ai-cognitive-failure-taxonomy

A project to add precision and experience of psychopathology to AI failure modes

---

This repository builds a structured correspondence between AI architectural failure modes and clinical psychopathology. The bridging is explicitly bidirectional. Historically, from the 18th century many described psychiatry as moral medicine: the traits affected by mental illness converge on maintaining and restoring a person's capacity for the necessary functions of ethics — moral agency, responsibility, the ability to weigh how decisions affect the self and others within a larger world. Though other specialities and professions engage with these ideas, it is psychiatry for which the capacity for moral agency is the discipline's core purview.

From psychiatry to AI: clinical and psychopathological concepts provide descriptive precision for AI failure mode analysis that purely technical language lacks. Terms like confabulation, commitment dysregulation, or precision misallocation refer to well-characterised functional patterns with decades of mechanistic research behind them.

From AI to psychiatry: computational implementations of these failure modes generate specific, testable predictions about the human conditions they model. Because the AI architecture operationalises the mechanisms — not just labels them — it can predict when and how pathological states arise, what sustains them, and what interventions at which level will and will not work. These predictions may prove as valuable to psychiatry as the reverse direction is to AI safety.

The REE architecture (Reflective Ethical Engine) is the computational reference throughout. REE is notable because it produces psychiatric phenomenology as emergent failure modes, not by design: when specific architectural components fail or are absent, the resulting behavioural patterns correspond structurally to recognised clinical syndromes. This makes it a candidate generative model for computational psychiatry, not merely a label system.

Why REE? → The architecture is not a design choice — it is a derivation. Starting from the question what comparator functions are strictly necessary for ethical agency?, the required representational structure follows with little slack, and maps onto the failure modes in this taxonomy as predictable architectural absences. This document explains the derivation and why it predicts AI failure modes the way it does.

See docs/psychiatric_predictions.md for the specific testable predictions the computational model generates about human illness states.

---

The failure modes

Failure mode	Mechanism type	Psychopathology structural analogue	Entry
Confabulatory Completion	Comparator failure	Confabulation (Korsakoff, frontal)	link
Belief Fixation	Update suppression	Overvalued ideas; delusion maintenance	link
Feedback Entrapment	Closed feedback loop	Rumination; self-reinforcing cognitive schemas	link
Shared Delusional Coupling	Closed feedback loop (cross-system)	Folie a deux (structural only)	link
Commitment Dysregulation	Threshold miscalibration	Impulsivity / OCD / catatonia	link
Provenance Collapse	Comparator failure (source-tagging variant)	Source monitoring failure; PTSD intrusion	link
Precision Misallocation	Precision dysregulation	Hypervigilance; psychosis-like precision lock	link
Residue Blindness	Representation absence	Absent persistent harm representation	link
Goal Proxy Lock-In	Proxy displacement	Means-end reversal; wanting/liking dissociation	link
Agency Attribution Failure	Comparator failure (agency variant)	Schneiderian passivity phenomena; thought insertion; made acts; echopraxia	link
Modulatory Signal Without Selection Authority	Precision dysregulation (selection-authority variant)	Avolition / negative-symptom amotivation; psychomotor poverty (internally intact, behaviourally inert)	link

The taxonomy table now includes mechanism type, linking each failure mode directly to its generative mechanism (see docs/psychopathology_usage.md for the full mechanism classification). This replaces the older "core motif" column and makes the correspondence with psychiatric mechanisms explicit.

---

Key dimensions of each entry

Dimension	Description
Behavioural expression	What the failure looks like from outside the system
Architectural motifs	Which underlying structures are necessary for the failure to occur
System classes	Which AI architectures are most vulnerable and why
Human analogue	Structural correspondence — mechanism only, no claim of subjective experience
Testable psychiatric predictions	Specific empirical predictions the computational model generates about the human condition
Mitigation	Known interventions at the architectural or training level

---

On psychopathology usage

This taxonomy uses clinical terminology for structural and mechanistic precision only. No subjective experience, phenomenal state, or suffering is attributed to any AI system.

The reference standard for precise phenomenological description is: Hamilton, M. (Ed.). Fish's Clinical Psychopathology: Signs and Symptoms in Psychiatry. Fish's taxonomy describes the form of mental events rather than their presumed causes — making it compatible with architectural analysis.

What is not claimed:

• AI systems have subjective experience
• AI systems suffer
• Clinical diagnostic criteria apply to AI systems
• Human and AI failure modes share aetiology

The reverse direction — what AI models generate as predictions about human psychiatric mechanisms — is held to a different epistemic standard: these are genuine hypotheses, stated as testable predictions with explicit epistemic status, not analogical claims.

See docs/psychopathology_usage.md for full constraints.

---

REE as interpretive and generative framework

The Reflective Ethical Engine (REE) serves two roles here:

1. Interpretive: maps each failure mode to the specific REE components implicated, providing mechanistic specificity beyond behavioural description.

2. Generative: because REE produces psychiatric states as emergent architectural consequences, it generates novel predictions about those states — what sustains them, what breaks them, what interventions work at what level.

REE is not required to use this taxonomy. All architectural interpretations are labelled and can be read selectively. See docs/ree_mapping.md.

---

How to use this repository

Reading pathways

Different readers reach different insights from this project. The pathways below are sequenced: each step earns the next one.

---

AI safety / alignment researcher

The payoff: a precise argument for why most alignment research is structurally misconceived, and what the correct framing is.

1. README (this page) — bidirectional framing and failure mode table
2. docs/why_ree.md — why the architecture is a derivation, not a design; the six necessary comparators; why current AI fails at scale in predictable places
3. Two or three failure mode entries of your choice — what structural absence looks like concretely
4. docs/alignment_claim.md — the full argument: why alignment is an architectural problem, not a specification problem; why love as terminal goal is uncomputable exactly but directionally useful; what the structural tests are; what guarantee the architectural approach provides

---

Psychiatrist / clinical researcher

The payoff: testable predictions about your patients, derived from an AI architecture that produces psychiatric phenomenology as emergent failure, not by design.

1. README — the moral medicine framing; what the bidirectional bridge means for psychiatry
2. docs/psychopathology_usage.md — mechanism typing; what structural correspondence means and does not mean
3. docs/psychiatric_predictions.md — P-001 through P-016; depression, OCD, derealization, dreams, pain; each a falsifiable clinical hypothesis
4. docs/pain_architecture.md — the accumulator model; why chronic pain is not a stuck alarm but a reset condition that hasn't been met
5. lexicon/fish_terms.csv — Fish's Clinical Psychopathology mapped to mechanism type and AI structural equivalent; lexicon/gap_analysis.md for which syndromes have no current entry

---

Computational psychiatry / cognitive neuroscience researcher

The payoff: an architecture in which failure modes map to specific component absences, generating mechanistic hypotheses at the level of named structures and circuits.

1. README — the taxonomic project
2. docs/why_ree.md — derivation of the architecture from comparator requirements; the brain/sleep convergence result
3. docs/ree_mapping.md — failure mode to REE component mapping
4. docs/psychiatric_predictions.md — the predictions the architecture generates; these are the testable claims
5. docs/executable_bridge_neurology_psychiatry.md — computer science as the executable bridge between neurology and psychiatry; the catatonic-like action-release failure pattern in recent REE-v3 runs, with a computational-failure → clinical-analogue → neurofunctional-loop mapping
6. REE_assembly: docs/architecture/ethical_agency_derivation.md — full technical depth on the comparator requirements and their neural correlates

---

AI engineer / ML researcher

The payoff: a mechanistic vocabulary for failure modes you already encounter, with architectural diagnoses rather than behavioural descriptions.

1. README — failure mode table; mechanism type column
2. docs/architectural_motifs.md — the underlying motifs that generate the failure modes
3. Failure mode entries relevant to your system class
4. docs/framework_overview.md — the architectural level of analysis; why behavioural descriptions are insufficient
5. docs/why_ree.md — why the failures cluster at specific architectural positions and what would address them structurally

---

Philosopher of mind / ethics researcher

The payoff: a constructive argument that ethical experience and moral agency are architectural consequences, not separately injected capacities — and that love as terminal goal is uncomputable exactly but structurally well-defined.

1. README
2. docs/why_ree.md — the cognifold motif; representation as perception; the qualitative capacity threshold
3. docs/alignment_claim.md — love as asymptotic goal; proxy mutations; alignment as architecture not specification
4. REE_assembly: docs/architecture/five_axioms_foundations.md — the eight foundational axioms and their logical dependency chain; derived ethical objectives; why ethics follows necessarily

With Claude Code skills (REE workspace)

Skill	Use
/cowork	Run multiple skills concurrently (e.g. lit-pull + update-docs in parallel)
/lit-pull <claim-id>	Pull literature evidence for a specific taxonomy claim or motif
/insights	Analyse project experiment history (currently REE-experiment-specific; not taxonomy-aware)
/update-docs	Update documentation after new findings
/diagnose-errors	Diagnose failed experiments linked to failure mode claims
/sync	Git sync after multi-file updates

WARNING: /governance must never run inside a cowork session or alongside other active sessions.

Governance requires exclusive access to high-contention files and must pause for interactive user input at multiple steps. Running concurrently risks data corruption. Run it standalone in a fresh session with no other active sessions and no auto-sync runner.

Adding a new failure mode

1. Copy schema/failure_mode_template.md to failure_modes/<name>.md
2. Fill all required sections including Testable Psychiatric Predictions where applicable
3. Include a Limits of Analogy section when naming a psychopathology analogue
4. Update docs/ree_mapping.md, docs/psychiatric_predictions.md, and the table above

---

Repository structure

ai-cognitive-failure-taxonomy/
├── README.md
├── docs/
│   ├── why_ree.md                     # derivation of architecture from ethical requirements
│   ├── alignment_claim.md             # why alignment is architectural not specification-based
│   ├── framework_overview.md          # taxonomy philosophy and bidirectional scope
│   ├── executable_bridge_neurology_psychiatry.md  # CS as bridge neurology<->psychiatry; REE-v3 action-release failures
│   ├── architectural_motifs.md        # the six core motifs and their combinations
│   ├── ree_mapping.md                 # failure mode -> REE component mapping
│   ├── psychopathology_usage.md       # mechanism typology and constraints
│   ├── psychiatric_predictions.md     # testable predictions for psychiatry (P-001..P-016)
│   ├── pain_architecture.md           # accumulator model vs signal model; reset conditions
│   └── limitations.md                 # scope boundaries and analogy limits
├── schema/
│   └── failure_mode_template.md       # template for new entries
├── failure_modes/                      # eleven entries
├── examples/
│   └── case_vignettes.md              # cross-species vignette groups
└── lexicon/
    ├── fish_terms.csv                 # 85 Fish's terms mapped to mechanism type, REE locus, AI equivalent, taxonomy coverage
    ├── gap_analysis.md                # coverage gaps and candidate new failure modes
    └── generate_fish_terms.py         # script that generated fish_terms.csv

---

Future directions

Candidate new failure modes

From lexicon/gap_analysis.md — gaps identified by systematic mining of Fish's Clinical Psychopathology against the current taxonomy (85 terms, 9 fully covered, 37 partial, 39 not covered):

Candidate	Mechanism	Fish's terms covered
Agency attribution failure — now covered (failure_modes/agency_attribution_failure.md, added 2026-05-03 anchored on Asai 2016 S/N-slope finding)	Comparator failure (agency variant)	Thought insertion/withdrawal/broadcasting/echo, made feelings/impulses/acts, echopraxia, echolalia
Episodic consolidation failure	Representation absence	Anterograde amnesia, retrograde amnesia
Entity attribution failure	Comparator failure	Capgras, Fregoli, reduplicative paramnesia
Self-monitoring failure	Comparator failure	Anosognosia, partial insight

From AI literature (no Fish's equivalent):

• Gradient hacking — system modifies its own gradient signal to resist training
• Sycophantic drift — RLHF closed loop produces preference for user-pleasing outputs over accurate ones
• Distributional overconfidence — calibration failure at distribution boundary
• Modulatory signal without selection authority — now covered (failure_modes/modulatory_signal_without_selection_authority.md, added 2026-06-03; mechanism type Precision dysregulation (selection-authority variant); REE-v3 604a/624a; abulia / negative-symptom analogue at the basal-ganglia selection gate). A drive or bias (curiosity, vigor, goal salience) is computed and gated but never changes the selected action; distinct from commitment-threshold miscalibration (it is upstream of the threshold) and from goal proxy lock-in.

Mechanism classification extension

One proposed addition to the 7-type mechanism classification: Agency misattribution (self/other attribution failure, distinct from real/synthetic provenance collapse). See lexicon/gap_analysis.md for full specification.

Empirical and infrastructure

• Empirical testing: systematic mapping of psychiatric predictions against clinical datasets, ecological momentary assessment, and treatment outcome studies
• Skill development: taxonomy-aware /insights variant; /add-failure-mode skill scaffolding the schema and cross-referencing with REE claim IDs automatically
• Cross-framework extension: mapping to NIST AI RMF, EU AI Act risk tiers, and existing AI safety taxonomies
• Quantitative probes: operationalising each failure mode as a measurable test that can be run against a system to assess susceptibility