Multiplicity, context, boundaries, experience and participation come to the fore. The widening is offered as an addition to the second wave, not a replacement of it.

Ten years later Bodker (2015) worries that the third wave threw out the baby with the bath water. It risked discarding the cooperation, learning and participation that second-wave work had built. Her conclusion is a proposal, not a verdict. Sharing through many common artefacts is itself how people participate. The field will not settle until it embraces whole lives, which include the older user and the reader working in a second language, rather than the dichotomy between work and its negation.

The experience turn made meaning and emotion the object

McCarthy and Wright (2004) argue that any account of user experience must take in the emotional, intellectual and sensual aspects of interaction. People do not merely use technology, they live with it. Their framework builds on the pragmatism of Dewey and the dialogism of Bakhtin. It treats felt experience as the thing to be analysed, rather than as a residue left once usability is measured. The move changes the object, not the method.

Bannon (2011) gives the shift a longer arc. HCI, he argues, has moved from evaluating interfaces, through designing systems, into making sense of a world mediated by computing. The focus is then human activity rather than a user interacting with a machine. Human-centred design, on his account, is a value commitment rather than a label. It places people, their concerns and their values at the front of development.

Three maps of one widening ask different things of theory

Harrison, Tatar and Sengers (2007) divide the field into three paradigms. The first comes from engineering and human factors, the second from cognitive science, treating the person as an information processor. The third, the phenomenological matrix, treats interaction as meaning making in which artefact and context are mutually defining. Paradigms co-exist and a new one does not disprove the old, they argue, but it carries different criteria for design, evaluation and appreciation.

Rogers (2012) charts the same terrain across classical, modern and contemporary periods, and asks what theory has been used for in each. Its role, she finds, has expanded from scientific testing to describing, explaining, critiquing and generating designs. Rogers (2004) finds the imported theories broadened the field's discourse and produced insightful accounts of HCI phenomena but had limited impact on design practice. Together they supply the unit's tension: a theory that broadens the subject alone leaves the question of function open, and one that changes function alone leaves the subject where it was.

A theory carries several powers, not one function

Halverson (2002) asks what CSCW should do with theory. She answers by giving any theory four attributes: descriptive power, rhetorical power, inferential power and application power. Applied to activity theory and distributed cognition, the scheme finds them less different than their partisans imply. Both emphasise cognition, include its social context, and rely on ethnographic data. Her contribution here is the refusal of a single function: a theory can describe a setting well and still say nothing about what to build.

A theory's value lies in the counterfactuals it pumps

Beaudouin-Lafon (2000) proposes instrumental interaction, extending direct manipulation. Interfaces become domain objects and interaction instruments, and the model defines a design space with properties for comparing techniques. It compares existing techniques and designs a text-search interface in which an instrument acts on other instruments. Oulasvirta and Hornbaek (2021) give the general account. The practical value of a theory, they argue, lies in counterfactual thinking: theories are speculation pumps that produce thought experiments of the form, if the design were like this, interaction would be like that. Their example concerns typing by speech in a car, where the consequent is that the driver's gaze stays on the road. A theory earns its place by making that speculation valid, and must direct it towards outcomes that are desirable and reliable. That function is neither prediction nor explanation, and it is the strongest contemporary answer to Rogers's (2004) complaint.

The object is still unsettled: what is interaction?

If a theory is to do any of these things it should say what interaction is, and here the field is least settled. Hornbaek and Oulasvirta (2017) begin from the observation that the term is field-defining yet confused, with few direct definitions. They recover the concepts in circulation, among them interaction as dialogue, transmission, optimal behaviour, embodiment and tool use. These concepts carry different scopes and construals of the human-computer causal relation. The paper ends with desiderata rather than a definition.

Contested question: does theory predict, explain, or generate design?

The positions can be stated cleanly. On the first, theory is an instrument of warrant, existing to make a checkable claim. The classical and modern periods measured it against performance at the interface (Rogers, 2012). On the second, theory is an instrument of interpretation: it makes a setting intelligible and supplies its own criteria. The third paradigm asks for this (Harrison, Tatar and Sengers, 2007), as does the experience turn when it makes felt meaning the object (McCarthy and Wright, 2004). On the third, theory is an instrument of generation, its value lying in the counterfactual thought experiments it pumps into design (Oulasvirta and Hornbaek, 2021). An interaction model with its own design space is the working example (Beaudouin-Lafon, 2000).

The disagreement is easily mistaken for a quarrel about schools. If a theory has four powers, then predict, explain and generate are not rival answers to one question but answers to different questions. A theory may be strong on one axis and silent on another. So read, the third wave did not defeat the second: it added functions and left the predictive ambition intact but narrower. That is why this module's earlier units still earn their place.

What would settle it is a comparison the literature has not run. Take one design decision. Ask each account for its answer before anything is built, and record which answers can be checked and which only make sense after the fact. Beaudouin-Lafon's (2000) model is the fair test case, since it both generates techniques and supplies properties for comparing them, so the design function can be scored rather than asserted. The complementary test is Rogers's (2004): whether a theory changed the practice of interaction design at all, a historical question with a checkable answer.

The evidence is argument rather than test on all three sides. Rogers (2012) and Harrison and colleagues (2007) synthesise and classify, Beaudouin-Lafon (2000) demonstrates a model in use, and Halverson (2002) compares two frameworks. The only validation study in the unit tested eighteen guidelines with forty-nine practitioners: evidence about guidance, not about what a theory is for. Fang, Yatani and Hornbaek (2026) supply the count the argument has lacked. Reviewing 615 CHI papers from 2015 to 2024 that use the term framework, they found more enthusiasm for proposing frameworks than for iterating on existing ones, and systematic validation rare: this unit's objection, measured. While the dispute is open, a designer should put one decision to each account before anything is built, recording which answers could be checked in advance and which only make sense afterwards. The strongest objection is that the three functions may not be separable, since a theory that generates a design is already predicting.

Where the argument is heading

The live territory for the question of theory's function is interaction with systems that learn. Amershi and colleagues (2019) consolidate eighteen design guidelines for human-AI interaction, validated with forty-nine practitioners against twenty AI-infused products, so the guidance comes from practice rather than from a model of the user. Yang and colleagues (2020) ask whether human-AI interaction is uniquely difficult to design, and locate the difficulty in the unpredictable errors AI systems make, which damage experience in ways ordinary system errors do not. Inie and colleagues (2026) review eighty HCI papers on critical thinking with or about AI and find the term rarely defined and measured inconsistently. The pattern is the older one: a new subject matter arrives, guidance precedes any account of it, and the question of what the guidance is a theory of is left for later.

The sources this unit rests on

Amershi, S. et al. (2019). Guidelines for Human-AI Interaction. Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3290605.3300233 Proposes eighteen design guidelines for human-AI interaction, validated with forty-nine practitioners who tested them against twenty AI-infused products. Evidence that design guidance can precede a theory.
Bannon, L. (2011). Reimagining HCI: toward a more human-centred perspective. Interactions 18(4), 50-57. doi:10.1145/1978822.1978833 Argues HCI has moved from evaluating interfaces into making sense of a world mediated by computing, and calls for an agenda giving primacy to human actors, values and activities. The unit's normative thread.
Beaudouin-Lafon, M. (2000). Instrumental interaction: an interaction model for designing post-WIMP user interfaces. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/332040.332473 Introduces instrumental interaction, extending direct manipulation, describing interfaces as domain objects and interaction instruments, and defining a design space with properties for comparing techniques. Theory that generates design.
Bodker, S. (2006). When second wave HCI meets third wave challenges. Proceedings of the 4th Nordic Conference on Human-Computer Interaction. doi:10.1145/1182475.1182476 Surveys second-wave theory facing the third wave, in which use contexts broaden, technology spreads into homes and culture, and topics such as multiplicity, context, boundaries, experience and participation come to the fore. It named the third wave.
Bodker, S. (2015). Third-wave HCI, 10 years later: participation and sharing. Interactions 22(5). doi:10.1145/2804405 Reconsiders the third wave ten years on, warns it risked discarding the cooperation, learning and participation built in the second wave, and argues that participation and sharing belong together. The author's own reassessment.
Fang, S., Yatani, K. and Hornbaek, K. (2026). What we talk about when we talk about frameworks in HCI. Proceedings of the CHI Conference on Human Factors in Computing Systems. doi:10.1145/3772318.3791400 Reviews 615 CHI papers from 2015 to 2024 that use the term framework. Enthusiasm for proposing new frameworks exceeds the willingness to iterate on existing ones, and systematic validation is rare.
Halverson, C. A. (2002). Activity Theory and Distributed Cognition: Or What Does CSCW Need to DO with Theories? Computer Supported Cooperative Work (CSCW) 11(1-2), 243-267. doi:10.1023/a:1015298005381 Compares activity theory and distributed cognition by proposing that theories carry descriptive, rhetorical, inferential and application power, and finds the two less different than partisans imply. Theory has several functions at once.
Harrison, S., Tatar, D. and Sengers, P. (2007). The three paradigms of HCI. In alt.chi at the SIGCHI Conference on Human Factors in Computing Systems. https://people.cs.vt.edu/~srh/Downloads/TheThreeParadigmsofHCI.pdf Divides HCI into human-factors, information-processing and phenomenological paradigms, the third treating interaction as meaning making. Paradigms co-exist, the authors argue, but carry different criteria for design and evaluation.
Hornbaek, K. and Oulasvirta, A. (2017). What Is Interaction? Proceedings of the 2017 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3025453.3025765 Argues the field-defining term interaction is confused and rarely defined, recovers concepts such as dialogue, transmission, embodiment and tool use, and lists desiderata rather than a definition. What the theory is a theory of remains unsettled.
Inie, N., Kristensen, A. M., Misfeldt, M. and Dalsgaard, P. (2026). A Critical Review of Critical Thinking in HCI Research About AI. International Journal of Human-Computer Interaction. doi:10.1080/10447318.2026.2722522 Reviews eighty HCI papers on critical thinking with or about AI, finding the term rarely defined and measured inconsistently. The forward edge, where study precedes specification.
McCarthy, J. and Wright, P. (2004). Technology as Experience. MIT Press. doi:10.7551/mitpress/6687.001.0001 Argues an account of experience must include the emotional, intellectual and sensual aspects of interaction, since people live with technology rather than merely using it. The statement of the experience turn.
Oulasvirta, A. and Hornbaek, K. (2021). Counterfactual Thinking: What Theories Do in Design. International Journal of Human-Computer Interaction 38(1). doi:10.1080/10447318.2021.1925436 Argues a theory's practical value lies in counterfactual thinking, treating theories as speculation pumps yielding thought experiments about what interaction would be like. A function beyond prediction.
Rogers, Y. (2004). New theoretical approaches for human-computer interaction. Annual Review of Information Science and Technology 38(1). doi:10.1002/aris.1440380103 Reviews the theories imported into HCI and finds they broadened the field's discourse but had limited impact on the practice of interaction design. Supplies the measure of design work.
Rogers, Y. (2012). HCI Theory: Classical, Modern, and Contemporary. Synthesis Lectures on Human-Centred Informatics 5(2). doi:10.2200/s00418ed1v01y201205hci014 Charts theoretical developments across classical, modern and contemporary periods and finds theory's role expanded from testing to describing, explaining, critiquing and generating designs. The unit's map of what theory is for.
Yang, Q., Steinfeld, A., Rose, C. and Zimmerman, J. (2020). Re-examining Whether, Why, and How Human-AI Interaction Is Uniquely Difficult to Design. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. doi:10.1145/3313831.3376301 Asks whether human-AI interaction is uniquely difficult to design and locates the difficulty in the unpredictable errors AI systems make, which damage experience. The new territory needs its own account.

Check your understanding & apply it · 4 questions, 1 essay

1A team tells a service board that its framework is theoretically grounded and therefore better than the usability testing the service has run so far. Which finding from the unit bears most directly on that claim?

2A designer asks what would happen if a car interface were operated by speech rather than by touch, and concludes that the driver's gaze would stay on the road. What kind of theoretical move is that?

3A framework describes a setting well and says nothing at all about what to build. What does the unit draw from that state of affairs?

4A review consolidates guidance for systems that learn, validating it with practitioners against popular AI-infused products. What does the unit say about guidance that arrives before an account of its subject?

5Essay A product team at a large public library service has adopted a framework called 'practice theory of search' from a research paper and wants it to guide the redesign of the catalogue. The team argues the framework is 'theoretically grounded' and therefore better than the usability testing they have run so far. Write an assessment for the library's service board. Say what the framework would have to do to count as a theory of interaction in the sense this unit uses, drawing on Rogers (2012) for the roles theory has played in the field and on Halverson (2002) for the several powers a theory can carry. Say which function the team is implicitly claiming, using Rogers (2004) and Hornbaek and Oulasvirta (2017). Set out the rival answers the unit's contested question identifies (Harrison, Tatar and Sengers, 2007; McCarthy and Wright, 2004; Oulasvirta and Hornbaek, 2021; Beaudouin-Lafon, 2000), and say what evidence the board should require before accepting the claim. Close by saying how the case would look different if the catalogue recommended results learned from use (Amershi et al., 2019; Yang et al., 2020).

The assessment should begin by separating three questions the board's brief runs together: what the framework is a theory of, what it is for, and what would count as evidence for it. On the first, Hornbaek and Oulasvirta (2017) show that even interaction itself is rarely defined directly, so the team should be asked what practice the framework describes and what it leaves as noise. On the second, Rogers (2012) records that theory in HCI has been used for testing, and also for describing, explaining, critiquing and generating designs, so 'theoretically grounded' names no particular function. Halverson's (2002) four attributes give the board a workable test, since the framework can be scored separately on descriptive, rhetorical, inferential and application power rather than accepted or rejected as a package. The team's claim that the framework beats usability testing implicitly asserts application power, that the theory tells designers what to build, which is exactly the claim Rogers (2004) found the imported theories largely failed to substantiate in practice. The contested question then sets out the rival answers: theory might make the setting intelligible and supply its own criteria (Harrison, Tatar and Sengers, 2007; McCarthy and Wright, 2004), or it might generate design through counterfactual thought experiments about what interaction would be like (Oulasvirta and Hornbaek, 2021), as Beaudouin-Lafon's (2000) interaction model does by defining a design space with properties for comparing techniques. The evidence the board should require follows from Halverson's scheme. Let the team name one design decision in the catalogue, state in advance what the framework predicts the change will do, and record whether the outcome matched; and additionally test whether the framework generated options the old process did not. A learned catalogue changes the target: the system's errors are unpredictable and damage experience in ways ordinary system errors do not (Yang et al., 2020), and design guidance for such systems has largely preceded any theory of them (Amershi et al., 2019), so the board should treat the framework's predictive claims as weaker, not stronger, in that setting.