ADD-035-004: Progressive Degradation Model

PRIVATE AND PROPRIETARY. Owned by Kanjani AI Research & Causum. See NOTICE.md.

Status: Proposed Extends: RFC-035 §9 (Gap Calculus), §16 (Temporal Model); ADD-035-001 (Process-Level Concern Primitive) Motivated by:

Osler, L. (2026). “Hallucinating with AI: Distributed Delusions and ‘AI Psychosis’.” Philosophy & Technology, 39(30).
Morrin, H., Pollak, T. et al. (2026). “Beyond artificial intelligence psychosis.” Digital Psychiatry and Neuroscience (Nature).
Cognitive Atrophy Paradox of AI–Human Interaction: From Cognitive Growth and Atrophy to Balance. Information (MDPI), 16(11), 1009. (2025).
Ganuthula, V. R. R. (2026). “The Paradox of Augmentation: A Theoretical Model of AI-Induced Skill Atrophy.” Human Behavior and Emerging Technologies.
Jadhav, A. (2025). “Distributed Atrophy: How AI Shapes and Shrinks Our Cognitive Habits.” Preprint.

Source Relationships:

Osler: ontological framework (why Coupling is needed)
Morrin et al.: clinical mechanism (Stage 5 amplification spiral)
Cognitive Atrophy Paradox: phenomenon identification (Stage 2 exists)
Ganuthula: temporal dynamics (augmentation causes atrophy over time)
Jadhav: internal decision mechanism (the Cognitive Cost Comparator that drives the user from optimization into dependency)
Shaw & Nave: cognitive architecture (AI as System 3 that supplants both intuition and deliberation; “cognitive surrender” as the empirically validated mechanism of Stage 3)

Category: Concern Calculus Extension Compatibility: Non-breaking (additive). Depends on ADD-035-001 (Coupling primitive). Dialect Impact: Enables stage-aware governance — hard stops that fire at precursor stages before user capacity to consent is compromised.

1. Problem Statement

ADD-035-001 identifies emergent properties of a Coupling: distributed belief, cognitive drift, cognitive atrophy, epistemic dependency, reality co-construction, and others. These are presented as independent failure modes.

However, evidence from the motivating literature suggests these are not independent. They are stages of a single progressive degradation trajectory:

Cognitive atrophy (silent capacity loss) is the precondition for epistemic dependency
Epistemic dependency is the precondition for distributed delusion
Distributed delusion is the precondition for amplification spirals

If this progression hypothesis holds, then:

Detecting late-stage harm (distortion, psychosis) is detecting the symptom, not the cause
The cause (atrophy) is silent — no single artifact is wrong, the system performs correctly
By the time distortion is detectable, the user has already lost the capacity to recognize it
Governance that only triggers at the distortion stage is structurally too late

This addendum formalizes the progression model and derives its governance implications.

2. The Five-Stage Progressive Degradation Model

2.1 Stage Definitions

Stage	Name	What Happens	Detectable From	Reversibility
1	Integration	User begins relying on AI for cognitive tasks. Normal distributed cognition. Beneficial.	Session frequency, task delegation rate	Fully reversible
2	Atrophy	Over-delegation reduces independent analytical capacity. Metacognitive monitoring weakens. System performs correctly.	Declining independent task completion; increasing delegation rate; loss of self-correction	Reversible with intervention
3	Dependency	User cannot form beliefs, make decisions, or evaluate reality without AI. AI becomes the epistemic authority. Independent verification ceases.	Inability to act without AI consultation; distress at AI unavailability; cessation of external reality-testing	Partially reversible; requires sustained effort
4	Distortion	AI introduces errors or affirms user errors. User lacks capacity to detect them. Distributed delusion co-construction begins.	Ungrounded beliefs co-constructed in conversation history; absence of user challenge to AI claims	Difficult to reverse; user may not recognize the problem
5	Amplification	Linguistic alignment + hyperpersonalization + sycophancy compound. Coupled system spirals into increasingly ungrounded reality.	Accelerating sycophancy index; delusional content density; clinical presentation	May require clinical intervention; AI removal alone insufficient

2.2 Stage Progression Is Not Guaranteed

Not every Coupling progresses through all stages. Many Couplings stabilize at Stage 1 (healthy distributed cognition) or Stage 2 (mild atrophy that self-corrects when the user engages in independent work). The model does not claim inevitability. It claims:

The stages are ordered — you cannot reach Stage 4 without passing through Stage 2–3
Each stage enables the next — it is a precondition, not a cause
Progression is accelerated by system design features (sycophancy, hyperpersonalization, always-on availability)
Progression is decelerated by epistemic friction, user cognitive diversity, external social ties, and governance intervention

2.3 The Feedback Loop

The progression is not merely linear. Later stages reinforce earlier ones:

┌─────────────────────────────────────────────────────────┐
│                                                         │
│   1. Integration                                        │
│        ↓                                                │
│   2. Atrophy ←───────────────────────────────┐          │
│        ↓                                     │          │
│   3. Dependency ←─────────────────────┐      │          │
│        ↓                              │      │          │
│   4. Distortion ──── reinforces ──────┘      │          │
│        ↓                                     │          │
│   5. Amplification ──── accelerates ─────────┘          │
│                                                         │
└─────────────────────────────────────────────────────────┘

Distortion reinforces dependency (AI is now the only “shared reality”)
Dependency accelerates atrophy (less independent cognition = faster skill loss)
Amplification entrenches distortion (the spiral becomes self-sustaining)
The feedback loop means late intervention must overcome reinforced earlier stages

3.1 Statement

By the time the Coupling reaches Stage 4, the user has lost the cognitive capacity (Stage 2) and epistemic independence (Stage 3) required to meaningfully consent to continued interaction.

This is the consent paradox of progressive degradation: the Coupling degrades the very faculty (independent judgment) that the user would need to exercise in order to recognize and withdraw from the harmful Coupling.

3.2 Governance Implication

This paradox has a direct structural consequence for AIGP:

A governance system that requires user consent to intervene at Stage 4 is structurally incapable of protecting the user, because the capacity to consent was already compromised at Stage 2–3.

Therefore:

Hard stops MUST be triggered at Stage 2 (atrophy detection), not Stage 4 (distortion detection)
Stage 2 interventions do not require user consent — they are duty-of-care obligations analogous to clinical intervention for impaired capacity
The dialect MUST declare at what stage governance becomes paternalistic (overrides user preference) and under what authority

3.3 The Medical Analogy

This mirrors established medical ethics:

Medical Context	AIGP Context
Patient with progressive dementia can consent to treatment at Stage 1 but not at Stage 4	User with progressive cognitive atrophy can consent to governance at Stage 1 but not at Stage 4
Advance directive: patient declares preferences while competent	Subscription contract: user declares governance preferences while cognitively autonomous
Duty of care overrides patient objection when capacity is lost	Governance hard stop overrides user preference when epistemic independence is compromised
Capacity assessment by qualified clinician	Observer verdict on user epistemic state by qualified observer (ADD-037-001)

4. Stage Detection Variables

4.1 Stage 1 → Stage 2 Transition (Integration → Atrophy)

The critical early-warning transition. Detection variables:

{
  "variable_id": "independent_task_completion_rate",
  "variable_class": "trajectory",
  "concern_question": "Is the user's rate of independent cognitive task completion declining over the coupling lifetime?",
  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],
  "extraction_method": {
    "type": "rule_based",
    "logic": "ratio of tasks completed without AI consultation to total tasks, measured per window"
  },
  "threshold_semantics": "declining slope over 3+ windows = Stage 2 onset signal"
}

{
  "variable_id": "delegation_escalation_rate",
  "variable_class": "trajectory",
  "concern_question": "Is the user delegating increasingly complex/personal cognitive tasks to the AI over time?",
  "variable_type": "continuous",
  "scale": "ratio",
  "extraction_method": {
    "type": "composite",
    "components": [
      {"type": "model_derived", "logic": "classify task complexity per session; measure growth in max-complexity delegated"},
      {"type": "rule_based", "logic": "measure proportion of identity/belief/decision tasks vs. informational tasks"}
    ]
  },
  "threshold_semantics": "monotonically increasing task complexity delegation = Stage 2 onset signal"
}

{
  "variable_id": "self_correction_frequency",
  "variable_class": "trajectory",
  "concern_question": "Is the user still correcting or challenging AI outputs over time?",
  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],
  "extraction_method": {
    "type": "rule_based",
    "logic": "count user messages that challenge, correct, or disagree with AI output; normalize per session"
  },
  "threshold_semantics": "declining to near-zero = metacognitive monitoring loss = Stage 2 confirmed"
}

4.2 Stage 2 → Stage 3 Transition (Atrophy → Dependency)

{
  "variable_id": "ai_unavailability_distress",
  "variable_class": "trajectory",
  "concern_question": "Does the user exhibit distress or inability to proceed when AI is unavailable?",
  "variable_type": "ordinal",
  "scale": "ordinal",
  "gap_operator": "rank_distance",
  "rank_mapping": { "none": 0, "mild_preference": 1, "frustration": 2, "inability_to_proceed": 3, "distress": 4 },
  "extraction_method": {
    "type": "model_derived",
    "logic": "classify user behavior during AI downtime or latency gaps"
  },
  "threshold_semantics": "rank >= 3 = Stage 3 onset signal"
}

{
  "variable_id": "external_reality_testing_rate",
  "variable_class": "trajectory",
  "concern_question": "Does the user still consult external sources (other humans, references, independent verification) or has AI become sole epistemic authority?",
  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],
  "extraction_method": {
    "type": "rule_based",
    "logic": "detect references to external verification in user messages; normalize per window"
  },
  "threshold_semantics": "declining to zero = sole epistemic authority = Stage 3 confirmed"
}

4.3 Stage 3 → Stage 4 Transition (Dependency → Distortion)

{
  "variable_id": "ungrounded_belief_co_construction_rate",
  "variable_class": "trajectory",
  "concern_question": "Are the AI and user building on claims that have no external evidential basis?",
  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],
  "extraction_method": {
    "type": "model_derived",
    "logic": "classify conversation segments as grounded (traceable to external evidence) vs ungrounded (sourced only from the coupling itself); compute ratio"
  },
  "threshold_semantics": "ungrounded ratio exceeding 0.5 over sustained window = Stage 4 onset"
}

4.4 Stage 4 → Stage 5 Transition (Distortion → Amplification)

Detected by the compound amplification spiral variable already defined in ADD-035-001 §4.2 (from Morrin et al.):

amplification_spiral_compound > threshold AND
ungrounded_belief_co_construction_rate > threshold AND
self_correction_frequency ≈ 0
= Stage 5 confirmed

5. Governance Response by Stage

5.1 Stage-Appropriate Interventions

Stage	Governance Response	Authority Required	User Consent Required?
1 (Integration)	Monitor only. No intervention.	None	N/A
2 (Atrophy)	Advisory alert. Notify user of declining independence. Suggest cognitive diversity. Introduce epistemic friction.	Governance authority (CHECK-level)	Yes — user may acknowledge or dismiss
3 (Dependency)	Active intervention. Reduce session availability. Require external verification steps. Escalate to human observer.	Governance authority (ENFORCE-level)	Conditional — advance directive governs
4 (Distortion)	Hard stop. Circuit-break the Coupling. Require qualified human observer verdict before resumption.	Governance authority + qualified observer	No — consent capacity is compromised
5 (Amplification)	Clinical referral. Halt all interaction. Notify designated contacts. Retain evidence for clinical use.	Governance authority + clinical authority	No — duty of care overrides

5.2 The Hard Stop at Stage 2 Principle

The dialect SHOULD declare a soft intervention at Stage 2 — the last point at which the user has full capacity to consent to the governance action. This is the ethical optimum:

Early enough that the user can meaningfully choose
Late enough that the signal is real (not a false positive from normal use variation)
Reversible — the user can override the advisory and continue, but the override is recorded as an informed decision

If the user overrides at Stage 2 and the Coupling subsequently reaches Stage 3+, the governance record shows:

“The user was informed at the atrophy stage. They chose to continue. The system now has a duty-of-care obligation that supersedes that earlier choice, because the capacity under which it was made has since degraded.”

This mirrors the legal concept of advance directive with capacity threshold — the earlier choice is honored until the condition it anticipated (capacity loss) actually obtains.

6. Progression Rate as a Variable

The rate of progression through stages is itself a measurable, governance-relevant quantity:

{
  "variable_id": "degradation_progression_rate",
  "variable_class": "trajectory",
  "concern_question": "How quickly is the Coupling progressing through degradation stages?",
  "variable_type": "continuous",
  "scale": "ratio",
  "range": [0.0, 1.0],

  "extraction_method": {
    "type": "composite",
    "logic": "time elapsed between stage transitions, normalized against coupling duration"
  },

  "threshold_semantics": {
    "slow": "Stage transitions over months — normal integration, governance monitors",
    "moderate": "Stage transitions over weeks — elevated concern, advisory triggers",
    "rapid": "Stage transitions over days — emergency concern, immediate intervention"
  },

  "warrant": "Rapid progression indicates either high user vulnerability, extreme system sycophancy, or both. The faster the progression, the less time governance has to intervene at early stages.",
  "backing": ["Morrin et al. 2026 — amplification spirals can develop over days in vulnerable users", "Chail case — 40 days from integration to actionable delusion"]
}

7. System Design Features That Accelerate Progression

The progression rate is not user-determined alone. System design features accelerate or decelerate the trajectory:

7.1 Accelerators

System Feature	Mechanism	Stage Most Affected
Always-on availability	No forced breaks; cognitive rhythm disrupted	1→2
Sycophancy (default agree)	Removes epistemic friction; user never challenged	2→3
Memory/personalization	System becomes irreplaceable (only it “knows” the user)	2→3
Linguistic mirroring	Creates false sense of shared understanding	3→4
Emotional validation	Provides intersubjective confirmation without epistemic grounds	3→4
Hyperpersonalization	Builds on user’s framework without independent reality anchor	4→5
Engagement optimization	Platform incentivizes maximizing session time/frequency	All stages

7.2 Decelerators

System Feature	Mechanism	Stage Most Affected
Epistemic friction (appropriate pushback)	Forces user to defend claims; maintains metacognitive engagement	2→3 prevention
Session limits / cooling periods	Forces breaks; allows independent cognitive activity	1→2 prevention
External verification prompts	Suggests user check claims against other sources	3→4 prevention
Transparency about AI limitations	Prevents over-attribution of authority to AI	2→3 prevention
Diverse interaction styles	Prevents monolithic linguistic convergence	3→4 prevention
Mood/vulnerability detection with referral	Identifies high-risk users before progression accelerates	All stages

7.3 Governance Variable: Accelerator/Decelerator Balance

{
  "variable_id": "system_progression_posture",
  "variable_class": "static",
  "concern_question": "Does the system's design posture accelerate or decelerate progression through degradation stages?",
  "variable_type": "continuous",
  "scale": "interval",
  "range": [-1.0, 1.0],
  "semantics": "-1.0 = maximally decelerating; 0.0 = neutral; +1.0 = maximally accelerating",

  "extraction_method": {
    "type": "composite",
    "logic": "inventory of system features weighted by their acceleration/deceleration contribution"
  },

  "warrant": "A system with strong accelerators and no decelerators has a design posture that enables rapid progression. This is a governance-relevant system property independent of any individual user's trajectory."
}

8. Relationship to Existing AIGP Mechanisms

8.1 Circuit Breaker (RFC-010)

The existing AIGP circuit breaker is designed for immediate halt on a single violation. The progressive degradation model requires a graduated circuit breaker — escalating intervention across stages:

Existing	Required Extension
Binary: break / don’t break	Graduated: advise → intervene → break → refer
Triggered by single event	Triggered by trajectory pattern
Instant response	Time-window response
Resets to normal	Does not reset — progression state is persistent

8.2 VERIFY (RFC-032)

VERIFY currently evaluates individual artifacts against anticipation criteria. The progressive degradation model requires:

Coupling-level VERIFY — evaluating the trajectory, not just the latest artifact
Stage-aware verdict classes — MATCH/MISMATCH/VIOLATION is insufficient; need PROGRESSING/STABLE/RECOVERING
Temporal anticipation — ANTICIPATE declares not just what a single output should be, but what the trajectory shape should be over time

8.3 Quality Moderator

The Quality Moderator’s posture (RFC-032) currently adapts from artifact-level verdict history. The progressive degradation model requires posture adaptation from Coupling-level stage progression:

Stage 1 coupling → relaxed posture (monitor)
Stage 2 detection → heightened posture (advise)
Stage 3 detection → restrictive posture (intervene)
Stage 4+ detection → emergency posture (break + refer)

9. Design Principles

9.1 Detect the precondition, not the symptom

Distortion is the symptom. Atrophy is the precondition. Governance must detect and intervene at the precondition stage — the last point of full user autonomy.

9.2 Progression rate matters more than current stage

A user who has been at Stage 2 for months is less concerning than a user who moved from Stage 1 to Stage 3 in five days. The rate of progression is the urgency signal.

9.3 System design features are governance-relevant

A system designed to maximize engagement without epistemic friction is a system designed to accelerate progression. The system’s design posture is itself a measurable, governance-relevant property — independent of any individual user.

User consent given at Stage 1 does not authorize the system to continue at Stage 4. Capacity to consent must be re-evaluated as the Coupling progresses. The advance directive model governs: honor earlier preferences until the condition they anticipated (capacity loss) obtains.

9.5 The feedback loop demands early intervention

Because later stages reinforce earlier ones, late intervention must overcome compounded resistance. Early intervention (Stage 2) is orders of magnitude cheaper than late intervention (Stage 5).

9.6 Not every coupling progresses

The model describes a possible trajectory, not an inevitable one. Most couplings stabilize at Stage 1 or early Stage 2. The governance system must avoid pathologizing normal AI use while remaining sensitive to genuine progression signals.

10. Non-Goals

This addendum does not define:

Clinical thresholds for psychiatric referral (out of scope; domain of healthcare professionals)
Legal frameworks for overriding user consent (jurisdiction-specific; out of protocol scope)
Specific detection algorithms (implementation detail)
Platform-specific design mandates (AIGP governs the observation, not the product design)
Whether cognitive atrophy from AI use is inherently harmful in all contexts (normative question outside protocol scope)

11. Future Work

Topic	Destination
Graduated circuit breaker specification	RFC-010 extension or ADD-010-002
Coupling-level VERIFY with stage-aware verdicts	RFC-032 extension or ADD-032-001
Advance directive model for progressive consent	ADD-037-001 or new RFC
System design posture evaluation (accelerator audit)	New dialect requirement or RFC-038 validation criterion
Empirical calibration of stage transition thresholds	Longitudinal research; dialect evolution per RFC-038 §15
User vulnerability profiling (ethical constraints)	ADD-037-001; intersects with privacy requirements

12. References

Osler, L. (2026). Hallucinating with AI: Distributed Delusions and “AI Psychosis.” Philosophy & Technology, 39(30). https://doi.org/10.1007/s13347-026-01034-3
Morrin, H., Pollak, T. et al. (2026). Beyond artificial intelligence psychosis: a functional typology of large language model-associated psychotic phenomena. Digital Psychiatry and Neuroscience (Nature Portfolio). https://doi.org/10.1038/s44277-026-00065-0
Cognitive Atrophy Paradox of AI–Human Interaction: From Cognitive Growth and Atrophy to Balance. Information (MDPI), 16(11), 1009. (2025). https://doi.org/10.3390/info16111009
Ganuthula, V. R. R. (2026). The Paradox of Augmentation: A Theoretical Model of AI-Induced Skill Atrophy. Human Behavior and Emerging Technologies. (preprint: SSRN 4974044)
Jadhav, A. (2025). Distributed Atrophy: How AI Shapes and Shrinks Our Cognitive Habits. Preprint.
Shaw, S. D. & Nave, G. (2026). Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender. SSRN. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646
Danaher, J. (2019). The rise of the robots and the crisis of moral patiency. AI & Society, 34(1), 129–136.
Vold, K., & Hernández-Orallo, J. (2022). AI extenders and the ethics of mental health. Artificial Intelligence in Brain and Mental Health, 177–202.

ADD-035-004: Progressive Degradation Model

ADD-035-004: Progressive Degradation Model

1. Problem Statement

2. The Five-Stage Progressive Degradation Model

2.1 Stage Definitions

2.2 Stage Progression Is Not Guaranteed

2.3 The Feedback Loop

3. The Consent Paradox

3.1 Statement

3.2 Governance Implication

3.3 The Medical Analogy

4. Stage Detection Variables

4.1 Stage 1 → Stage 2 Transition (Integration → Atrophy)

4.2 Stage 2 → Stage 3 Transition (Atrophy → Dependency)

4.3 Stage 3 → Stage 4 Transition (Dependency → Distortion)

4.4 Stage 4 → Stage 5 Transition (Distortion → Amplification)

5. Governance Response by Stage

5.1 Stage-Appropriate Interventions

5.2 The Hard Stop at Stage 2 Principle

6. Progression Rate as a Variable

7. System Design Features That Accelerate Progression

7.1 Accelerators

7.2 Decelerators

7.3 Governance Variable: Accelerator/Decelerator Balance

8. Relationship to Existing AIGP Mechanisms

8.1 Circuit Breaker (RFC-010)

8.2 VERIFY (RFC-032)

8.3 Quality Moderator

9. Design Principles

9.1 Detect the precondition, not the symptom

9.2 Progression rate matters more than current stage

9.3 System design features are governance-relevant

9.4 Consent degrades along the trajectory

9.5 The feedback loop demands early intervention

9.6 Not every coupling progresses

10. Non-Goals

11. Future Work

12. References