In 2016, AlphaGo played move 37 in game 2 against Lee Sedol. Commentators called it a mistake. It won the game. The network had not walked a rule tree out loud. It had ranked a pattern from millions of board positions.
Tap to enlargeTwo Modes of Human Thought
Daniel Kahneman split that kind of snap judgment from slow arithmetic in Thinking, Fast and Slow (2011). System 1 is the glance: a chess player seeing a move, a firefighter leaving a room before the floor goes. System 2 is 347 times 28 in your head. AI labs spent decades on System 2: expert systems, theorem provers, planners. Those tools died on messy rooms and incomplete rules.
Expert intuition is pattern recognition trained by experience, operating on representations too complex to articulate but too useful to ignore.
The current generation of AI, particularly large neural networks, may represent the first meaningful attempt at System 1 replication. They process information in a manner that is parallel, associative, and pattern-driven rather than sequential and rule-based.
What expert intuition is
Gary Klein's naturalistic decision-making research (1998) studied how experts in high-stakes environments (firefighters, military commanders, ICU nurses) make rapid decisions under uncertainty. His findings challenged the rationalist assumption that good decisions require deliberate analysis.
Klein found that experts rarely compare options systematically. Instead, they recognize the current situation as belonging to a category they have encountered before and immediately generate a plausible course of action. If the action satisfies a quick mental simulation, they execute it. If not, they modify it or generate another. The entire process takes seconds.
This recognition-primed decision model treats intuition as extensive domain experience compressed into rapid pattern matching. The expert has seen so many variations of a problem that the underlying structure of a new instance is immediately recognizable, even when surface features differ.
K. Anders Ericsson's research on deliberate practice (1993) suggests that expert-level intuition typically requires approximately 10,000 hours of domain-specific experience. This provides an empirical benchmark: intuition is the result of extensive training that encodes complex patterns into fast, automatic retrieval systems. The question for artificial systems is whether analogous pattern encoding can be achieved computationally, and if so, whether it constitutes "intuition" in any meaningful sense.
The relevance to AI is direct. If expert intuition is pattern recognition on compressed experience, and if neural networks perform pattern recognition on compressed training data, then the structural analogy is not trivial. The outputs may be functionally similar even if the mechanisms differ.
Where AI Already Demonstrates Intuition-Like Behavior
Several AI systems have produced outputs that resemble human intuitive judgment: fast, accurate, and not easily explicable through sequential reasoning.
AlphaGo's Move 37 (2016). In the second game against Lee Sedol, AlphaGo played a move on the fifth line that professional Go commentators initially dismissed as a mistake. Within 15 moves, the strategic value of the placement became clear. AlphaGo's evaluation network had identified a positional advantage that was invisible to human experts analyzing the board position by position. The system arrived at a judgment that matched (and exceeded) human intuition without following human-legible reasoning.
Protein structure prediction. AlphaFold (2020) predicts three-dimensional protein structures from amino acid sequences with accuracy approaching experimental methods. The system recognizes patterns in the relationship between sequences and structures, learned from approximately 170,000 known structures, rather than reasoning through biochemical principles step by step. The output resembles the intuitive judgment of an experienced structural biologist, arrived at through a different process.
Medical image interpretation. Deep learning systems for radiology, pathology, and dermatology match or exceed specialist-level performance in detecting specific conditions (diabetic retinopathy, certain cancers, skin lesions). These systems recognize visual patterns associated with disease rather than applying diagnostic criteria sequentially, in a manner that is structurally analogous to how experienced clinicians develop a "clinical eye" through years of case exposure.
AI systems that produce intuition-like outputs do so through statistical pattern matching on training data, not through the neurobiological processes that produce human intuition. The functional similarity (fast, accurate, difficult to explain) does not establish mechanistic equivalence. Whether the AI "understands" its domain in any sense comparable to human expert understanding remains an open philosophical question. What can be established empirically is that the outputs are reliable enough to be useful, and that the process by which they are generated shares structural features with human intuitive cognition.
Knowledge as a Network
The connection between intuition and knowledge structure can be modeled through network theory. Human knowledge is organized as a dense, interconnected semantic network where concepts are nodes and relationships are edges with varying strengths.
A standard analytical approach traverses this network methodically: premise to conclusion, step by step. An intuitive approach assesses the network topology globally, identifying emergent patterns and novel connections between distant concepts. The "gut feeling" of an expert is, in computational terms, a highly refined graph-traversal algorithm that has been optimized through experience to identify non-obvious paths.
This has implications for how AI systems can be designed to produce new outputs rather than merely recombinant ones. Current language models trained on human-generated text encode a representation of the human knowledge graph. When they generate novel connections between previously unrelated concepts, the process is structurally analogous to what happens when a human expert has an intuitive insight that connects ideas across domains.
Measuring innovativeness. The degree to which a solution relies on non-obvious semantic paths (connecting previously unrelated parts of the knowledge graph) versus well-trodden paths (replicating existing associations) can serve as a proxy for innovativeness. A system that consistently produces connections scored as "high distance" in the knowledge graph may be exhibiting something functionally analogous to creative intuition.
The Limits of Artificial Intuition
Human intuition has well-documented failure modes. Kahneman himself catalogued the systematic biases that System 1 produces: anchoring, availability bias, representational heuristics, overconfidence in pattern-matched judgments that happen to be wrong.
AI systems inherit analogous weaknesses. A system trained on biased data recognizes biased patterns. A system that excels in-distribution (on data similar to its training set) may fail catastrophically out-of-distribution (on data that differs in ways its pattern recognition cannot detect). Hallucination in large language models, the generation of confident, plausible, and factually incorrect outputs, is structurally analogous to the overconfidence bias in human System 1 processing.
The critical difference: human experts can, with effort, engage System 2 to check their intuitions. They can step back, apply deliberate analysis, and override a gut feeling that does not survive scrutiny. Current AI systems generally lack this metacognitive capacity. They produce outputs without the ability to evaluate whether those outputs are reliable.
The hybrid architecture. The most effective current AI systems may be those that combine System 1 and System 2 processing. Chain-of-thought prompting (2022-23) and test-time compute models (2024-25) approximate System 2 by forcing the model to reason step by step before producing an answer. The combination of fast pattern recognition (the base model) with slow deliberation (the reasoning chain) mirrors the dual-process architecture of human cognition.
The path to artificial intuition may require building systems that can combine fast pattern recognition with slow deliberation, and that can calibrate their confidence in each mode's outputs.
Implications for Problem-Solving
If artificial intuition is achievable, even in the narrow functional sense described here, it shows up first in a few kinds of work.
Cross-domain innovation. The most valuable insights often connect knowledge from different fields. A system that can identify non-obvious parallels between, say, materials science and epidemiology, or between urban planning and network theory, may produce the kinds of cross-pollinating insights that drive breakthrough innovation. Human experts are constrained by the breadth of their individual experience. An AI system trained on the full breadth of human knowledge has, in principle, access to a broader knowledge graph.
Decision support under uncertainty. In domains where decisions must be made faster than System 2 allows (emergency medicine, financial trading, military operations), artificial intuition may become an extra System 1 for human decision-makers. The system recognizes patterns in real-time data that human System 1 has not been trained to detect, and surfaces them for human evaluation.
Scientific discovery. If pattern recognition on large datasets can identify relationships that human scientists have not noticed (as AlphaFold did for protein structures), artificial intuition may accelerate discovery in domains where the relevant patterns are hidden in datasets too large for human analysis.
Chain-of-thought prompting and test-time compute already pair fast pattern recognition with a slow check, which is the part of Kahneman's dual-process model that transfers. The training-set advantage is largest across domains, because a human expert is bounded by one career of cases. Until a system can tell when its own match is unreliable, a person still has to run that check. Point the stack at datasets too large for a person to hold, then keep the override.