Man-Machine Systems

Abstract

A man-machine system is any arrangement in which a human operator and a machine share the work of a task, their functions interrelated and both necessary for the system to meet its goal. It is the organizing idea of human-factors engineering: rather than studying the person or the device alone, it treats the pair as one unit whose performance depends on how sensing, decision, and action are divided between them. This article sets out the field's founding problem — how to allocate functions between people and machines — and traces its development through automation, situation awareness, mental workload, trust, and human error. It shows why adding automation does not simply subtract work, and why the joint system's reliability depends less on the machine's raw capability than on how well its behaviour matches the operator's understanding of it.

Keywords: man-machine systems, human-automation interaction, situation awareness

The phrase sounds dated — it belongs to the era of the control room and the cockpit dial — but the problem it names is more current than ever. Every partially automated car, every clinical alarm, every recommendation the operator can accept or override is a man-machine system, and each raises the same question the field has asked since the 1950s: given a task, which parts should the human do, which should the machine do, and how should the two be joined so that the whole performs better than either alone? The answers have grown more subtle over seventy years, but the enduring lesson is that the join itself — the interface, the allocation of authority, the operator's model of what the machine is doing — is where systems succeed or fail.

What Man-Machine Systems Are

A man-machine system — today more often called a human-machine or human-automation system — is, in the words of the MeSH definition, a system in which the functions of the human and the machine are interrelated and necessary for the operation of the system. The defining commitment is holistic: the object of study is neither the operator's psychology nor the machine's engineering on its own, but the coupling between them. A pilot and an autopilot, a radiologist and a detection algorithm, a driver and a lane-keeping system are each a single system in this sense, and each will perform well or badly according to how the work is shared and how the interface conveys what the machine is doing.

This framing grew out of the human-factors and ergonomics tradition that emerged during and after the Second World War, when it became clear that even well-trained operators made predictable errors with poorly designed equipment — that the fault lay in the coupling, not the person. The field's task is therefore design as much as description: to arrange displays, controls, and automation so that the joint system is efficient, resilient to error, and usable by a real human with finite attention and memory (Norman, 2013). Three concerns recur throughout: how to allocate functions between human and machine, how to keep the operator aware of the system's state, and how to manage the workload and trust that determine whether the human's contribution helps or hinders.

Types of Man-Machine Systems

MeSH files man-machine systems beneath two broader descriptors — Ergonomics, the study of fitting work to the human, and Technology — and lists one narrower topic beneath it in turn. Neither parent nor child yet has a dedicated article on this site. Such a classification is an indexing hierarchy rather than a theory of the field: it records how the literature is catalogued, and its single child is best read as one specialized branch rather than an exhaustive taxonomy of the many kinds of human-machine system. The subareas that actually organize the field — function allocation, automation, situation awareness, trust — cut across it and are treated in their own sections below.

Function Allocation and the Fitts List

The field's founding problem is function allocation: given a task, which functions should be assigned to the human and which to the machine? The first systematic answer was the Fitts list of 1951, which catalogued what humans do better (perceiving patterns, improvising, exercising judgement in novel situations) and what machines do better (fast repetitive computation, applying great force, storing information reliably) — a scheme later nicknamed MABA-MABA, for “men are better at / machines are better at.” The list gave designers a first principled basis for dividing labour rather than automating whatever happened to be technically convenient.

Paul Fitts also gave the field one of its few quantitative laws. Studying the human motor side of the system, he showed that the time to move to and acquire a target is a lawful function of the movement's difficulty: the farther the target and the smaller it is, the longer the movement takes, and the relationship is logarithmic (Fitts, 1954). Fitts's law lets a designer predict how long an operator will take to reach a control or a screen target, and it remains a working tool of interface design, examined in the first demonstration and the worked example below. The Fitts list itself has aged less well as a design method — static allocation ignores that the best division of labour can change with the situation — but as the statement of the field's central question it has never been superseded.

Levels and Stages of Automation

Function allocation is rarely all-or-nothing. Automation can be partial, and it can be applied to different parts of a task, so the modern treatment replaces the binary human-or-machine choice with a two-dimensional scheme. One influential model distinguishes four stages of information processing that automation can support — information acquisition, information analysis, decision selection, and action implementation — and, within each stage, a continuum of levels from fully manual to fully autonomous (Parasuraman et al., 2000). A collision-warning system automates acquisition and analysis but leaves the decision to the driver; an autopilot automates action implementation under human supervision. Describing a system means specifying, stage by stage, how high the automation reaches.

The level chosen has systematic consequences. A meta-analysis integrating the stages-and-levels framework found that higher and later-stage automation reliably reduces operator workload and improves routine performance while the automation works — but that the same design degrades the operator's ability to detect and recover from automation failures, because a person who has been relieved of the decision is poorly placed to take it back (Onnasch et al., 2014). This lumberjack effect — the higher the automation, the harder the fall when it fails — is the central trade-off of the field, and it explains why the taxonomy has itself been the subject of continued scrutiny and refinement over the decades (Vagia et al., 2016). The choice of level is not a matter of using as much automation as the technology allows; it is a design decision about how much of the human's engagement to preserve.

Situation Awareness and Mental Workload

Why does high automation leave the operator unable to recover? The answer lies in two cognitive constructs the field has made precise. The first is situation awareness (SA): the operator's internal model of what is happening, formalized as three levels — perceiving the relevant elements in the environment, comprehending their meaning, and projecting their near-future state (Endsley, 1995). Good decisions depend on good SA, and automation that takes the human out of the control loop starves SA at all three levels: the operator no longer perceives the raw data, no longer builds comprehension by acting, and so cannot project what will happen when the automation quits. This out-of-the-loop problem is the mechanism behind the lumberjack effect, and it is the core of the classic ironies of automation: the more a designer automates, the more the human is left with precisely the monitoring and fault-recovery tasks people perform worst, so that automating away the easy work makes the residual human role harder and more critical, not easier (Bainbridge, 1983).

The second construct is mental workload — the demand a task places on the operator's limited attentional resources. Wickens's multiple-resource theory holds that these resources are not a single pool but several, defined along dimensions such as processing stage and perceptual modality, so that two tasks interfere more when they draw on the same resource than when they draw on different ones (Wickens, 2008). The practical upshot for man-machine systems is that workload can be managed by design — offloading a visual task to the auditory channel, say — and that both overload and underload are hazards: too much demand causes errors of capacity, while too little, as under high automation, causes the vigilance decrements and disengagement that leave an operator unprepared. A well-designed system keeps workload in a moderate band and preserves the situation awareness that recovery requires (Rasmussen, 1983).

Trust, Reliance, and Human Error

Whether an operator actually uses automation, and whether they should, turns on trust. Trust in automation is the attitude that an agent will help achieve one's goals under uncertainty, and it governs reliance: operators lean on automation they trust and ignore automation they do not (Lee & See, 2004). The design goal is not maximal trust but calibrated trust — trust that matches the automation's true reliability. When trust exceeds reliability the result is misuse: the operator over-relies, fails to monitor, and is caught out by failures, the automation-induced complacency documented across aviation and process control (Parasuraman & Riley, 1997). When trust falls short of reliability the result is disuse: the operator ignores or disables automation that would have helped, as happens with alarms that cry wolf too often.

These reliance failures are one strand of the broader problem of human error, which the field treats not as operator fault but as a systematic product of the design. Reason's taxonomy distinguishes slips and lapses — failures in executing a good plan — from mistakes, failures in the plan itself, and locates the origins of many errors in latent conditions built into the system long before the operator arrived (Reason, 1990). Rasmussen's complementary skill-rule-knowledge framework ties error types to the level at which behaviour is controlled, from automatic skilled action through rule-following to effortful problem-solving (Rasmussen, 1983). Both frameworks carry the same design implication: because errors are predictable consequences of how the joint system is arranged, they are reduced by redesigning the coupling — better feedback, error-tolerant interfaces, appropriate automation — rather than by exhorting operators to try harder (Sheridan & Parasuraman, 2005).

Interactive demonstrations

The three demonstrations below let the reader work with core ideas of the field. The first applies Fitts's law to a pointing task, showing how movement time grows with the difficulty of the movement; the second moves a task across the stages and levels of automation and traces the workload–awareness trade-off; the third shows how the calibration of trust to reliability determines appropriate reliance, misuse, and disuse.

Worked Example

Consider the pointing task in the first demonstration, which makes Fitts's law concrete. Fitts found that the time to move to a target depends on an index of difficulty that combines the movement's amplitude and the target's width. In his 1954 formulation the index of difficulty is ID = log₂(2A / W) bits, where A is the amplitude of the movement (the distance to the target) and W is the target's width along the line of motion (Fitts, 1954). Movement time is then a linear function of that index: MT = a + b × ID, where a and b are constants fitted to the operator and the device.

Take an interface with an intercept a = 0.100 s and a slope b = 0.150 s per bit, and a control at amplitude A = 160 mm. For a wide target, W = 80 mm, the index is ID = log₂(2 × 160 / 80) = log₂(4) = 2.0 bits, so MT = 0.100 + 0.150 × 2.0 = 0.400 s. Halving the target to W = 40 mm raises the index to log₂(8) = 3.0 bits and the time to 0.550 s; halving again to W = 20 mm gives log₂(16) = 4.0 bits and 0.700 s; and a small W = 10 mm target gives log₂(32) = 5.0 bits and MT = 0.100 + 0.150 × 5.0 = 0.850 s. The throughput, a measure of the operator's effective bandwidth, is ID / MT = 5.0 / 0.850 = 5.88 bits per second at the hardest setting.

The point for design is the logarithmic form. Because movement time depends on the ratio of distance to width, a fixed increase in difficulty — doubling the distance or halving the target — adds a constant increment of time, not a proportional one, and the cheapest way to speed a movement is almost always to make the target bigger rather than to move it closer. The demonstration recomputes ID and MT live as the amplitude and width are changed; the four points above fall on the straight line of Figure 1, which is exactly the relationship Fitts's law asserts. These constants are illustrative of the law's structure, not measurements from any one device.

Figure 1

Discussion

Seen as a whole, the study of man-machine systems is a sustained argument against a seductive intuition: that a system is improved by making its machine more capable. The field's central findings all qualify that idea. Automating a function does not simply remove the human's work; it changes the human's task from doing to monitoring, a task people do poorly, and it can leave them unable to resume control when it matters most (Onnasch et al., 2014). More reliable automation does not guarantee better joint performance, because performance depends on whether the operator's trust is calibrated to that reliability (Lee & See, 2004). And a more powerful interface is not a better one if it exceeds the operator's workload capacity or starves their situation awareness (Endsley, 1995).

The unifying principle is that the human and the machine form one system, so the design variable that matters is the coupling between them, not the capability of either part. This is why the field's recommendations are so often about feedback, transparency, and the preservation of engagement rather than about raw automation: an operator who can see what the machine is doing and why can trust it appropriately, recover from its failures, and keep enough of the loop to stay competent (Norman, 2013). As automation grows more autonomous, that principle does not weaken — it becomes the whole problem, because the more the machine can do alone, the more its occasional need for the human depends on a human who has been kept ready.

Current Directions

Three lines of work are currently active. The first is the rethinking of automation taxonomies for an era of learning, adaptive systems. The stages-and-levels scheme was built for automation whose behaviour is fixed and knowable, and current work asks how the framework must change when the machine's competence varies with context and is not fully transparent even to its designers — a critique that questions how well the classic levels describe modern autonomy at all (Kaber, 2018). The second is the road to full autonomy, where the accumulated lessons of human-automation research are being applied to self-driving vehicles and other highly autonomous systems, whose intermediate stages reproduce the out-of-the-loop and handover problems the field has studied for decades, now with far less time for the human to recover (Endsley, 2017).

The third is human-robot interaction, in which the machine is an embodied, sometimes social agent rather than a display and a control. Here new variables enter, including the machine's physical form and apparent agency: a meta-analysis of anthropomorphism found that giving a robot humanlike features has real but bounded effects on how people perceive and cooperate with it, effects that interact with the trust and reliance dynamics established in the automation literature (Roesler et al., 2021). Across all three, the open problem is the one the field began with, made sharper by capable machines: how to keep a human meaningfully in a loop that the machine can increasingly run without them.

Common Misconceptions

“More automation always makes a system safer and more efficient.”

Higher automation reliably cuts workload and improves routine performance, but it degrades the operator's ability to detect and recover from failures — the lumberjack effect — so the right level is a design trade-off, not a maximum (Onnasch et al., 2014).

“A more reliable automated aid will always be used well.”

Reliance depends on trust, and only calibrated trust yields good use; when trust exceeds reliability operators over-rely and miss failures, and when it falls short they ignore aids that would have helped (Lee & See, 2004).

“Human error is the operator's fault.”

Errors are largely predictable products of how the joint system is designed, arising from latent conditions and poor feedback; they are reduced by redesigning the coupling, not by blaming or exhorting the operator (Reason, 1990).

“Once a task is automated, the human's job is simply easier.”

Automation changes the human's task from doing to monitoring and can starve situation awareness, so an operator who is out of the loop is often less able to handle the moments when the automation needs them (Endsley, 1995).

Glossary

Automation-induced complacency.

The reduced monitoring of a reliable automated system that follows from over-trust, leaving the operator slow to detect the occasions when it fails.

Automation.

The execution by a machine of a function previously carried out by a human operator; can be applied at different information-processing stages and to different degrees.

Fitts's law.

The empirical relationship that the time to acquire a target is a linear function of an index of difficulty, itself logarithmic in the ratio of movement amplitude to target width.

Function allocation.

The design decision of which functions in a task are assigned to the human operator and which to the machine; the founding problem of man-machine systems.

Haptic technology.

Technology that conveys information through the sense of touch — force feedback, vibration, and tactile displays; the narrower topic MeSH files beneath man-machine systems.

Human error.

A failure of a planned action to achieve its goal; analysed as slips and lapses (execution failures) versus mistakes (planning failures), and traced to latent conditions in the system.

Index of difficulty.

In Fitts's law, the quantity log₂(2A / W) in bits, combining movement amplitude A and target width W into a single measure of a movement's demand.

Levels of automation.

The continuum, within each processing stage, from fully manual to fully autonomous, describing how much authority the machine holds over that part of the task.

Man-machine system.

A system in which the functions of the human and the machine are interrelated and both necessary for its operation; the human-machine or human-automation system.

Mental workload.

The demand a task places on an operator's limited attentional resources; both overload and underload impair performance, and interference depends on whether tasks share a resource.

Multiple-resource theory.

Wickens's account that attentional capacity is several distinct resources rather than one pool, so two tasks interfere more when they draw on the same resource.

Out-of-the-loop problem.

The degradation of an operator's situation awareness and manual competence under high automation, which leaves them poorly placed to take over when the automation fails.

Situation awareness.

The operator's internal model of the task environment, comprising the perception of relevant elements, the comprehension of their meaning, and the projection of their future state.

Skill-rule-knowledge framework.

Rasmussen's classification of behaviour into automatic skilled action, rule-following, and effortful knowledge-based problem-solving, each with a characteristic error type.

Trust in automation.

The operator's attitude that automation will help achieve their goals under uncertainty; appropriate reliance requires trust calibrated to the automation's true reliability.

Key Researchers

Mica R. Endsley. American engineer who developed the leading three-level model of situation awareness — perception, comprehension, and projection — and its measurement, and applied it across aviation and autonomy. ORCID · Wikipedia · Wikidata

Paul M. Fitts (1912–1965). American psychologist who founded the engineering-psychology study of the human motor system; Fitts's law quantifies the speed–accuracy trade-off in aimed movement, and the Fitts list framed function allocation. Wikipedia · Wikidata

John D. Lee. American industrial engineer whose work with See on trust in automation set the framework for designing appropriate reliance between operators and automated systems. Google Scholar · Faculty page

Donald A. Norman (b. 1935). American cognitive scientist and usability engineer whose work on affordances, mappings, and the gulfs of execution and evaluation defined human-centred design. Wikipedia · Wikidata · Google Scholar

Raja Parasuraman (1950–2015). Indian-American cognitive neuroscientist who co-authored the canonical model of types and levels of automation and shaped the study of automation-induced complacency and vigilance. GMU In Memoriam

Jens Rasmussen (1926–2018). Danish system-safety researcher whose skill-rule-knowledge framework and abstraction hierarchy became foundational to cognitive systems engineering and the analysis of human error. Wikipedia · Wikidata

Thomas B. Sheridan (b. 1929). American engineer, a founder of the study of supervisory control and telerobotics, and co-author of the types-and-levels-of-automation model. Wikipedia · Wikidata

Christopher D. Wickens. American engineering psychologist whose multiple-resource theory of attention and work on mental workload and human-automation interaction are central to the field. ORCID · Faculty page

Frequently Asked Questions

What is a man-machine system in simple terms?
It is any task setup where a person and a machine work together and both are needed — a pilot and an autopilot, a driver and a lane-keeping system. The key idea is that they form one system, so how well it works depends on how the work is split and how the interface connects them, not just on how good the machine is.

What is Fitts's law?
It is the empirical finding that the time to move to and hit a target grows with an index of difficulty that combines how far the target is and how small it is. The relationship is logarithmic, so making a target bigger usually speeds a movement more than moving it closer (Fitts, 1954).

What are levels of automation?
Automation is not all-or-nothing. It can be applied to different stages of a task — acquiring information, analysing it, choosing an action, carrying it out — and, within each, at levels ranging from fully manual to fully autonomous. Describing a system means saying how high the automation reaches at each stage (Parasuraman et al., 2000).

What is situation awareness?
It is the operator's up-to-date model of what is happening: perceiving the relevant elements, understanding what they mean, and projecting what will happen next. Good decisions depend on it, and automation that removes the human from the control loop tends to erode it (Endsley, 1995).

Why does automation sometimes make performance worse?
Because it changes the human's job from doing to monitoring, a task people do poorly, and it can leave them out of the loop — with degraded awareness and rusty skills — exactly when the automation fails and needs them to take over. Higher automation eases routine work but makes recovery harder (Onnasch et al., 2014).

What is trust in automation, and why does it matter?
Trust is the operator's belief that automation will help them under uncertainty, and it drives whether they rely on it. Good performance needs trust calibrated to the automation's real reliability: over-trust leads to complacent misuse, and under-trust leads to disuse of aids that would have helped (Lee & See, 2004).

How is a man-machine system different from ergonomics?
Ergonomics (human factors) is the broad discipline of fitting work, tools, and environments to human capabilities; MeSH files man-machine systems beneath it. The man-machine-systems concept is the specific idea within that discipline of treating a human–machine pair as a single interdependent unit.

What is haptic technology?
It is technology that communicates with the operator through touch — force feedback, vibration, and tactile displays — adding a channel beyond sight and sound. In MeSH it is the one narrower topic filed beneath man-machine systems.

References