Recognition-Primed Decision Making — When Expert Intuition Is Reliable

Recognition-Primed Decision Making

TL;DR: Recognition-Primed Decision making (RPD), formalized by Gary Klein from naturalistic field studies of firefighters, military commanders, and ER nurses, describes how experts make rapid high-stakes decisions: they recognize the situation as similar to past prototypes, generate one course of action, mentally simulate its outcome, and act if the simulation looks good. Not parallel comparison — serial pattern-matching. In their joint 2009 paper “Conditions for Intuitive Expertise: A Failure to Disagree,” Klein and Kahneman agreed on the resolution: intuitive expertise is reliable only when (a) the environment is sufficiently regular for valid cues to exist, and (b) the expert has had prolonged opportunity to learn the regularities through rapid, unambiguous feedback. When either condition fails, confident intuition becomes systematic error. The frame predicts where AI pattern-matching will and will not be trustworthy — for exactly the same reasons.

What it is

Recognition-Primed Decision making (RPD) is Gary Klein’s 1989 model — popularized in Sources of Power (1998) — for how experts actually make decisions in time-pressured, high-stakes environments. It came from the Naturalistic Decision Making (NDM) tradition: field studies of fireground commanders, military officers, ER nurses, chess masters, neonatal ICU staff, and other domains where decisions matter and the conditions are too messy for textbook decision theory.

The headline finding: experts do not generate a list of options and compare them. They:

Recognize the situation as similar to one or more prototypes from past experience.
Generate one course of action — the first one suggested by pattern matching.
Mentally simulate that action’s likely outcome.
If the simulation looks acceptable: act. If not: adjust the plan, simulate again. Iterate serially.

This is serial single-option evaluation, not parallel multi-option comparison. It explains why expert firefighters can decide to evacuate a building in seconds where a decision-theory model would require minutes of explicit option comparison.

Three Variations

Klein identified three variations of the basic model:

Recognition of typical situation → immediate response. (“If X then Y.”) Veteran firefighter recognizes a flashover-imminent room, evacuates without conscious deliberation.
Diagnosis of unfamiliar situation → known action. Expert can’t immediately classify the situation; spends cycles diagnosing, then applies a familiar action once classified.
Known situation, no known action. Expert recognizes the situation but no action template matches; uses mental simulation to generate and evaluate a novel approach.

The Klein-Kahneman 2009 Synthesis

For decades, Klein’s NDM tradition appeared to disagree with Kahneman & Tversky’s Heuristics and Biases (HB) tradition on a fundamental question: can intuition be trusted?

HB tradition (Kahneman): intuition is biased, overconfident, systematically error-prone. Famous demonstrations of base-rate neglect, anchoring, availability, representativeness — all show intuition failing.
NDM tradition (Klein): expert intuition is fast, accurate, and crucial in real high-stakes settings. Firefighters and chess masters demonstrably make great decisions intuitively.

In their 2009 American Psychologist paper “Conditions for Intuitive Expertise: A Failure to Disagree,” Klein and Kahneman agreed on a synthesis. The finding both traditions had been pointing at, from opposite sides:

Intuitive expertise is reliable when, and only when, two conditions both hold:

The environment provides high-validity cues — the relationships between situation features and correct outcomes are stable and learnable. (Chess: yes. Firefighting most of the time: yes. Stock-picking: no. Long-range political prediction: no.)
The expert has had prolonged opportunity to practice with rapid, unambiguous feedback. Without feedback, no genuine pattern library can be built — only the feeling of expertise.

When both conditions hold, intuition is recognition, and Klein is right. When either fails, intuition is substitution and overconfidence, and Kahneman is right.

The most important Kahneman-Klein quote:

“Subjective experience is not a reliable indicator of judgment accuracy.”

Experts in low-validity environments still feel confident. Their confidence is uncorrelated with their accuracy. Stock pickers, political pundits, and many strategy consultants are the canonical examples — high subjective certainty, mediocre or chance-level prediction.

A practical typology

Environment	Feedback	Klein-Kahneman verdict	Examples
High-validity	Rapid, unambiguous	Intuition is reliable	Chess masters, firefighters on routine fires, ER triage, weather prediction (next day)
High-validity	Slow or noisy	Skill develops slowly or partially	Investment banking M&A, surgery (some specialties), agricultural decisions
Low-validity	Rapid feedback	Confidence develops; accuracy doesn’t	Day trading, sports betting, pop hits prediction (?)
Low-validity	Slow or noisy	Pure overconfidence	Macroeconomic forecasting, long-range geopolitics, clinical psychology, fund management

The implication: be skeptical of expert intuition in the bottom row. Trust expert intuition in the top row. Track which row a domain occupies before delegating to a human expert — or to an AI.

Why it matters for the wiki

The RPD framework, paired with the Klein-Kahneman synthesis, supplies the theoretical foundation for several wiki claims that previously rested on observation alone.

It predicts where AI pattern-matching will succeed and fail

LLMs are also doing a form of pattern-matching, just from training data instead of personal experience. The Klein-Kahneman conditions apply to AI for the same reasons they apply to humans:

High-validity environment + abundant training data → AI pattern-matching is reliable. (Customer-support response patterns, code completion in popular languages, well-structured writing tasks. This is exactly the glossary/jagged-frontier inside-frontier zone.)
High-validity environment but sparse training data → AI may extrapolate badly even though the underlying patterns are learnable. (Niche professional domains with few documented cases.)
Low-validity environment → AI confidently outputs answers that are no more accurate than chance. (Long-range market prediction, geopolitical forecasting, novel strategy decisions in unprecedented contexts.) This is where the glossary/jagged-frontier outside-frontier failure mode is theoretically predicted.

The Klein-Kahneman frame explains why AI is jagged: AI inherits the same conditions for reliable pattern-matching that humans do. Where the underlying environment is unpredictable, no amount of training compensates.

It refines the strategy-vs-execution claim

comparisons/strategy-vs-execution-ai previously argued that strategy stays human-leveraged. The Klein-Kahneman frame sharpens this: strategy in low-validity environments will struggle for both humans AND AI. The reason humans should keep doing it isn’t that humans are reliably good at it — it’s that the responsibility needs to be located somewhere accountable, and AI’s confident-but-not-validated output in low-validity domains is dangerous.

This is a meaningful refinement: it removes the accidental implication that human strategists are reliably accurate. Some are; some aren’t. The right test is the Klein-Kahneman conditions, applied to the specific domain.

It supplies the dual-process theory’s missing context

glossary/dual-process-thinking (Kahneman) described System 1 / System 2 thinking and the failure modes of fast intuition. RPD describes when System 1 works. Together, they tell a complete story:

System 1 (fast, pattern-matching) is reliable when the environment is high-validity and you’ve had feedback. Klein.
System 1 is unreliable when the environment is low-validity or feedback is missing. Kahneman.
System 2 (slow, deliberate) is the fallback when System 1 doesn’t have a good pattern available. Both agree.

The wiki’s existing dual-process page covers half of this. RPD covers the other half.

For the questions/ai-as-personal-advisor open question

The framework is directly applicable: an AI advisor will be useful in high-validity, abundant-data domains and dangerous in low-validity ones. The user’s instinct to trust the advisor will track confidence, not accuracy. Same failure mode as trusting a human expert in a low-validity domain — but easier to do because AI confidence is universal and instant.

Honest limits

Klein’s original RPD field studies (firefighters, military) are domains with rapid, unambiguous feedback. Generalization to slower-feedback domains (medicine, finance, strategy) is exactly what Klein and Kahneman themselves caution against. Do not assume RPD applies wherever experts feel confident.
“High-validity environment” is a slippery construct. Practitioners in dubious-validity domains (fund managers, clinical psychologists) often believe their environments are higher-validity than they are. Self-assessment is unreliable.
The model describes what experts do; it doesn’t tell beginners how to acquire the pattern library. Klein devotes parts of Sources of Power to this — deliberate practice with feedback. There are no shortcuts that the data supports.
Klein-Kahneman 2009 is a theoretical synthesis, not a quantitative meta-analysis. It is broadly accepted but has been challenged (e.g., on whether the boundary conditions are sufficient or just necessary).
Applying the frame to AI is an extrapolation. It is consistent with empirical findings (jagged frontier, hallucination patterns) but the original framework is about humans.
The framework is descriptive of individual expertise. Group decision-making and organizational pattern-matching are related but distinct topics not covered.

glossary/dual-process-thinking — Kahneman half of the duo; together they describe both when fast intuition works and when it fails
glossary/jagged-frontier — Klein-Kahneman explains why the AI capability frontier is jagged (high-validity tasks are inside; low-validity tasks are outside)
comparisons/strategy-vs-execution-ai — strategy in low-validity environments is hard for both humans and AI; the reasoning needs sharpening, not just human-credit
questions/ai-as-personal-advisor — when can AI advice be trusted? Klein-Kahneman gives the conditions
glossary/honest-assessment — honest assessment as a content strategy is a low-tech application of the same humility about confidence vs. accuracy
glossary/automation-eats-execution — execution work tends to be high-validity; strategy often spans into low-validity territory
glossary/agent-engineering — Karpathy’s “jagged intelligence” (model-side) inherits the same conditions-for-reliability that Klein-Kahneman established for human pattern-matching
glossary/agent-adoption-frictions — trust accumulates fastest in tasks meeting both Klein-Kahneman conditions (high-validity + rapid feedback); the user-side trust friction is structurally predicted by this foundation
marketing/ai-tells-in-sales-copy — operators detect AI tells faster than conscious analysis because pitch copy is exactly the high-validity environment with rapid feedback that makes pattern-matching reliable
marketing/ai-human-voice-prompting — same mechanism extended to platform algorithms. LinkedIn 360Brew, X Grok ranking, and email spam filters are algorithmic pattern-matchers operating in high-validity environments (massive labeled training data) with rapid feedback. They detect AI patterns by the same reliability conditions Klein-Kahneman established for human experts
glossary/hallucination — what happens when the model is not in a high-validity environment but continues to pattern-match anyway. Klein-Kahneman’s conditions predict where hallucination is rare (inside frontier) vs. reliable (outside frontier)

Key takeaways

Experts make decisions by serial pattern-matching against past prototypes, not by parallel option comparison.
Three RPD variations: typical-situation if-then, diagnose-then-act, known-situation novel-simulation.
Klein-Kahneman 2009 synthesis: intuition is reliable when the environment is high-validity AND the expert has had practice with rapid, unambiguous feedback. Both conditions are required.
“Subjective experience is not a reliable indicator of judgment accuracy.” — direct quote.
The frame predicts where AI pattern-matching will succeed (high-validity + abundant data) and fail (low-validity environments). AI is jagged for the same structural reasons human intuition is.
Strategic work in low-validity domains is hard for both humans and AI. The reason to keep humans on it isn’t reliable accuracy — it’s accountability and judgment about when not to trust the pattern.
Pairs with glossary/dual-process-thinking for the complete picture: when System 1 works, when it doesn’t.

Sources

Klein, G. (1998). Sources of Power: How People Make Decisions. MIT Press. — Book-length treatment of RPD with field-study cases. Foundational.
Kahneman, D., & Klein, G. (2009). Conditions for intuitive expertise: A failure to disagree. American Psychologist, 64(6), 515–526. — The synthesis paper. Both authors agree on conditions for reliable intuition. This is the load-bearing citation for boundary conditions.
Klein, G. (1989). Recognition-primed decisions. In W. Rouse (Ed.), Advances in Man-Machine Systems Research, 5, 47–92. — Original formal RPD model.
Kahneman, D. (2011). Thinking, Fast and Slow. — Companion treatment of the heuristics-and-biases side. See glossary/dual-process-thinking.