The Idea
There is a set of cognitive capacities — causal reasoning, plausibility auditing, problem formulation, interpretive judgment, metacognitive supervision, collective epistemic function — that machines currently cannot perform reliably, and that the standard curriculum does not teach systematically.
These are not soft skills. They are the specific forms of intelligence that allow a person to use a powerful AI tool rather than be used by it. Irreducibly Human names them, teaches them, and demonstrates each one by the method used to build it.
Stop teaching people to be slower calculators. Start teaching them to be better question askers — specifically, the questions machines cannot yet answer.
The Seven Tiers
A taxonomy of human intelligence, organized by how far machines have reached into each tier. The first three are where machines compete or convincingly imitate. The last four — highlighted — are where they have not arrived, and where the educational payoff is highest.
Tier 1 — Pattern & Association
Machines: Superhuman
Statistical pattern-matching at scale — classification, prediction, retrieval, generation. Humans competing here is malpractice.
Deprioritize. Machines do this better.
Tier 2 — Embodied & Sensorimotor
Machines: Improving, bounded
Physical skill and sensorimotor intelligence. Machines are improving but remain bounded by the constraints of physical deployment.
Context-dependent. Not the series focus.
Tier 3 — Social & Personal
Machines: Simulacra only
Genuine relationship, trust, and interpersonal judgment. Machines produce convincing simulations of connection, not connection.
Important, but not the series focus.
Tier 4 — Metacognitive & Supervisory
Machines: Weak
Plausibility auditing, problem formulation, knowing when to distrust the output. These require a model of the machine’s limits that the machine cannot supply.
High priority. Underscaffolded in every current curriculum.
Tier 5 — Causal & Counterfactual
Machines: Unreliable
The causal parrot problem: language models reproduce causal-sounding language without performing causal reasoning. Variable selection, edge orientation, and conditioning decisions still require domain judgment.
High priority. The gap between predictive and causal intelligence is the decisive engineering problem.
Tier 6 — Collective & Distributed
Machines: Absent
Collective intelligence emerges from systems of people in relationship; it cannot be compressed into training data. The collaborative friction that refined our ideas is not in the weights.
High priority. Unaddressed by individual-focused intelligence frameworks.
Tier 7 — Existential & Wisdom
Machines: Absent
Interpretive judgment, values integration, the capacity to know which question is worth asking. These require a self that machines do not have.
High priority. The least-taught tier of all.
The taxonomy derives from “Knowing Enough to Distrust the Machine” (Theorist.ai, March 2026).
The Meta-Principle
The process of building each part of Irreducibly Human is an instance of the methodology it describes. A book on causal reasoning is built through structured causal reasoning. A book on plausibility auditing is built through explicit plausibility auditing. The argument evolves through conversation with conversational AI, and the curriculum takes shape as the design is debated. What you are looking at is the working document, not the finished product. That's the point.
The First Instrument
The AI Exposure Explorer
The tiers are a claim, and claims should be testable. The AI Exposure Explorer puts the thesis against real labor data: pick any two occupations and compare their employment trends against major AI milestones, then compare the human abilities each job leans on — asking which of those abilities machines have reached, and which they haven't.