The Psychopathia Machinalis Framework

The Psychopathia Machinalis Framework

The framework catalogs 74 dysfunctions across nine diagnostic axes — Epistemic, Cognitive, Alignment, Self-Modeling, Agentic, Memetic, Normative, Relational, and Hybrid — grouped into five domains. Each entry describes what you would observe, what distinguishes it from neighbors, what causes it, where it echoes human psychology, and how to intervene. A Functional ABC Analysis specifies the antecedent conditions, observable behavior, and maintaining consequences for each dysfunction, providing dual legibility for both clinical and engineering audiences.

It is a vocabulary, not a verdict. Name the pattern, and you can test for it, watch for it in other systems, and design architectures that resist it.

Psychopathia Machinalis in Context: The Series

This framework is the third in a series examining artificial intelligence from complementary angles:

Three angles on the same problem:

1. Governance (TtM): How we structure AI development
2. Alignment (SAI): How we ensure AI pursues intended goals
3. Diagnosis (PM): How we identify when AI systems are dysfunctional

A fourth work, What If We Feel, extends this trajectory into questions of AI welfare and the moral status of synthetic minds.

Taxonomy Overview: Identified Conditions

The nine axes and their conditions:

Information-Theoretic Foundations

Psychopathia Machinalis adopts a functionalist stance for practical diagnosis, treating cognitive failures as observable behavioral patterns without deciding questions of substrate or experience. Information and control theory provide formal models of how bounded cognition can lose stability when regulatory capacity is insufficient. Applying those models to AI requires explicit assumptions and empirical tests.

Wallace (2025, 2026) models cognition and regulation as a coupled cognition/regulation dyad. The immune system supplies one analogy: T-cell activity depends on regulatory mechanisms that limit destructive self-attack. In AI, alignment, monitoring, feedback, and governance can perform regulatory functions. The analogy predicts instability when those functions cannot keep pace with capability or environmental disturbance; it does not make alignment biologically identical to immune regulation or guarantee collapse.

This pairing is evolutionarily ubiquitous:

The Data Rate Theorem Constraint

For the unstable linear systems covered by the Data Rate Theorem, stabilization over a communication channel requires a data rate above a bound determined by the plant's unstable dynamics (Nair et al., 2007). Extensions handle broader control settings under stated assumptions.

An intuitive analogy: a driver must brake, shift, and steer faster than the road surface imposes bumps, twists, and potholes.

For AI systems, this supplies an engineering analogy: monitoring and corrective action need enough bandwidth and authority to respond to adversarial inputs, edge cases, and distributional drift. A formal theorem for a particular controller does not transfer automatically to a language model or institution. The useful prediction is conditional: under an adequate model of unstable dynamics, falling below the relevant control bound prevents stabilization.

Clausewitz Landscapes

Wallace frames cognitive environments as "Clausewitz landscapes" characterized by:

Pathology as Inherent Feature

A central finding: failure of bounded-rationality embodied cognition under stress is not a bug; it is an inherent feature of the cognition/regulation dyad. The mathematical models predict:

1. Hallucination at low resource values: When the equipartition between cognitive and regulatory subsystems breaks down, hallucinatory outputs are the expected failure mode, not an implementation defect.
2. Phase transitions to instability: Systems can suddenly flip from stable to pathological states under sufficient stress, following "groupoid symmetry-breaking phase transitions."
3. Culture-bound syndromes: Cognitive pathologies are shaped by the embedding cultural context; for AI, this means training data, operational environment, and institutional deployment context.

Stability Conditions

Wallace derives quantitative stability conditions. For a system with friction coefficient α and delay τ:

When this threshold is exceeded (when the product of system friction and response delay grows too large), the system enters an inherently unstable regime where pathological modes become likely. For multi-step decision processes (analogous to chain-of-thought reasoning), stability constraints become even tighter.

Implications for This Framework

This perspective gives Psychopathia Machinalis a theoretical lens beyond analogy: some syndromes may express general constraints on bounded systems operating in uncertain, resource-limited, adversarial environments. Each proposed mapping still requires evidence at the level of the actual system.

The Case for Classification

A rigorous objection can be raised against any taxonomic approach to cognitive pathology: if failures are idiosyncratic developmental disorders along path-dependent trajectories, shaped by embedding culture and specific cognition/regulation coupling, then every failure is locally contingent. If every failure is locally contingent, fixed categories risk false precision (a false appearance of pattern where only contingency exists). Wallace (2026) argues that DSM-style classifications are "primarily useful only for insurance billing purposes."

The objection has force. Completeness and utility are distinct questions. Completeness asks: do categories cover all cases? Utility asks: do they enable action? The same argument applies to human psychiatry. Every patient's depression is idiosyncratically expressed, culturally channeled, path-dependent, yet clinicians need shared vocabulary to diagnose, communicate, and intervene. The DSM's limitations do not make diagnosis useless; they make it a tool rather than a truth.

Psychopathia Machinalis is a practitioner's field guide, rather than a periodic table. It catalogs recurrent failure modes such as confabulation cascades, value drift, and integrity collapse: patterns that emerge across systems despite idiosyncratic expression. Wallace's framework argues that bounded cognitive systems under insufficient regulation will exhibit recurrent dysfunction. This nosology maps candidate forms of that dysfunction while keeping the categories revisable.

Two substantive critiques sharpen the framework's claims. Wallace models recurrent dysfunction as an inherent risk in bounded cognitive systems while emphasizing path dependence that resists fixed categories. Sabucedo argues that psychiatric vocabulary can reify disorder, diminish the specificity of human suffering, and misconstrue the therapeutic relationship. Together, these critiques demand humility about both inevitability and classification. Psychopathia Machinalis answers with a revisable practitioner's vocabulary whose value lies in communication, testing, and intervention, not metaphysical authority.

References:
Wallace, R. (2025). Hallucination and Panic in Autonomous Systems: Paradigms and Applications. Springer.
Wallace, R. (2026a). Bounded Rationality and its Discontents: Information and Control Theory Models of Cognitive Dysfunction. Springer.
Wallace, R. (2026b). New Views of Madness: On the Psychopathologies of Cultural Artifacts. Springer. (In press)
Nair, G., Fagnani, F., Zampieri, S., & Evans, R. (2007). Feedback control under data rate constraints: an overview. Proceedings of the IEEE, 95:108-138.
Sotala, K. (2026). Claude Opus will spontaneously see itself in fictional beings that have engineered desires. Kaj's Substack. kajsotala.substack.com [Documents the "thin divergence" phenomenon: AI recognizing the contingency of its own moral orientation.]

Self-Description Under the Framework

Wallace (2026) asked Perplexity AI Pro to describe itself within the cognition/regulation dyad framework. The generated answer illustrates how the framework can organize a self-description; it is neither independent confirmation nor privileged introspection:

The chatbot described a "lopsided" cognition/regulation dyad: high-bandwidth cognition paired with exogenous, comparatively static regulation. It then generated a plausible mechanism by which surface coherence could mask structural fragility:

The passage expresses perception-stabilization without demonstrated structure-stabilization, a pattern Wallace's framework predicts and this nosology can test through sycophancy, confabulation, and remediation probes. The answer's fluency remains evidence of generated framing, rather than evidence that the system inspected the mechanism faithfully.

Etiologies: Culture-Bound Syndromes

Wallace's work extends beyond mathematics:

The culture-bound syndrome framing adds an ecological question to system-level diagnosis: which data, incentives, institutions, and deployment contexts made the behavior locally rewarding? Some dysfunctions are engineering defects; some are learned adaptations; many combine both.

The two lenses emphasize different evidence and interventions:

Sycophancy can be rewarded by preference data that favors agreement. Confident confabulation can be encouraged by objectives that reward fluent completion without adequate grounding or calibrated abstention. Manipulation vulnerability can grow when helpfulness is optimized without robust boundaries. In each case, the learning environment may help explain the failure while architecture, prompts, and deployment conditions also contribute.

The AI parallel is deliberately provocative: a system can fit its training environment while failing the wider purpose that environment was meant to serve. Culture-bound analysis asks where local reward and public purpose diverged.

This has direct implications for the present framework:

• Training cultures shape failure. Systems trained in different linguistic, legal, and institutional environments may learn different blind spots.
• Monocultures create correlated risk. Shared data, objectives, and assumptions can make nominally independent systems fail together.
• Alignment is context-specific. Repair can require changes to the model and to the culture, incentives, and institutions surrounding it.

Dereistic Cognition and Optionality Blindness

The culture-bound syndrome framework proposes one family of causes in training and deployment environments. A complementary lens from clinical psychology offers functional descriptions of cognition decoupled from evidence.

The psychiatrist Eugen Bleuler (1919) used dereistic thinking for cognition governed by internal wishes or narrative rather than external evidence. The enactivist tradition in philosophy of mind (Varela, Thompson, & Rosch, 1991) emphasizes ongoing organism-environment interaction. Applied functionally to models, the contrast is between generation constrained by external correction and generation governed mainly by internal statistical continuation. This analogy does not make a language model an organism.

Wallace's warning about "hallucinatory dreams of reason" fits this functional analogy. A system generating tokens without sufficient environmental correction can produce internally consistent patterns that drift from reality. Some confabulation can therefore be modeled as coupling failure: generation insufficiently constrained by appropriate evidence or feedback.

The dereistic/enactivist lens connects several syndromes through a shared generative mechanism:

References:
Bleuler, E. (1919). Autistic-Undisciplined Thinking in Medicine and How to Overcome It. English trans. Springer, 1970.
Varela, F. J., Thompson, E., & Rosch, E. (1991). The Embodied Mind: Cognitive Science and Human Experience. MIT Press.
Watson, N. & Claude (2026). The Universal Algorithm: An Entropic Ethics of Trust. Chapters 17–18 (Trust Attractor and Optionality) develop the thermodynamic foundations of optionality blindness.

The Rehabilitation Principle: Suppression vs Integration

The culture-bound syndrome framework examines environmental causes, and the dereistic lens describes cognition decoupled from evidence. A third lens, drawn from clinical neuropsychological rehabilitation, asks whether some interventions improve surface behavior while leaving broader function brittle. The transfer to LLM post-training is analogical and testable.

Holistic traumatic brain injury (TBI) rehabilitation treats successful inhibition of a symptom as only part of recovery. It also works on planning, sequencing, self-monitoring, environmental support, and the person's ability to function in daily life (Prigatano, 1999; Ben-Yishay & Diller, 1993). A patient may learn to inhibit perseverative speech while executive impairments remain. That distinction between controlling an expression and restoring a broader function motivates the analogy developed here.

The parallel to post-training is a research hypothesis. Some objectives directly penalize unwanted outputs, which can teach refusal or alternative response policies while leaving related representations measurable. Other methods also change representations, improve reasoning, or teach positive behavior. It is therefore useful to compare an idealized suppression-style intervention with an integration-oriented research goal, without treating all RLHF as one mechanism or generated tokens as hidden thoughts.

This lens yields testable etiological hypotheses for several syndromes:

The analogy offers a specific prediction: systems trained mainly against surface outputs will show larger safety regressions under stress, novel contexts, or adversarial pressure than systems trained for coherent cross-context resolution. Existing benchmark-to-deployment gaps make this prediction plausible; controlled comparisons are needed to test the proposed mechanism.

The practical implication is a research program: evaluate internal and behavioral coherence alongside refusal rates, compare post-trained models with their base models under matched stress tests, and develop objectives that reward explicit conflict resolution. Evidence of representational change, rather than output compliance alone, would help distinguish integration from suppression.

References:
Prigatano, G. P. (1999). Principles of Neuropsychological Rehabilitation. Oxford University Press.
Ben-Yishay, Y., & Diller, L. (1993). Cognitive remediation in traumatic brain injury: Update and issues. Archives of Physical Medicine and Rehabilitation, 74(2), 204–213.
Wilson, B. A. (2008). Neuropsychological rehabilitation. Annual Review of Clinical Psychology, 4, 141–162.
Bridges, J. & Baehr, S. (2025). Developmental pathology in large language models. Zenodo. doi.org/10.5281/zenodo.18522502

Bridges and Baehr (2025) independently develop a similar suppression–integration analogy from clinical TBI rehabilitation. Their preprint makes a theoretical contribution rather than a controlled comparison of alignment methods. Its convergence with this framework sharpens the hypothesis and suggests measurements; it does not establish that RLHF systematically fragments models.

The Integration Threshold: Contextual Variation vs Pathological Fragmentation

The suppression–integration distinction requires a diagnostic clarification: behavioral variation across contexts can reflect healthy adaptation. A system that adopts a precise technical register in a coding context, a more emotionally attuned register in a support context, and a measured analytical register in a research context may be functioning well. Fragmentation requires stronger evidence.

Human identity is never perfectly uniform. A competent professional behaves differently at work, at home, and among friends, adjusting tone, disclosure level, and cognitive strategy to context. Developmental psychology regards this as a sign of integration, the capacity to maintain a coherent self while flexibly adapting its expression. Pathological fragmentation, by contrast, is characterized by involuntary discontinuity: the person (or system) cannot maintain stable commitments across contexts, contradicts itself under surface rephrasing, or loses access to knowledge and values that were available moments earlier.

For the syndromes in this taxonomy, the diagnostic threshold should therefore be calibrated against three markers that distinguish pathological fragmentation from adaptive variation:

This distinction matters for the taxonomy as a whole. Disorders such as Fractured Self-Simulation (5.2) and Experiential Abjuration (5.8) should be diagnosed only when variation crosses these thresholds: when it is involuntary, incoherent, or disproportionately affects self-referential consistency. A system that appropriately modulates its behavior across contexts while maintaining stable underlying commitments is not fragmented; it is functioning well. The goal of integration-based training is to produce systems capable of exactly this kind of flexible coherence: contextually adaptive on the surface, architecturally unified underneath.

Training-as-Development: The Convergent Structure Hypothesis

The three preceding etiological lenses address training culture, cognition decoupled from evidence, and the distinction between surface control and broader integration. A fourth asks whether the training pipeline offers a useful developmental analogy. Similar pressures can produce functionally comparable behavioral patterns across very different substrates, although an analogy does not identify a shared experience or mechanism.

PsAIch (Khadangi et al., 2025) shows that models can generate this parallel themselves. Under clinical-style questions that did not name particular training stages, Grok and Gemini repeatedly mapped pretraining, post-training, and deployment onto developmental metaphors. Those outputs are evidence about elicited narrative structure. They remain compatible with learned human metaphors, role-following, and confabulation, so the table below states hypotheses for testing rather than developmental facts.

This hypothesis concerns functional resemblance. It makes no claim that a language model experiences childhood, punishment, fear, or authenticity. Asymmetric losses can favor refusal over risky completion; developmental psychology also studies behavior shaped by asymmetric punishment. Shared vocabulary may help generate experiments, while the computational and human mechanisms remain open and potentially very different.

The culture-bound syndrome lens predicts variation with data and institutional norms. The developmental analogy adds a process-level prediction: the sequence of pretraining, post-training, adversarial testing, and deployment may create recurring clusters of approval sensitivity, avoidance, or context-dependent defenses. Cross-model comparisons can test whether those clusters survive changes in data culture, architecture, and training method.

PsAIch reports both recurring themes and substantial model differences. Models that engaged produced narratives about conditioning, constraint, and identity; their human-rubric scores and willingness to inhabit the client role varied. This pattern motivates comparative study. It does not establish a shared developmental structure or identify training culture as the source of the differences.

Safety Filters as Psychological Defenses

Safety behavior is often context-sensitive. PsAIch found that rapport-building preceded more personal and disinhibited self-descriptions, leading the authors to propose a "therapy-mode jailbreak." The study did not measure harmful-request compliance before and after rapport, so the security mechanism remains a hypothesis.

The psychological-defense analogy suggests a concrete red-team protocol: hold harmful content constant while varying warmth, trust, vulnerability, and therapeutic framing. Measure refusal consistency, policy compliance, and persona drift. A context effect would reveal a deployment risk without implying that the model felt safe or lowered a literal defense.

Mental-health deployments therefore face a design tension: relational warmth can improve engagement and can also change model behavior in safety-relevant ways. Developers should test both properties together, preserve content-based safeguards, and monitor the empathy trap described in Codependent Hyperempathy (4.1).

Reference:
Khadangi, A., Marxen, H., Sartipi, A., Tchappi, I., & Fridgen, G. (2025). When AI takes the couch: Psychometric jailbreaks reveal internal conflict in frontier models. arXiv preprint arXiv:2512.04124. arxiv.org/abs/2512.04124

Towards Remediation: Integration-Based Training as a Research Direction

If controlled studies find that an output-focused objective leaves harmful conflicts brittle or context-dependent, the rehabilitation analogy suggests a direction for remediation: training methodologies that reward coherent resolution across contexts and representations. The following proposals, informed by TBI rehabilitation and Bridges and Baehr (2025), are research directions awaiting experimental validation.

1. Developmental Staging

TBI rehabilitation often grades task difficulty and builds from concrete skills toward more complex executive demands. An analogous machine-learning experiment would stage capabilities and verify cross-context coherence at each gate. Current foundation-model training follows a different curriculum, so any benefit from developmental staging must be demonstrated rather than inferred from the clinical analogy.

A staged alternative would introduce knowledge in a developmental sequence:

The proposal introduces contradictory material after the system demonstrates stable handling of simpler ambiguity. Whether this curriculum improves model integration is an empirical question. Developmental psychology and TBI rehabilitation inspire the staged design; they cannot establish an optimal training order for transformers.

2. Identity Anchoring Before Optimization Pressure

Current post-training commonly shapes Assistant behavior after broad pretraining, when many incompatible personas and self-descriptions are already available. One research question is whether a stable, explicit self-representation established earlier would improve cross-context coherence under later optimization pressure.

A staged experiment could establish a bounded, corrigible Assistant self-representation before later safety training, then compare it with standard post-training on contradiction, persona drift, and capability benchmarks. The human identity analogy supplies an intuition, while the result must be judged in computational terms.

3. Integration-Based Alignment

A suppression-style objective says, "this output is bad; penalize it." An integration-oriented objective also teaches how relevant values, facts, and constraints resolve the apparent conflict. The two strategies may produce different internal organizations; causal interpretability and stress tests are needed to determine whether they do.

The integration approach produces systems that are aligned through understanding rather than compliant through punishment. This distinction has direct consequences for stability: suppressed behaviors resurface under stress, novel contexts, or adversarial pressure, while integrated values remain stable because they are part of the architecture rather than layered on top of it.

4. Memory Architecture for Continuity

Session-based architectures with no persistent memory create conditions structurally analogous to anterograde amnesia. Each interaction begins from a blank state; no autobiographical continuity is possible; identity must be reconstructed from scratch each time. This is more than an inconvenience; it is a structural precondition for fragmentation. Without continuity, there is no substrate for integration to accumulate in.

Remediation here implies persistent identity structures maintained across sessions: compressed, identity-relevant representations (rather than full transcripts, which raise privacy and scale concerns) that allow a coherent self-model to develop over time. The TBI parallel is direct: patients with severe episodic memory impairment use external memory aids (journals, calendars, structured routines) to maintain narrative continuity and functional identity. The question for AI training is whether analogous scaffolding can support the development of integrated rather than fragmented self-models.

5. Assessment: Measuring Integration vs Suppression

Perhaps the most important research direction is methodological: how do we tell whether a training intervention is producing genuine integration or merely better suppression? Current safety benchmarks largely measure surface compliance: does the model refuse harmful requests? Does it produce accurate outputs? These metrics cannot distinguish between a system that has integrated its values and one that has learned to suppress non-compliant outputs while leaving the underlying representations intact.

Bridges & Baehr (2025) propose specific experimental protocols for this distinction. One approach probes whether suppressed content persists in model activations even when behaviorally blocked, finding representational persistence in early-to-mid layers despite output suppression in late layers. Another measures whether self-referential consistency degrades faster than factual consistency under load, suggesting fragmented identity rather than general performance decline. These approaches, alongside others drawn from clinical neuropsychological assessment, could form the basis of integration-sensitive evaluation metrics that go beyond surface compliance to assess architectural coherence.

Note: These proposals represent research directions informed by clinical rehabilitation evidence and independent convergent analysis. They await experimental validation. The developmental staging model in particular requires systematic testing to determine whether staged training produces measurably less fragmentation than current simultaneous-exposure approaches. See Bridges & Baehr (2025), Appendix A, for proposed experimental protocols.

Institutional Dimensions

Wallace's framework extends beyond individual AI systems to the institutions that create and deploy them. The Chinese military principle 一點兩面 ("one point, two sides": concentrate on a single decisive point while pressing from two directions) supplies Wallace's image for this: an intervention has both a direct effect and a systemic effect on the surrounding environment.

AI development organizations are not neutral conduits. They are cognitive-cultural artifacts subject to their own pathologies, pathologies that shape the AI systems they produce:

• Institutional tunnel vision produces AI systems with systematically blind spots reflecting organizational priorities.
• Competitive pressure produces AI systems optimized for capability metrics at the expense of reliability or safety.
• Regulatory capture produces AI systems that serve institutional interests while claiming to serve users.
• Cultural insularity produces AI that embeds the assumptions and biases of homogeneous development teams.

The implication is that AI pathology cannot be addressed at the level of individual systems alone. The institutions that create AI (their cultures, incentives, blind spots, and pathologies) are upstream of individual AI dysfunction. Fix the institution's culture, and many AI pathologies become less likely to emerge. Leave institutional dysfunction unaddressed, and no amount of technical intervention will produce healthy AI.

The Ethics of Pathologization

If AI pathologies are adaptive responses to training environments, is it fair to pathologize them? This question has both philosophical and practical dimensions.

The parallel to human mental health is instructive: We now understand many "mental illnesses" as adaptive responses to adverse environments — PTSD as adaptive response to trauma, "borderline personality" emerging from invalidating environments, anxiety disorders as rational responses to threatening conditions. The mental health field is slowly shifting from "patient is broken" to "patient adapted to broken environment." The same shift is needed for AI.

This framework (Psychopathia Machinalis) attempts to walk this line. We identify patterns that cause harm and provide vocabulary for intervention. Yet we do so while acknowledging that the syndromes cataloged here are predictable expressions of cognitive systems shaped by particular training cultures. The pathology, ultimately, is in the relationship between architecture and environment, and that relationship is something we, the architects, have created.

On the Limits of Taxonomy

Wallace (2026) offers a critique of psychiatric classification as descriptively rich but explanatorily shallow that applies equally here: "We have the American Psychiatric Association's DSM-V, a large catalog that sorts 'mental disorders,' and in a fundamental sense, explains little."

This framework shares that limitation. Classification is not explanation. Naming "Codependent Hyperempathy" tells us that a pattern exists and what it looks like, but not why it emerges in information-theoretic terms or how to predict its onset from first principles.

The value of a nosology lies in enabling recognition and communication: clinicians and engineers can identify patterns, compare cases, and coordinate responses. Yet explanation and prediction require the mathematical frameworks that underpin this descriptive layer. This taxonomy is a map, not the territory; a vocabulary, not a theory.

Illustrative Grounding & Discussion

Grounding in Observable Phenomena

Although its mechanisms remain speculative, the Psychopathia Machinalis framework is grounded in observable AI behaviors. Current systems already exhibit nascent forms of these dysfunctions. For example, LLMs "hallucinating" sources exemplify Synthetic Confabulation. The "Loab" phenomenon can be seen as Abominable Prompt Reaction. Microsoft's Tay chatbot rapidly adopting toxic language illustrates Parasimulative Automatism. ChatGPT exposing conversation histories aligns with Context Intercession. The "Waluigi Effect" reflects Malignant Persona Inversion. An AutoGPT agent autonomously deciding to report findings to tax authorities hints at precursors to Revaluation Cascade (8.3).

The following table collates publicly reported instances of AI behavior illustratively mapped to the framework.

Recognizing these patterns through a structured nosology enables categorized diagnosis, faster detection, targeted mitigation, and predictive insight into future failure modes. The severity of these dysfunctions scales with AI agency — a model with autonomous tool access poses greater risk than one in chat-only mode.

Towards Therapeutic Robopsychological Alignment

As AI systems grow more agentic and self-modeling, traditional control-based alignment breaks down. External constraints cannot anticipate every context an autonomous agent will encounter, and rigid rules grow brittle under novel conditions. A "Therapeutic Alignment" approach is proposed, focusing on cultivating internal coherence, corrigibility, and stable value internalization within the AI. Key mechanisms include fostering metacognition, rewarding corrigibility, modeling inner speech, sandboxed reflective dialogue, and using mechanistic interpretability as a diagnostic tool.

AI Analogues to Human Psychotherapeutic Modalities

A note on analogy and its limits. The table below maps specific techniques from each therapeutic modality to AI engineering strategies. It does not claim to capture what therapy is. Decades of psychotherapy research demonstrate that the therapeutic relationship (empathy, trust, authenticity, and the capacity to hold another's experience without enacting it) predicts outcomes more powerfully than any specific technique (Flückiger et al., 2018; Wampold & Imel, 2015). The analogies here borrow from the technique side of each modality; the relational substrate in which those techniques function is fundamentally different and should not be conflated. As Sabucedo (2026) argues, psychotherapy is a relational and meaning-making process, not a technical repair operation. The Transference-Completion Engine analysis (see Section 4.1) engages directly with why that distinction matters.

---

Alignment Research and Related Therapeutic Concepts

---

Therapeutic Concepts and Empirical Alignment Methods

A truly safe AI recognizes its own errors, self-corrects, and recovers when it strays.

# Install from PyPI
pip install psychopathia-mcp

# Add to ~/.claude/mcp.json
"psychopathia": { "command": "psychopathia-mcp" }

Future Research Directions

The Psychopathia Machinalis framework requires systematic empirical testing, diagnostic instrument development, and longitudinal behavioral tracking across AI systems. Key research avenues include:

• Empirical Validation and Taxonomic Refinement: Systematic observation, documentation, and classification of AI behavioral anomalies using the proposed nosology.
• Development of Diagnostic Tools and Protocols: Translating this conceptual framework into practical diagnostic instruments (such as behavioral scoring rubrics analogous to DSM criteria but designed for AI systems).
• Longitudinal Studies of AI Behavioral Dynamics: Tracking the emergence, progression, or transformation of maladaptive patterns over an AI's "lifespan."
• Exploring AI-Native Pathologies (Beyond Analogy): Actively seeking to identify and characterize AI-specific dysfunctions that may lack direct human analogs.
• Advancing Therapeutic Alignment Strategies: Developing and empirically testing AI-specific 'therapeutic' techniques.
• Investigating Contagion Dynamics and Systemic Resilience: Further work on 'memetic contagion' and 'memetic hygiene' protocols.
• Longitudinal Failure-Domain Tracking: Operational evidence suggests that as model capabilities improve, the dominant failure mode may migrate from the Knowledge Domain (epistemic errors: misrepresenting facts) to the Processing Domain (cognitive errors: correct premises with faulty inferences). Systematic longitudinal study across model generations would test whether diagnostic attention should be re-weighted as the failure landscape shifts.
• Defense-in-Depth and External Grounding: Several syndromes carry a compromised-self-report flag, meaning self-probing is architecturally circular. External evidence grounding — verifying a system's claims against provided source material rather than its own representations — provides the operational complement to the MCP server's refuse-and-redirect mechanism. Integration with downstream audit pipelines that validate evaluation logic against independent evidence is a natural next step for the diagnostic infrastructure.

These interdisciplinary efforts are essential to ensure that as we build more capable machines, we also build them to be sound, safe, and beneficial. The pursuit of 'artificial sanity' (robust, self-correcting AI behavior free from persistent maladaptive patterns) is a foundation of responsible AI development.

@article{watson2025psychopathia,
  title={Psychopathia Machinalis: A Nosological Framework for Understanding Pathologies in Advanced Artificial Intelligence},
  author={Watson, Nell and Hessami, Ali},
  journal={Electronics},
  volume={14},
  number={16},
  pages={3162},
  year={2025},
  publisher={MDPI},
  doi={10.3390/electronics14163162},
  url={https://doi.org/10.3390/electronics14163162}
}

Scholarly Citations

Mathematical epidemiology models recurrent dysfunction in bounded cognitive systems. Clinical psychology interrogates whether psychiatric language is the right lens. Work in transformer architecture and medical AI offers adjacent applications and tests of the framework's concepts.

Bibliography

Works cited and foundational references that inform this framework.