Attention

In plain terms, attention is how the mind picks out a small slice of the world to deal with properly while letting the rest slip by. At every instant the senses deliver far more information than the brain can fully process, so attention is the system that decides what gets through and how limited mental effort is shared out. This article explains what attention is and why it exists, the major theories of how it selects what to process (early versus late), how it shares limited capacity across competing tasks, how it is steered in space and bound into whole objects, the brain networks that support it, and the striking ways it fails.

Attention is the set of mechanisms by which the cognitive system selects some information for privileged processing and allocates its limited resources, under a constant interplay between current goals (top-down control) and the pull of salient events (bottom-up capture). It exists because processing capacity is sharply limited: only a small fraction of the information on the retina, or in the air, can be used to guide behaviour at any moment (Carrasco, 2011; Desimone & Duncan, 1995). This page is the overview for the topic and the hub for the rest of the section — selective attention, divided attention, sustained attention, feature integration theory, visual search, and the others. The field is best organised around two questions that recur in every theory: the problem of selection (what gets through, and at what stage) and the problem of capacity (how a fixed pool of resources is divided), with control — the balance of goals against salience — threaded through both.

A Response to Limited Capacity

It is tempting to think we perceive the world more or less completely — that we see what is in front of us and hear what is around us. We do not. The senses gather vastly more than the brain can analyse in full, and the machinery downstream of the receptors is the bottleneck. Donald Broadbent, building the first formal information-processing account of the mind, described the organism as a limited-capacity communication channel that must select among inputs because it cannot pass them all (Broadbent, 1958). The same logic still frames the field: selection is necessary precisely because there are severe limits on how much can be processed, limits rooted in the high metabolic cost of neural activity (Carrasco, 2011).

Attention is the name for the mechanisms that manage this scarcity. It is not one thing but a family of related abilities — holding a state of readiness, choosing what to admit, sharing effort between tasks, and resolving conflict when habits pull the wrong way. The sections below take these in turn, starting with the oldest and most studied question: when two messages arrive at once, how does the mind follow one and not the other?

Selective Attention and the Cocktail Party

The founding experiments used sound. Colin Cherry asked how a listener follows one conversation at a noisy party while others speak at the same time — the situation he named the cocktail party problem (Cherry, 1953). In his dichotic listening task, a different spoken message was played to each ear and the listener was asked to shadow one of them, repeating it aloud word by word. Listeners did this well, but afterwards could report almost nothing about the unattended ear: they noticed whether it carried speech and whether the voice was male or female, but not the words, not the meaning, and not even a switch from English to another language. Something was filtering the rejected channel out.

Yet the filter is not airtight. Neville Moray showed that a list of words repeated many times in the unattended ear left no trace in memory — but that one kind of item reliably broke through: the listener's own name (Moray, 1959). The own-name effect is a small fact with large consequences, because it tells us the unattended channel is not simply switched off; some of its content is analysed enough for an important word to capture attention. Where, then, does the selection happen? That question — the locus of selection — drove two decades of theory. (For the wider topic, see selective attention and the cocktail party effect.)

Where Does Selection Happen? Bottleneck Theories

The shared image behind these theories is a bottleneck: a stage at which the many things processed in parallel must narrow to the one or few that can be handled serially. Theories differ over where on the processing stream that bottleneck sits.

Broadbent's filter theory placed it early. On this account an early filter selects a channel by its simple physical properties — which ear, which pitch, which location — and only the selected channel passes on for analysis of meaning; the rest is blocked before it is recognised (Broadbent, 1958). Early selection neatly explains why Cherry's listeners knew the unattended voice's gender (a physical property) but not its words. It struggles, though, with the own-name effect: if the rejected channel is blocked before meaning, how does your name get through?

Anne Treisman repaired the model by softening the filter. In her attenuation theory, the unattended channel is not blocked but turned down — attenuated — so its signal is weakened rather than removed (Treisman, 1960). Words have different thresholds for recognition, and important or expectable words (your name, or a word the context predicts) have such low thresholds that even an attenuated signal can trip them. Treisman found that when the two messages were swapped between the ears mid-sentence, listeners would briefly follow the meaning across to the wrong ear, exactly as a threshold account predicts. Selection is early but graded, not all-or-none.

A more radical alternative moved the bottleneck to the end. J. A. Deutsch and Diana Deutsch proposed late selection: every input is analysed for meaning, and selection happens only afterwards, at the stage of deciding what to respond to or commit to memory (Deutsch & Deutsch, 1963). On this view nothing is filtered out perceptually; the apparent loss of the unattended message reflects a later failure to select it for report, not a failure to process it.

The table below lays the two poles side by side.

Reconciling the Debate: Perceptual Load

For forty years the early-versus-late argument ran without a decisive winner, because experiments sometimes showed unattended distractors being processed and sometimes did not. Nilli Lavie's load theory resolved the impasse by proposing that both can be right, depending on the task (Lavie, 1995; Lavie et al., 2004). Perception has a limited capacity that is nonetheless used involuntarily and to the full: whatever capacity a task leaves unused spills over and processes irrelevant stimuli. So when the relevant task carries high perceptual load — a crowded, demanding display — there is no spare capacity, distractors go unprocessed, and selection looks early. When the task carries low load — an easy display — spare capacity automatically leaks onto distractors, which are processed and can interfere, and selection looks late. The locus of selection is not fixed; it shifts with load.

The standard tool for measuring this leakage is the response-competition (flanker) task of Barbara and Charles Eriksen, in which a central target is flanked by letters that map to the same response or the opposite one (Eriksen & Eriksen, 1974). Incompatible flankers slow responses — proof that the flankers were processed despite being irrelevant — and load theory predicts, correctly, that this interference shrinks as the load of the central task rises. Selection, on this account, is a by-product of how fully a task consumes capacity. Load theory also distinguishes this perceptual limit from a separate role for higher cognitive control in keeping goals active against distraction (Lavie et al., 2004).

The Spotlight: Orienting Attention in Space

Selection has a geography as well as a stage. In vision, attention behaves like a spotlight that can be aimed at a region of space and, crucially, can move independently of where the eyes point — so-called covert attention. Michael Posner made this measurable with the spatial cueing paradigm: a cue draws attention to a location, and a target then appears either there (a valid trial) or elsewhere (an invalid trial) (Posner, 1980). Valid cues speed detection and invalid cues slow it, and the size of that benefit and cost traces the spotlight's effect even though the eyes never move. Decades of psychophysics since have shown that covert attention does not merely speed responses but genuinely sharpens early vision, improving contrast sensitivity and spatial resolution at the attended location (Carrasco, 2011).

Two ways of moving the spotlight are distinguished. Endogenous orienting is voluntary and goal-driven — you choose to attend somewhere, as when a central arrow tells you where a target will appear; it is relatively slow and sustained. Exogenous orienting is reflexive and stimulus-driven — a sudden flash or motion at the edge of vision grabs attention automatically, quickly, and briefly. (These have their own pages: endogenous attention and exogenous attention.) A signature of the reflexive system is inhibition of return: shortly after attention has been pulled to a location and then withdrawn, responses to that location become slower than to fresh ones, as if the system is biased against re-checking where it has just been — useful for searching new places (inhibition of return).

Binding Features: Feature Integration and Visual Search

Once attention is aimed at a location, what does it do there? Anne Treisman's feature integration theory gives an influential answer: attention is the glue that binds separate features into objects (Treisman & Gelade, 1980). The theory proposes two stages. First, a preattentive stage registers basic features — colour, orientation, size, motion — automatically, in parallel, across the whole visual field, with each feature held in its own map. Then a second stage uses focused attention, location by location, to combine the features present there into a single bound object.

The theory's reach comes from a sharp behavioural prediction about visual search. When the target is defined by a single feature — a red item among green ones — it is detected in the preattentive, parallel stage: it seems to pop out, and the time to find it barely changes as more distractors are added (a flat search slope). This is feature search and the pop-out effect. But when the target is defined by a conjunction of features — a red vertical bar among red horizontal and green vertical bars, where no single feature isolates it — parallel processing cannot find it, and attention must be moved from item to item to bind features and check each one. Search time then climbs steadily with the number of distractors: the signature of serial, attention-demanding conjunction search. A further prediction is that when attention is overloaded or diverted, features can be combined wrongly, producing illusory conjunctions — reporting a red X and a green O as a green X and a red O. The demonstration below contrasts the two kinds of search; later models such as the guided search model refine how feature information steers the spotlight.

Divided Attention and Mental Resources

So far the theme has been choosing one thing over another. But often we try to do two things at once, and the question becomes how a limited supply of attention is divided. Daniel Kahneman framed attention as a single pool of mental effort or capacity that can be allocated flexibly among activities, with the total amount rising and falling with arousal (Kahneman, 1973). On this model, two tasks can be combined as long as their joint demand stays within the available capacity; push past it and performance on one or both breaks down. This explains graceful dual-tasking when demands are light and sharp interference when they are heavy.

Christopher Wickens argued that a single pool is too simple, and proposed multiple resource theory: attention draws on several partly separate resources rather than one (Wickens, 2008). Tasks interfere most when they compete for the same resource — two visual tasks, or two verbal ones — and interfere less when they use different resources, which is why people can drive (visual-spatial) while listening (auditory-verbal) better than they can watch two displays at once. The practical payoff is workload prediction: knowing which resources two tasks share predicts when their combination will overload an operator. (The dedicated pages are divided attention and multiple resource theory.)

Automaticity and the Stroop Effect

Some processing demands attention and some does not, and the difference is central to how attention works. Walter Schneider and Richard Shiffrin drew the line between controlled processing — slow, serial, effortful, capacity-limited, and flexible — and automatic processing — fast, parallel, effortless, and hard to suppress, built up through consistent practice (Schneider & Shiffrin, 1977). Automaticity is what lets a skilled reader take in a word at a glance or a skilled driver steer without deliberation. Its cost is rigidity: an automatic process runs off whether or not you want it to.

The classic demonstration is the Stroop effect (Stroop, 1935). Name the ink colour of a colour word and you are fast when the two agree (the word RED in red ink) but markedly slower when they conflict (the word RED in blue ink). Reading is so automatic that the word's meaning is processed despite being irrelevant, and it competes with the controlled task of naming the ink — so the conflicting word steals time and provokes errors. The Stroop effect is the textbook signature of automaticity intruding on a controlled response, and the conflict it creates is exactly what the executive side of attention exists to resolve. (See automatic processes and the Stroop effect.)

The Attention Networks of the Brain

The abilities described above are not a single faculty, and an influential synthesis from cognitive neuroscience maps them onto distinct attention networks. Michael Posner and Steven Petersen distinguished three: an alerting network that achieves and maintains a state of readiness, an orienting network that selects information from sensory input (the spotlight of the cueing studies), and an executive control network that resolves conflict among responses — the system taxed by the Stroop task (Posner & Petersen, 1990; Petersen & Posner, 2012). These functions can be measured and dissociated, which is why the framework has been so durable.

Two further accounts fill in how selection is controlled. Maurizio Corbetta and Gordon Shulman described two interacting frontoparietal systems: a dorsal network that applies goal-directed, top-down selection (deciding where to attend), and a ventral network that detects salient or unexpected events and acts as a circuit breaker, reorienting the dorsal system toward something important — the neural counterpart of the endogenous-versus-exogenous distinction (Corbetta & Shulman, 2002). And at the level of single representations, Robert Desimone and John Duncan proposed biased competition: objects in a scene compete to be represented, and attention works by biasing that competition in favour of whatever is currently relevant, so the winner is determined jointly by stimulus strength and by top-down goals (Desimone & Duncan, 1995). The biased competition model and the figure most associated with this work, Michael Posner, have their own pages.

When Attention Fails: Inattentional and Change Blindness

If attention is necessary for full processing, then withdrawing it should make even obvious things vanish from awareness — and it does. Ulric Neisser and Robert Becklen first showed this with superimposed videos: observers tracking one of two overlapping events routinely failed to notice salient happenings in the other (Neisser & Becklen, 1975). The most famous version is Daniel Simons and Christopher Chabris's gorilla study: viewers counting basketball passes were so absorbed that about half failed to see a person in a gorilla suit walk into the middle of the scene, stop, and beat their chest — inattentional blindness, the failure to see an unexpected object when attention is engaged elsewhere (Simons & Chabris, 1999).

A related failure concerns change. Ronald Rensink and colleagues showed that when a brief blank flicker is inserted between two versions of a scene, people are remarkably poor at spotting even large differences between them — change blindness — which implies we do not hold a complete, detailed internal picture of the world and must attend to a thing to register that it has changed (Rensink et al., 1997). A third limit is temporal: in a rapid stream of items, a second target is often missed if it follows the first within about half a second, the attentional blink. Together these phenomena make the central point vivid — attention is not a luxury that sharpens an already-complete perception; it is a precondition for consciously perceiving even the obvious (change blindness, inattentional blindness).

A Worked Example: A Phone Call Behind the Wheel

Put the pieces together in one everyday scene: driving while holding a phone conversation. The conversation and the road compete for selection, and the conversation is hard to ignore because, like the cocktail party, speech is processed even when you mean to set it aside. As traffic thickens, the driving task's perceptual load rises; load theory says that should, if anything, leave less capacity for the conversation — but the conversation is self-paced and keeps demanding its share anyway. In resource terms the two are partly compatible: driving is largely visual-spatial and talking is auditory-verbal, so by multiple resource theory they interfere less than two visual tasks would. That is the grain of truth in hands-free systems. But they are not free of cost, because both draw on the same central, executive capacity, so braking reactions still lengthen. Skilled driving survives the split only because much of it is automatic (lane-keeping, gear changes) — yet the moment something novel happens, a child stepping out, the situation needs controlled attention, which the conversation has been consuming. The result is the failure mode that matters: the driver can look straight at the hazard and still not see it, inattentional blindness at the wheel. The lesson of decades of attention research is that the bottleneck is the mind, not the hands — which is why hands-free conversation does not make driving safe.

Why It Matters: From Driving to Design

The same principles shape a great deal of applied work. Road safety is the worked example writ large: distraction research rests directly on divided attention, capacity, and inattentional blindness. Interface and information design exploits the spotlight and its limits — a well-chosen colour or motion makes an alert pop out preattentively, while a cluttered screen raises perceptual load and slows everything; designers manage salience so that what matters captures attention and what does not stays quiet. Safety-critical monitoring — air-traffic control, baggage screening, anaesthesia — depends on sustained attention and vigilance, which decline over long, low-event watches, a robust limit that staffing and alerting systems are built around, and on workload models that predict when an operator's resources will be overloaded (Wickens, 2008). And attention is central to clinical questions: it is the core difficulty in attention-deficit/hyperactivity disorder (see ADHD) and is selectively disrupted after brain injury, as in hemispatial neglect, where the link between attention and awareness becomes starkly visible.

Key Researchers

Donald Broadbent (1926–1993) — built the first formal information-processing model of selective attention, the early-selection filter theory (Broadbent, 1958).

Colin Cherry (1914–1979) — pioneered the dichotic listening method and framed the cocktail party problem (Cherry, 1953).

Anne Treisman (1935–2018) — developed attenuation theory and, later, feature integration theory, two of the most influential ideas in the field (Treisman, 1960; Treisman & Gelade, 1980).

Daniel Kahneman (1934–2024) — advanced the capacity, or effort, model of attention as a single flexible resource (Kahneman, 1973).

Michael I. Posner — University of Oregon (emeritus); created the spatial cueing paradigm and, with Petersen, the attention-networks framework (Posner, 1980; Posner & Petersen, 1990). Faculty

Nilli Lavie — UCL Institute of Cognitive Neuroscience; developed perceptual load theory, reconciling early and late selection (Lavie, 1995). Faculty

Marisa Carrasco — New York University; established through psychophysics how covert attention alters early vision (Carrasco, 2011). Faculty

Key Terms

Frequently Asked Questions

What is attention in psychology?
Attention is the set of mechanisms that select some information for fuller processing and share out the brain's limited resources, steered by current goals and by salient events. It exists because the senses deliver far more than can be processed at once, so the system must choose what to admit and how to divide effort (Broadbent, 1958; Carrasco, 2011).

What is the difference between early and late selection?
Early-selection theories say attention filters information by simple physical features before its meaning is analysed, so unattended input is blocked or weakened early (Broadbent, 1958; Treisman, 1960). Late-selection theory says everything is analysed for meaning and the filtering happens later, when deciding what to respond to or remember (Deutsch & Deutsch, 1963). Load theory shows which one applies depends on how demanding the task is (Lavie, 1995).

What is the cocktail party effect?
It is the ability to follow one conversation among many competing voices, first studied by Cherry with the dichotic listening task (Cherry, 1953). Listeners shadowing one message notice almost nothing about an unattended one — except certain meaningful items, most reliably their own name, which break through and reveal that the unattended channel is not entirely shut out (Moray, 1959).

Why does the Stroop effect happen?
Because reading is automatic. When you try to name the ink colour of a conflicting colour word, the word's meaning is processed whether you want it to be or not, and it competes with the colour-naming response, slowing you down and causing errors (Stroop, 1935). It is the classic sign of an automatic process intruding on a controlled one (Schneider & Shiffrin, 1977).

What is inattentional blindness?
It is the failure to notice an unexpected but fully visible object when your attention is occupied by another task. In the best-known demonstration, about half of viewers counting basketball passes did not see a person in a gorilla suit cross the scene (Simons & Chabris, 1999). It shows that attention is needed to consciously perceive even obvious events.

Can we really pay attention to two things at once?
Sometimes, within limits. Attention is a limited capacity, so two tasks can be combined only while their joint demand stays within what is available (Kahneman, 1973). Interference is worst when the tasks use the same kind of resource and lighter when they use different ones, which is why listening while driving is easier than watching two things at once — though it is never truly cost-free (Wickens, 2008).

References

The three interactive figures on this page — the Posner cueing, visual-search, and Stroop demonstrations — generate their trials and compute their reaction times live in your browser; no experimental dataset is bundled with the page. The empirical claims in the text are sourced to the references above.