The information-processing bet: predicting the routine before the prototype

Card, Moran and Newell made an unusual promise for a field that learned by building things. In The Psychology of Human-Computer Interaction (1983) they set out the model human processor. They argued that a description of an interface plus a model of the user yields a quantitative prediction of performance. That prediction arrives before the interface exists. This first of ten units argues that the bet earns its place because it is checkable, and the nine that follow each take one of its limits in turn. The same commitment confines the tradition to the skilled, uninterrupted, single-user task it was built for. Its durable contribution is a way of pricing routine interaction, not a theory of interaction.

The bet: a number before the artefact

The founding claim is that interaction is tractable because the human side of it is a processor with measurable constants. Card, Moran and Newell (1983) treated the person as an information-processing system with perceptual, cognitive and motor stages. The interface was a description from which a method could be read. Together the two gave a prediction of time on task before a prototype existed. Newell and Card (1985) stated the ambition: psychology should give HCI an engineering-grade science, because a technical field tends to drive out softer ones. They named the objections critics would raise. The result would be too low level, too limited in scope, too late and too difficult to apply.

The keystroke-level model sums operator times into a prediction

The clearest instrument is the keystroke-level model (Card and colleagues, 1980). It predicts the time a practised user takes on a given task in a given system. The analyst codes the method into a few operators and sums their times. The operators are K for a keystroke, P for pointing, H for homing the hands, D for drawing, M for mental preparation and R for the system's response. The authors gave standard times, among them 0.20 seconds for a keystroke by an average skilled typist, 1.1 seconds for pointing and 1.35 seconds for a mental operation, and five rules for placing the mental operators. They also coded the replacement of a five-letter word one line below the last edit on two text editors: 4 mental preparations and 15 keystrokes, 8.4 seconds, on the line-oriented system, against 2 mental preparations, 8 keystrokes, 2 homing moves and 1 pointing action, 6.2 seconds, on the display-based one. Against data from 10 systems the model's prediction error was 21 per cent for individual tasks. A screen reader is outside the operator set.

The model is exact for a handed method, and the family disagrees

The model prices the execution of a method the analyst supplies, for an expert, without error. It says nothing about acquisition, choice between methods or recovery. Its authors noted that up to a quarter of an expert's time can go on correcting errors. They excluded those tasks.

John and Kieras (1996) applied four variants, the keystroke-level model, the original GOMS formulation, NGOMSL and CPM-GOMS, to one text-editing task. They differ in the assumptions each makes about unobservable perceptual and cognitive operations, and so in what they predict: scoring the same steps, the parallel CPM-GOMS model gives 2.21 seconds where NGOMSL gives 6.18 seconds. A family of techniques with different assumptions is a different object from a shared theory.

A prediction made before the data held

Gray, John and Atwood (1993) built GOMS models of telephone toll and assistance operators on two workstations, and ran a four-month trial that collected data on 72,450 calls. The models predicted that the proposed new workstation would be slower than the incumbent, contrary to the expectation of those commissioning it. The field data confirmed the prediction. The difference varied with call type, and performance on the new workstation stabilised after the first month. The increase in time was worth almost $2 million a year. The models predicted the first two results and explained all three, and the prediction had been made before the outcome was known.

The boundary: the model is only as good as the method it is handed

The keystroke-level model returns an exact number for the method it is given. Bhavnani and John (1996) tested whether that condition holds when the method is under the user's control. In their field study of computer-aided drafting, even experienced users relied on suboptimal strategies resembling manual drafting. Their example is a figure of three identical arches: efficient work draws one shape's elements, groups them and copies the group, a sequence they name detail-aggregate-manipulate. Their users instead worked operation by operation across the whole drawing, an order inherited from manual drafting, where the compass and the risk of smudging fixed the sequence. The strategic knowledge for using the tool well, they argue, is neither defined nor taught, and because the drawings still come out clean the inefficiency is never noticed. The model can be exactly right about a procedure no efficient user follows, and silent about the person who would have been faster.

The tradition's own review records the limits it cannot pass

The most candid limits come from inside. Olson and Olson (1990) confirmed the tradition's cumulative record: the basic parameters held and the framework extended to error, learning and concurrent processes. The models do not extend to the broader context of computer use. They fail to capture fatigue, individual differences or mental workload. They cannot decide which functions software should contain, judge acceptability, or predict the change in working life a system brings.

The last defences are modest and specific. John (2011) asked whether predictive models, built to compare designs on execution time, can also yield design recommendations. Novice modellers using CogTool reached a 34 per cent Any-Two Agreement in their recommendations, in the upper quartile of studies of the evaluator effect. That effect is the finding that different people using one method find different problems. The paper argues that model-grounded recommendations may still affect measurable performance more reliably. Barnard and May (2000) separately argued that task analysis had become too narrow for knowledge-intensive, integrated work and proposed an analysis of broader scope.

Contested question: is routine performance a plan or an assembly?

The literature disagrees about what the routine in "routine task" is. One position treats it as a compiled method an expert executes faithfully. Card, Moran and Newell (1980) were explicit that if the method is unknown no reasonable prediction is possible. The proper problem, they wrote, is to predict the time given the method. Olson and Olson (1990) reported that a decade of coordinated work confirmed the basic parameters and extended them.

The other position holds that the routine is assembled in the moment from what the artefact allows. Gray and Boehm-Davis (2000) introduced microstrategies. These are the limited ways elementary perceptual, cognitive and motor operations can be combined for an interface, such as moving to a button and clicking it. Two studies found that users deploy alternative microstrategies that shave milliseconds from routine behaviour. The choice is sensitive to small interface features, and millisecond-level analysis can inform design.

The positions are closer than they look. Microstrategies document variation within the range the keystroke-level model already tolerates, so they qualify it without refuting it. Conversely, the model's "given method" is exactly the simplification the microstrategies expose. What would settle the dispute is a comparison the literature has not run. On one interface, measure how much of the between-user variance is explained by microstrategy choice against method choice. Test whether the model's fixed operator times still bound the range when the interface invites different microstrategies. Gray and Boehm-Davis (2000) supply the apparatus. The head-to-head test against a GOMS model is missing, and its absence is the finding.

The evidence on the two sides differs in kind. The compiled-method side rests on parameters checked against ten systems and one prospective field test, where the prediction was made before the data and held. The microstrategy side rests on two laboratory studies of millisecond variation, with no field deployment and no head-to-head test against a GOMS model. While the dispute is open, a designer should compare only methods already fixed. A predicted advantage below the model's own error stays unproven. The strongest objection is to the unit's confidence that checkability saves the bet. The analyst chooses the method after seeing the task, so a poor prediction is repaired by re-coding the method. John and Kieras (1996) report 2.21 and 6.18 seconds for the same steps.

Where the argument is heading

Oulasvirta, Kristensson, Bi and Howes (2018) apply abstraction, automation and analysis to generate interfaces, using optimisation and machine learning to find designs optimal against an objective. The analyst who sums operator times by hand gives way to a fit-and-optimise loop, and the commitment to a checkable prediction survives. Salvucci (2009) shows the method reaching a case the serial model excluded, integrating a driver-behaviour model with task models to predict distraction in vehicle interfaces. Mozannar, Bansal, Fourney and Horvitz (2024) carry the same accounting into AI-assisted programming. Coding 21 programmers' Copilot sessions, they found verifying suggestions the costliest activity, at 34.3 per cent of session time, with more than half of task time going on Copilot-related work. The routine being priced is produced partly by the system, so the method is no longer the analyst's to fix. The bet survives as a way of putting a price on routine work, but it no longer holds one method still.

The sources this unit rests on

• Card, S. K., Moran, T. P. and Newell, A. (1980). The keystroke-level model for user performance time with interactive systems. Communications of the ACM 23(7). doi:10.1145/358886.358895 Predicts an expert's task time by coding the method into operators K, P, H, D, M and R and summing their times. Against 10 systems its error was 21 per cent for individual tasks.
• Card, S. K., Moran, T. P. and Newell, A. (1983). The Psychology of Human-Computer Interaction. Lawrence Erlbaum. doi:10.1201/9780203736166 The book that founded the information-processing programme, setting out the model human processor and the claim that a description of an interface plus a model of the user yields a prediction before the interface is built.
• Newell, A. and Card, S. K. (1985). The Prospects for Psychological Science in Human-Computer Interaction. Human-Computer Interaction 1(3). doi:10.1207/s15327051hci0103_1 Argues that psychology should give HCI an engineering-grade science and names the objections critics would raise: the resulting knowledge would be too low level, too limited in scope, too late to affect technology and too difficult to apply.
• John, B. E. and Kieras, D. E. (1996). The GOMS family of user interface analysis techniques: comparison and contrast. ACM Transactions on Computer-Human Interaction 3(4). doi:10.1145/235833.236054 Applies four GOMS variants, the keystroke-level model, the original formulation, NGOMSL and CPM-GOMS, to one text-editing task. They differ in architectural assumptions and predictive power, estimating 2.21 against 6.18 seconds for the same steps and 14.38 against 16.38 seconds for the full text-editing task.
• Gray, W., John, B. and Atwood, M. (1993). Project Ernestine: Validating a GOMS Analysis for Predicting and Explaining Real-World Task Performance. Human-Computer Interaction 8(3). doi:10.1207/s15327051hci0803_3 Built GOMS models for telephone operators on two workstations and tested them in a four-month trial of 72,450 calls. The models predicted the proposed workstation would be slower and the data confirmed it, at a cost of almost $2 million a year.
• Gray, W. D. and Boehm-Davis, D. A. (2000). Milliseconds matter: an introduction to microstrategies and to their use in describing and predicting interactive behaviour. Journal of Experimental Psychology: Applied 6(4). doi:10.1037/1076-898x.6.4.322 Introduces microstrategies, the limited ways elementary perceptual, cognitive and motor operations can be combined for a given artefact. Two studies found users deploy alternatives that shave milliseconds from routine behaviour, sensitive to small interface features.
• Olson, J. R. and Olson, G. (1990). The growth of cognitive modelling in human-computer interaction since GOMS. Human-Computer Interaction 5(2). doi:10.1207/s15327051hci0502&3_4 Reviews a decade of cognitive modelling after GOMS, reporting that coordinated work confirmed and extended the basic parameters, with formal modelling accounting for errors and learning time. It limits the approach: the models miss context, fatigue, individual differences and mental workload.
• Bhavnani, S. K. and John, B. E. (1996). Exploring the unrealized potential of computer-aided drafting. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/238386.238538 Analysed real-world computer-aided drafting and found that experienced users rely on suboptimal strategies resembling manual drafting. It argues the strategic knowledge for using the tool well is neither defined nor taught, so the model is exact only for the method it is given.
• John, B. E. (2011). Using predictive human performance models to inspire and support UI design recommendations. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. doi:10.1145/1978942.1979088 Asks whether novice modellers can extract design recommendations from predictive models, which had mainly compared designs on execution time. They reach a 34 per cent Any-Two Agreement, in the upper quartile of evaluator-effect studies, while the paper argues model-grounded recommendations may affect performance more reliably.
• Mozannar, H., Bansal, G., Fourney, A. and Horvitz, E. (2024). Proceedings of the CHI Conference on Human Factors in Computing Systems. doi:10.1145/3613904.3641936 Coded 21 programmers' Copilot sessions against a taxonomy of activities. Verifying suggestions was the costliest state, at 34.3 per cent of session time, with more than half of task time going on Copilot-related work.
• Barnard, P. J. and May, J. (2000). Towards a theory-based form of cognitive task analysis of broad scope and complexity. In Schraagen, J. M., Chipman, S. F. and Shalin, V. L. (eds), Cognitive Task Analysis. Lawrence Erlbaum. doi:10.4324/9781410605795-20 Argues that task analysis rests on decomposing action, tractable for simple technologies but too restricted for intricate, knowledge-intensive work, and is therefore the tradition's own answer to its critics.
• Oulasvirta, A., Kristensson, P. O., Bi, X. and Howes, A. (eds) (2018). Computational Interaction. Oxford University Press. doi:10.1093/oso/9780198799603.001.0001 The textbook statement of computational interaction, which applies abstraction, automation and analysis to model interaction and generate interfaces using optimisation, machine learning, control theory, formal methods and cognitive models. It is the predictive tradition's successor.
• Salvucci, D. D. (2009). Rapid prototyping and evaluation of in-vehicle interfaces. ACM Transactions on Computer-Human Interaction 16(2). doi:10.1145/1534903.1534906 Presents Distract-R, which lets designers prototype in-vehicle interfaces and evaluate them with a cognitive model of driver behaviour, predicting driver performance and distraction in three studies. It reaches into multitasking and safety, the case the serial model excluded.