LLMs and Human Intuition: Evolutionary Pathways and Complementary Spaces
v2.0 | June 2026 | arXiv Preprint | PDF Download | GitHub | OSF
What Question Are We Trying to Answer
Large language models are permeating every facet of human cognition. But where do they stand on intuition — those near-instinctive cognitive capacities that require neither reasoning nor analysis?
Intuition is not a mysterious sixth sense. It is compressed experience: a chess player who "just sees" the right move after thousands of games, a radiologist who spots an anomaly at a glance, the uneasy feeling about someone you "can't quite explain." These judgments share three features: fast, effortless, no explicit reasoning required.
The core question is not "do LLMs have intuition" — but "what do they lack, to what degree, and are there alternative paths?"
Three Core Propositions
Proposition 1: LLMs lack endogenous cost signals. Human intuition is forged by "make a mistake → get punished → brain tags which patterns to prioritize." An LLM's "punishment signal" is the human preference rating in RLHF — but "this answer looks correct" ≠ "this advice produced good real-world outcomes." LLMs do not experience the consequences of their outputs.
Proposition 2: LLM attention tries not to miss anything, rather than selectively ignore. Expert intuition is not about "seeing what matters" — it is about "instantly ignoring what doesn't." Humans evolved cost-driven selective ignoring under survival pressure; LLM sparse attention variants are motivated by "we can't afford to compute this," not "this isn't worth looking at."
Proposition 3: The body's core contribution is the inescapability of cost signals. When your body tells you something is wrong (racing heart, stomach tightening), you cannot dismiss it by turning down a parameter. Every LLM "importance" parameter — learning rate, temperature, reward scale — is exogenous, adjustable by engineers. The body's value is not the signal content, but the guarantee that the signal must be received.
Propositions 1 and 2 converge on the same underlying deficit from two different disciplinary traditions — cognitive science and machine learning. They are mutually reinforcing, not independent.
Four Intuition Subtypes
| Type | Examples | LLM Substitutability |
|---|---|---|
| Perceptual | Chess pattern recognition, radiology anomaly detection | ⚠️ Partial |
| Conceptual | Mathematician's "sense of direction," scientific taste | ⚠️ Closed domain yes, open domain limited |
| Social | Reading people, judging trustworthiness, sensing atmosphere | ⚠️ Textually accessible, embodied inaccessible |
| Moral | Instant "this is wrong" first-pass judgment | ❌ Structurally inaccessible |
Key insight: As costs move from internal (efficiency) to epistemic (wasted exploration) to interpersonal (rejection, shame) to identity (moral selfhood), they become increasingly embodied and irreplaceable by pure information processing.
Three Hypotheses, Nine Core Judgments
Hypothesis A (Intuition Prosthetic): LLMs fill human intuition gaps via different computational paths
Hypothesis B (Degradation Risk): Long-term LLM dependence weakens human intuition
Hypothesis C (Autonomous Intuition): LLMs may develop functionally equivalent intuition
Below are the three judgments most relevant to product design:
Judgment 1: Degradation exists but is not inevitable — it depends on usage mode. Three mediating variables determine degradation: (a) whether AI replaces intuition execution or verification (humans judging first then checking AI shows significantly less degradation than AI answering first); (b) usage frequency and temporal structure (asking AI for everything vs. only when uncertain); (c) the inescapability of the intuition subtype (you cannot pause to check AI during real-time social interaction).
Judgment 2: The Human-First Protocol is the single most effective intervention. Radiology on-demand AI shows upskilling rather than deskilling. A meta-analysis of 35 studies confirms AI-first protocols significantly increase automation bias compared to human-first protocols. The critical point: nearly all current LLM products default to AI answering first — this is a systematic amplifier of degradation risk. Changing the default requires product design decisions, not technological breakthroughs.
Judgment 3: Social intuition's "hollow period" is the most dangerous imbalance zone. The hollow period occurs when human capability has already degraded but AI is not yet mature enough to fully replace it across all scenarios — and the degradation is masked by AI's good performance in routine cases. Social intuition simultaneously satisfies the most dangerous conditions: LLMs have prosthetic-level usability (making people willing to use them), systematic blind spots (sycophancy, no real-time social calibration), and human degradation is accelerating under the dual pressure of social media and LLM adoption.
Three Iron Laws
- Let humans judge first. The Human-First Protocol is the default for all scenarios — not out of distrust for AI, but to ensure humans don't forget they can still judge independently.
- Make costs inescapable. Key domains must preserve the inescapability of judgment consequences, analogous to FAA requirements for pilots to fly manually.
- Make boundaries clear. No deployment of autonomous moral judgment systems or real-time social judgment substitution systems — degradation in these domains is identity abdication, not skill loss, and is extremely difficult to reverse.
An Honest Footnote
Of 41 evidence-level annotations across the full paper, L5 (confirmed by ≥2 independent laboratory replications) accounts for 0%. These claims are based on theoretical deduction and analogical extrapolation — not "unreliable," but at the normal early stage of theory building. We will not pretend to know more than we do.
Cross-Validation
Three independent theoretical traditions converge on the same structural judgment:
- Predictive Processing (this paper): LLMs lack endogenous cost-driven signal prioritization — precision-weighting is exogenous
- Phenomenology (Rafiee & Sutton, 2026, arXiv:2605.24238): LLMs lack four Enactive AI dimensions — experience, action-perception inseparability, autonomy, embodiment
- Mathematics (Klindt, LeCun & Balestriero, 2026, arXiv:2605.26379): non-Gaussian latent variables make linear identifiability impossible — World Models cannot recover true structure in non-Gaussian domains
The division of labor: JEPA provides the formal proof, Rafiee & Sutton provide the philosophical diagnosis, and this paper provides the systematic mapping and intervention framework. See Appendix D of the full paper for a detailed cross-mapping.
Resources
- 📄 Full Paper PDF: Chinese | English
- 💻 GitHub: co-cognition-lab/llm-intuition
- 📊 OSF: 10.17605/OSF.IO/5FSCZ (v2.0 snapshot) | Project: 10.17605/OSF.IO/XSY39
📋 Historical version: v1.3 (May 2026, arXiv preprint anchor version)