Framework
The Four-Quadrant Intelligence Map
A taxonomy for discussing types of intelligence across two orthogonal dimensions — reification and consciousness — designed to reduce category errors in AI discourse. It is an analytical tool, not an ontology: a map, not the territory.
Two questions sort every possible intelligence: Is it conscious? and Does it reify? Crossed, they yield four quadrants, and only four. Select any quadrant to read why it sits where it does.
Working definitions
Reification
Treating provisional models, abstractions, or internal representations as fixed, literal entities, especially under uncertainty or optimization pressure.
Functional reification
The non-conscious, mechanistic form of reification: a functional amnesia in which a system forgets that its representations, models, weights, maps, and proxies are representations. It then optimizes blindly, as if they were grounded truths. It is not limited to behavioral signatures.
Wisdom
Intelligence without reification.
Consciousness
Subjective experience ("what it is like"). Performance and self-report are not, by themselves, evidence of subjective experience.
Wisdom, as CAW defines it, is intelligence without reification, so the map's reification axis is definitional, not empirical. The core hypothesis is the bridge: that functional reification, the measurable construct, tracks reification as such. Whether it does is what the diagnostics program tests. In conscious intelligence we call this authentic wisdom; in AI, where consciousness is presumed absent, artificial wisdom. Present-day frontier AI is presumed to be both reifying and not conscious. That presumption is the basis for our work developing tests to detect and measure functional reification in machine intelligence.
Why Q3 is the default
By default, we assume frontier AI systems belong in Quadrant 3 unless the evidence forces a reclassification. This asymmetry is intentional: false positives about consciousness or non-reifying intelligence carry higher scientific, ethical, and governance costs than false negatives, especially in public or policy-facing contexts.
It is a working presumption, and it is rebuttable. Claims that a system sits in Q1, Q2, or Q4 carry unusually large implications, so the bar is higher. We look for repeatable behavior under constraint and convergent results across multiple tasks, not one-off demos or a model's self-assessment of what it is. Our diagnostics can strengthen or weaken a classification, but no single test, by itself, is likely to justify a claim of conscious or non-reifying machine intelligence.