Discipline 01 / Reference

Cognitive Science

Cognitive science is the interdisciplinary study of mind and intelligence: how perception, attention, memory, language, and judgment actually operate, in people and in the systems built to imitate them. It borrows method from psychology, neuroscience, linguistics, computer science, philosophy, and anthropology, and it treats cognition as something measurable rather than something only describable. This page covers what the field measures, the constraints it has established, and how those constraints show up inside an organization.

Field
Cognitive science
Draws on
Psychology, neuroscience, linguistics, computer science, philosophy, anthropology
Formalized
1956 onward; the term was coined in 1973
Central claim
Cognition is information processing, and its limits are measurable and largely shared across people
Primary methods
Controlled experiment, semi-structured interview, instrumented measurement, process tracing
Applied here to
Decision architecture, tacit knowledge capture, human-in-the-loop design
01

What the field actually measures.

The useful part of cognitive science for an organization is not its theories of mind. It is the set of constructs that have been operationalized well enough to measure, and that behave consistently enough across people to design against. Six of them do most of the work.

ConstructWhat it isHow it is measuredWhy it matters at work
Working memoryThe small set of items a person can hold and manipulate at once. Current estimates put it near four chunks, not the seven of popular memory.Span tasks, dual-task interference.Any step requiring more than a few conditions held in mind will be performed wrong under time pressure, whatever the training.
Cognitive loadTotal mental effort a task imposes, split between the task itself and the way it is presented.Validated self-report scales, secondary task cost, pupillometry.Load spent decoding a bad interface or unclear ownership is load unavailable for judgment.
Vigilance decrementMeasurable decline in detection accuracy during sustained monitoring, observable within the first half hour.Signal detection tasks over extended periods.Roles that consist of watching for rare events degrade on a known curve. Staffing that ignores the curve is staffing a fiction.
Recognition and recallRecognizing a correct option is markedly easier and more reliable than producing it from memory.Forced-choice compared against free recall.This is why checklists work. Asking people to remember a procedure sets a harder task than showing it to them.
Metacognitive calibrationThe gap between how confident someone is and how accurate they are.Confidence ratings scored against outcomes.Miscalibration is the mechanism behind most escalation failures. The person who should have asked did not know they should have.
Situation awarenessWhether an operator's mental model matches the actual state of the system they are responsible for.Freeze-probe and query techniques.Automation often improves routine performance while quietly degrading this, which is why handover is where automated systems fail.
02

Eight constraints that hold across people.

These are findings, not opinions about human nature. Each has held up across decades of replication, and each is a design constraint rather than a flaw to be trained away. Open any of them for the field definition and where it surfaces at work.

Bounded rationalityOpen

Herbert Simon's finding that people do not optimize across options. They satisfice: they search until something clears an acceptable bar, then stop. Search is expensive, and stopping early is usually the rational move given that cost.

In an organization

Your process assumes a comparison of alternatives. What happens is that the first workable option gets taken. Designing for the comparison you wish occurred produces a process nobody follows.

Automation biasOpen

The tendency to accept an automated recommendation without independent verification, and to miss errors the system makes. It strengthens as the system becomes more reliable, because reliability trains the operator out of checking.

In an organization

Once a tool is usually right, people stop checking it. The moment checking matters most is exactly the moment nobody is doing it any more.

Skill atrophyOpen

Capability degrades when it stops being exercised. Lisanne Bainbridge named the version that matters here the out-of-the-loop problem: automate the routine cases and the operator loses the practice that made them competent at the hard ones.

In an organization

The team that reviewed every case for years and now reviews only exceptions has lost the baseline that made them good at recognizing an exception.

Illusion of explanatory depthOpen

People believe they understand mechanisms in far more detail than they do, and the belief survives until they are asked to explain the mechanism step by step. The explanation attempt is what reveals the gap, to the explainer as much as to anyone else.

In an organization

Ask three people to describe the same process end to end. The divergence between the three accounts is the finding, and it is usually larger than any of them expect.

Tacit knowledgeOpen

Michael Polanyi's observation that we know more than we can tell. Expertise compiles into pattern recognition that the expert can act on but cannot decompose into rules on request. It is recoverable, but only by watching the work and asking about specific cases rather than general procedure.

In an organization

The expert cannot write the procedure because the procedure is not what they are doing. Ask them to document it and you get a description of the official version they also know.

Transactive memoryOpen

Groups distribute knowledge across members and maintain a shared index of who knows what. The group's capability is therefore not the sum of individual knowledge; it includes the routing information that connects a question to the right person.

In an organization

Losing one person can remove the index rather than a single entry. The knowledge is still in the building and nobody can find it any more.

Diffusion of responsibilityOpen

As the number of people who could act rises, the probability that any particular one does falls. Each person's sense of personal obligation is divided by the size of the group that shares it.

In an organization

A decision assigned to a committee is assigned to nobody. The reliable fix is naming one accountable person, not adding a reminder to the group channel.

AnchoringOpen

The first quantity presented biases every judgment that follows, including judgments by people who know about the effect and are trying to correct for it. Awareness reduces the size of the effect and does not remove it.

In an organization

The first estimate spoken in a planning meeting sets the range for the whole session, regardless of where it came from or how much thought went into it.

03

Anatomy of a decision moment.

Most process failures are not failures of intent or of training. They occur at a specific instant, where a particular person holds particular information under particular pressure, and the documented path costs more than the alternative. Naming that instant precisely is what makes it fixable, because the moment can be redesigned even when the person cannot be changed.

Triggeran event arrivesDecisionmomentWHO DECIDESWHAT THEY KNOW AT THAT INSTANTWHAT PRESSURE THEY ARE UNDERDocumented pathwhat the procedure saysDIVERGENCE MEASURED HEREActual pathwhat people do
Figure 1. The unit of analysis. Divergence between the two right-hand paths is not a compliance problem to be corrected. It is evidence about the conditions at the decision moment, and it is usually the cheapest available description of what is actually wrong.
04

How the work is done.

Behavioral findings do not transfer directly from a paper to your organization. The constructs are general; the conditions are local. The sequence below establishes the local facts before anything is redesigned.

01ObserveWatch the work happen02InterviewSurface the tacit rules03InstrumentLoad and calibration04ModelTest before deploying05RedesignAlter the moment
Figure 2. Observation precedes interview deliberately. Asking first anchors the answers to the official version, and the official version is what everyone can already recite.
  1. 01

    Observe the work as performed

    Watch the task happen without intervening, and record where it diverges from documentation. Divergence is the raw material. Treating it as a violation at this stage ends the study, because people stop showing you the real version.

  2. 02

    Interview around specific cases

    Semi-structured interviews anchored to particular incidents rather than general procedure. Asking how something is normally done retrieves the official account; asking about last Thursday's difficult case retrieves the reasoning that was actually used.

  3. 03

    Instrument load, confidence, and authority

    Short validated instruments administered in context measure cognitive load, operator confidence against outcome, and whether the person accountable for a decision holds the authority to make it. Misalignment on the third of these predicts failure earlier than any performance metric.

  4. 04

    Model the change before deploying it

    Agent-based simulation runs the proposed redesign against the observed behavior, which surfaces second-order effects such as load displaced onto an adjacent role. Interventions that fix one moment by breaking another are common and are cheaper to find here.

  5. 05

    Redesign the moment, not the person

    Change what information is present, who holds authority, how the option is presented, and what the correct action costs relative to the alternative. Training is the last resort rather than the first, because training decays and structure does not.

05

What this produces, and what it does not.

Behavioral work attracts overclaiming more than most disciplines, partly because its findings are easy to state and hard to apply. The right-hand column is the more useful one to read first.

Produces
  • A map of the decision moments in a process, each with its actor, its available information, and its pressures named explicitly.
  • Measured cognitive load and confidence calibration, with the specific steps where either exceeds a workable range.
  • Job aids designed against recognition rather than recall, which is the difference between a checklist that gets used and one that gets filed.
  • Review points placed where authority actually sits, rather than where the org chart says it should.
  • A documented baseline of current capability, so that atrophy after an automation change is detectable rather than inferred later.
Does not produce
  • Personality assessment, typing, or anything that sorts your people into categories. That is a different field with a considerably weaker evidence base.
  • A prediction of what a named individual will do. These constructs describe populations and conditions, not persons.
  • A substitute for fixing incentives. Where the reward structure opposes the desired behavior, no amount of interface design will hold.
  • Reliable findings from a single round of interviews. Self-report without observation reproduces the official account.
  • A case for removing people from decisions. The finding is almost always about the conditions surrounding a judgment, not the judgment itself.
06

Where the field came from.

Cognitive science began as a rejection of behaviorism's refusal to model what happens between stimulus and response. A cluster of papers in 1956 established that the interior could be studied with formal methods, and the field assembled around that possibility over the following two decades. The applied branch that matters most here, the study of skilled people working under real conditions, arrived later and largely from aviation and process control.

  1. Herbert A. Simon, A Behavioral Model of Rational Choice (1955). Bounded rationality stated formally. The foundation for every subsequent account of how people actually choose under constraint.
  2. George A. Miller, The Magical Number Seven, Plus or Minus Two, Psychological Review (1956). The paper that made capacity limits a measurable quantity. Later work revised the number downward, which strengthens rather than weakens the design implication.
  3. Michael Polanyi, The Tacit Dimension (1966). The source of the observation that we know more than we can tell, and the reason knowledge capture is an interview problem rather than a documentation problem.
  4. Ulric Neisser, Cognition and Reality (1976). An argument from inside the field that laboratory findings must be tested against how cognition works in ordinary settings. The case for ecological validity.
  5. Lisanne Bainbridge, Ironies of Automation, Automatica (1983). Six pages on why automating the easy parts of a job makes the remaining parts harder. Written about process plants, and describing most current AI deployments precisely.
  6. Gary Klein, Sources of Power (1998). How experts actually decide under time pressure, based on fieldwork with firefighters and military commanders. The origin of recognition-primed decision making.