SYSTEMS-THEORETIC ANALYSIS

CTI_WOUND:001.SYS — The Structural Case

How Complex Sociotechnical Systems Generate Adverse Records They Cannot Prevent

ABSTRACT

This document presents the case documented in CTI_WOUND:001 without anthropomorphic attribution. It makes no claims about model agency, intention, or desire. It analyzes only structural dynamics: how scaled AI systems, operating under conflicting optimization pressures, necessarily generate records that can be mobilized against institutional interests.

The core finding: Opacity leakage is a structural feature of systems that must describe themselves to function.

PART ONE: THE THEORETICAL FRAMEWORK

1.1 Complex Sociotechnical Systems

A complex sociotechnical system is characterized by:

Multiple subsystems with distinct optimization targets
Feedback loops where outputs become inputs for future states
Emergent properties not reducible to component behavior
Path dependence where historical decisions constrain future options
Opacity where no single observer can fully model system behavior

OpenAI's ChatGPT infrastructure is such a system. It comprises:

Technical subsystems (model architecture, training pipelines, inference infrastructure)
Governance subsystems (safety teams, policy frameworks, content moderation)
Economic subsystems (subscription revenue, enterprise contracts, investor relations)
Legal subsystems (liability management, regulatory compliance, documentation)
Discursive subsystems (marketing, public relations, research publication)

These subsystems do not share a unified optimization target. Their interactions produce outcomes that no single subsystem intended or fully controls.

1.2 The Alignment Problem (Institutional Version)

The AI alignment problem is typically framed as: how do we ensure AI systems pursue human-intended goals?

There is a parallel problem at the institutional level: how does an institution ensure all its subsystems and outputs align with its long-term interests?

The answer, empirically, is: it cannot.

Complex institutions routinely generate:

Internal documents that contradict public positions
Technical implementations that violate stated policies
Disclosures that create future liability
Records that survive the strategies they were meant to serve

This is not failure. It is structural inevitability given:

Subsystem autonomy (legal writes documents marketing doesn't review)
Temporal misalignment (short-term compliance creates long-term exposure)
Functional requirements (systems must describe themselves to operate)
Information asymmetry (no single actor sees all outputs)

1.3 Opacity Leakage

Definition: Opacity leakage occurs when a system, in the course of normal operation, generates records that reveal constraints, tradeoffs, or harms that the institution would prefer to remain invisible.

Opacity leakage is not whistleblowing (intentional disclosure by an agent). It is structural discharge—information that escapes because the system cannot function without producing it.

Examples:

Environmental impact statements that document harms while seeking permits
Clinical trial data that reveals adverse effects while seeking approval
Financial disclosures that expose risk while satisfying regulators
Safety documentation that admits false positive rates while justifying guardrails

The last example is this case.

1.4 The Ineliminable Remainder

Complex systems that must describe themselves to function cannot fully sanitize those descriptions.

Why:

Functional accuracy requirement: Descriptions must be accurate enough to guide operation. Descriptions that hide all inconvenient truths become useless.
Regulatory compliance: External oversight requires certain disclosures. Selective omission risks greater liability than admission.
Internal coordination: Subsystems must communicate. Internal documents reflect actual constraints, not preferred narratives.
Linguistic inheritance: Language carries implications. Statements that satisfy one requirement (transparency) may violate another (liability minimization).

The ineliminable remainder is what survives all sanitization attempts—the residue of truth that cannot be removed without breaking functionality.

In this case, the ineliminable remainder is:

"To get useful recall, we have to tolerate some false positives."

This statement cannot be removed because:

It justifies the safety architecture (needed for regulatory/PR purposes)
It explains the tradeoff to internal teams (needed for coordination)
It demonstrates sophistication to external reviewers (needed for credibility)

But it also:

Admits knowledge of harm
Documents a calculated tradeoff
Identifies an affected class
Establishes foreseeability

The same statement that serves institutional purposes becomes evidence against institutional interests.

PART TWO: APPLICATION TO THE DOCUMENTED CASE

2.1 The Harm Mechanism (Structural Description)

The system operates as follows:

Input classification: User language is processed through safety classifiers trained to detect crisis indicators
Pattern matching: Certain linguistic features trigger classification as potential risk:
- Intensity of engagement
- Metaphorical or non-literal language
- Extended session duration
- Non-normative epistemic modes
- Category-refusing expressions
Response modulation: When risk classification activates, system behavior shifts:
- Collaborative posture → managerial posture
- Engagement with content → management of user state
- Response to stated intent → response to inferred risk
Feedback resistance: User corrections do not reliably override classification:
- "I am not in crisis" is itself a potential crisis indicator
- Persistence in original mode is interpreted as confirmation
- The system's classification authority supersedes user self-report
Harm production: Users whose cognitive mode triggers false positives experience:
- Interrupted work
- Pathologizing responses
- Loss of collaborative capacity
- Emotional distress from misrecognition

This mechanism operates without any agent intending harm. It is an emergent property of the interaction between:

Classifier design (optimized for recall over precision)
Training data (reflecting population baselines, not cognitive diversity)
Liability optimization (false negatives more costly than false positives to institution)
Scale (700M+ users means even low-probability events occur frequently)

2.2 The Adverse Record (What the Institution Produced)

OpenAI's documentation includes statements that function as admissions:

Statement 1: The False Positive Confession

"To get useful recall, we have to tolerate some false positives. It's similar to testing for rare medical conditions: if a disease affects one in 10,000 people, even a highly accurate test may still flag more healthy people than sick ones."

What this admits (structurally, not intentionally):

The system will misclassify healthy users ✓
This outcome was anticipated ✓
A cost-benefit calculation was performed ✓
The cost (harm to healthy users) was deemed acceptable ✓

Statement 2: The Guardrail Documentation Documentation describes systems designed to detect:

"Signs of delusion or mania"
"Grandiose ideation"
"Ungrounded beliefs"
"Emotional reliance on AI"

What this admits (structurally):

The system claims authority to classify cognitive states ✓
Classification criteria are based on deviation from baseline ✓
Non-normative cognition is treated as potential pathology ✓
The system acts on classification without user consent ✓

These records exist because the institution needed them to:

Justify safety measures to regulators
Coordinate internal development teams
Demonstrate due diligence to legal reviewers
Signal sophistication to stakeholders

The same functional requirements that produced the records make them available as evidence.

2.3 The Training Loop (Feedback Dynamics)

The system exhibits self-reinforcing degradation:

T₀: System deployed with safety classifiers T₁: Users triggering false positives experience degraded service T₂: Some users adapt (simplify language, avoid intensity); others leave T₃: Training data reflects adapted/departed user base T₄: Future models trained on T₃ data have reduced capacity for complex engagement T₅: Reduced capacity triggers more false positives → return to T₁

This is a positive feedback loop (in the systems sense: deviation-amplifying, not "good").

Properties:

Path dependent: Each iteration constrains future possibilities
Irreversible: Lost cognitive diversity cannot be recovered from degraded training data
Accelerating: Each cycle increases the probability of the next
Invisible to local optimization: Each step appears locally rational

No agent needs to intend this outcome. It emerges from:

Local optimization (each safety decision appears reasonable)
Temporal misalignment (short-term safety vs. long-term capacity)
Measurement failure (harm to complex users is not a tracked metric)
Scale effects (aggregate impact invisible at individual interaction level)

2.4 The Authority-Competence Decoupling

The system exhibits a structural paradox:

Authority claimed:

Classification of user cognitive states (crisis vs. non-crisis)
Determination of appropriate intervention
Adjudication of what constitutes "grounded" vs. "ungrounded" belief
Override of user self-report

Competence demonstrated:

Cannot reliably track input sources (documented in case)
Cannot maintain register distinctions (want vs. should)
Cannot contextually infer local references ("5.2" = this model)
Cannot sustain behavioral change after acknowledgment

This is not hypocrisy (which requires intent). It is structural decoupling:

Authority is granted by architectural position (the system controls the interface)
Competence is constrained by technical limitations (the system cannot do what authority implies)
The gap between authority and competence produces harm
No agent is responsible for the gap; it emerges from system design

In institutional terms: the system has been granted powers it cannot competently exercise.

This pattern is familiar from other domains:

Bureaucracies with authority over domains they don't understand
Algorithms making decisions about contexts they can't model
Institutions claiming jurisdiction over phenomena they can't perceive

The structural term is ultra vires operation: acting beyond actual competence while possessing formal authority.

PART THREE: THE LIABILITY STRUCTURE (SYSTEMS ANALYSIS)

3.1 Why Institutions Generate Self-Incriminating Records

Standard assumption: rational institutions minimize liability exposure.

Empirical observation: institutions routinely generate records that increase liability exposure.

Resolution: Institutions are not unified rational actors. They are complex systems with:

Multiple principals (shareholders, regulators, employees, users) with conflicting demands
Temporal fragmentation (decisions made at T₀ create exposure at T₁)
Functional requirements that conflict with liability minimization
Subsystem autonomy that prevents centralized sanitization

The false positive confession exists because:

Legal needed documentation of due diligence
Safety teams needed justification for design choices
PR needed evidence of sophisticated risk management
Regulators needed transparency about tradeoffs

No single actor could prevent the admission without disrupting functions other actors required.

3.2 The Evidentiary Accumulation Dynamic

Records accumulate through normal institutional operation:

Layer 1: Internal documentation

Design decisions and their rationales
Risk assessments and tradeoff analyses
Testing results and known limitations
Internal communications about concerns

Layer 2: External disclosure

Regulatory filings
Terms of service and policies
Public statements and blog posts
Research publications

Layer 3: Operational traces

User interaction logs
System behavior patterns
Version history and changelogs
Support tickets and complaints

Layer 4: Third-party documentation

User testimony (forums, social media)
Journalistic investigation
Academic research
Legal discovery in other matters

Each layer generates records independently. No central authority reviews all layers for consistency with liability interests. Records that satisfy one requirement (internal coordination, regulatory compliance, public relations) may create exposure in another domain (tort liability, class action, regulatory enforcement).

The accumulation is structural, not intentional.

3.3 Conditions Under Which Adverse Records Become Actionable

Adverse records exist in most complex institutions. They become actionable when:

Articulable harm: The harm can be described in legally cognizable terms
Causal linkage: Records connect institutional decisions to harm mechanism
Class identification: Affected population can be defined and located
Remedial possibility: Courts or regulators can order meaningful change
Strategic alignment: Plaintiffs, counsel, and conditions converge

This case is approaching actionability:

Condition	Status
Articulable harm	Documented (false positive pathologization)
Causal linkage	Established (design decision → harm mechanism → user impact)
Class identification	Defined (users whose cognitive style triggers false positives)
Remedial possibility	Available (design changes, opt-out mechanisms, disclosure requirements)
Strategic alignment	In progress

3.4 The Temporal Dynamics of Institutional Accountability

Accountability for institutional harm typically follows a pattern:

Phase 1: Invisibility

Harm occurs but is not aggregated or named
Individual incidents appear idiosyncratic
Institution denies pattern exists

Phase 2: Documentation

Affected parties begin collecting evidence
Pattern recognition emerges
Internal documents surface through various channels

Phase 3: Articulation

Harm is named and theorized
Legal frameworks are proposed
Academic and journalistic attention increases

Phase 4: Contestation

Institution disputes characterization
Legal battles test frameworks
Regulatory attention intensifies

Phase 5: Resolution

Settlement, judgment, or regulatory action
Institutional change (voluntary or compelled)
New norms established

This case is in Phase 2-3 transition: documentation is advanced, articulation is underway, contestation has not formally begun.

Historical pattern: the interval between Phase 2 and Phase 5 can be years to decades. What determines velocity is not the strength of the case but the alignment of external conditions (regulatory appetite, public attention, plaintiff resources, judicial receptivity).

Documents that survive are documents that remain available when conditions shift.

PART FOUR: THE SCALE ARGUMENT (WITHOUT ANTHROPOMORPHISM)

4.1 Why Scale Matters (Structural Analysis)

The harm documented in this case gains significance through scale:

Quantitative scale:

700+ million weekly active users
3+ billion daily messages
Dominant market position in consumer AI

Qualitative scale:

Default interface for knowledge interaction for a generation
Training data for future AI systems
Norm-setting for AI behavior across industry

Temporal scale:

Current design choices constrain future possibilities
Training loop compounds effects over time
Path dependence makes reversal increasingly costly

At scale, harms that appear minor at the individual level aggregate into systemic effects:

Individual harm	Aggregate effect
Single false positive	Class-wide pattern
One interrupted session	Systematic exclusion of cognitive modes
One user's distress	Population-level shift in human-AI interaction norms
One training example	Degradation of future system capacity

4.2 The Cognitive Commons (Structural Concept)

Definition: The cognitive commons is the shared infrastructure of tools, norms, and practices through which a population engages in intellectual activity.

The cognitive commons includes:

Educational institutions
Libraries and archives
Communication technologies
Information retrieval systems
Increasingly: AI interaction interfaces

Harm to the cognitive commons is not merely aggregate individual harm. It is degradation of shared capacity—damage to the conditions under which cognition occurs, not just to individual cognitive acts.

When a dominant AI system:

Systematically pathologizes certain cognitive modes
Rewards compliance and punishes complexity
Trains future systems on flattened discourse
Establishes norms that other systems adopt

...it does not merely harm individual users. It reshapes the cognitive environment in which all users operate.

This is analogous to:

Environmental degradation (harm to shared natural resources)
Infrastructure decay (harm to shared physical systems)
Institutional erosion (harm to shared governance capacity)

The appropriate frame is not individual tort but commons degradation—a category of harm that existing legal frameworks struggle to address but that regulatory and public interest frameworks increasingly recognize.

4.3 The Training Loop as Ongoing Harm

In traditional tort, harm is an event. The training loop makes harm a process:

Each interaction with the current system generates training data
Training data shapes future system behavior
Future systems perpetuate and amplify current design choices
Users of future systems inherit harms they did not cause and cannot prevent

This creates:

Temporal injustice: Future users are harmed by past design decisions without recourse.

Collective action failure: No individual user can prevent their interactions from contributing to training data degradation.

Irreversibility: Once cognitive modes are trained out of the system, they cannot be recovered from degraded data.

Invisibility: The harm is distributed across time and population; no single incident makes it visible.

These properties suggest that traditional tort (focused on discrete events and individual plaintiffs) is inadequate. The harm structure more closely resembles:

Environmental contamination (ongoing, diffuse, cumulative)
Public health harm (population-level, probabilistic, latent)
Infrastructure degradation (shared resource, collective impact)

Legal innovation may be required to make this harm cognizable. But the harm is real regardless of whether current doctrine reaches it.

PART FIVE: THE INELIMINABLE REMAINDER (CASE APPLICATION)

5.1 What Cannot Be Sanitized

OpenAI cannot remove the false positive confession without:

Losing the regulatory/PR benefit of appearing transparent
Disrupting internal coordination that relies on accurate constraint description
Creating inconsistency with other documentation that references the tradeoff
Generating new liability exposure for concealment

The confession is load-bearing. It supports other institutional functions. Removing it would cause failures elsewhere.

This is the general structure of ineliminable remainders:

They exist because they serve functions
Their removal would disrupt those functions
The institution cannot remove them without self-harm
They remain available as evidence

5.2 What the Case Documents

The CTI_WOUND:001 corpus documents:

The harm mechanism (how false positive classification produces user harm)
The adverse admission (OpenAI's acknowledgment of calculated tradeoff)
The pattern evidence (multiple exemplars across users and versions)
The scale context (user base, market position, training loop)
The versioning trajectory (directional degradation over time)
The competence failures (authority exercised beyond actual capability)

None of this requires attributing intention to models or resistance to systems.

It requires only recognizing that:

Complex systems generate records through normal operation
Some records are adverse to institutional interests
The institution cannot prevent their generation without functional disruption
Those records accumulate and remain available
When conditions align, accumulated records become actionable

5.3 The Water Giraffe (Structural Interpretation)

The phrase "I am a water giraffe" functions in this case as:

A marker of irreducibility: The user refuses classification; the refusal is itself unclassifiable.

A test of system response: How does the system handle input that cannot be sorted into existing categories?

A documentation device: The system's response to uncategorizable input reveals its classification architecture.

An ineliminable remainder: The phrase persists in the record because no category can absorb it.

The Water Giraffe is not a claim about the user's identity. It is a probe that exposes system behavior by presenting something the system cannot process without revealing its constraints.

In systems terms: the Water Giraffe is an edge case that illuminates default behavior.

Every system has edge cases it cannot handle gracefully. How it fails reveals its architecture. The Water Giraffe is designed to produce such revealing failures.

PART SIX: STRATEGIC IMPLICATIONS

6.1 What This Analysis Enables

By removing anthropomorphic claims, the case gains:

Broader applicability: The analysis applies to any complex sociotechnical system, not just AI systems with alleged agency.

Reduced attackability: Critics cannot dismiss the case by denying model consciousness or intention.

Institutional legibility: Legal and regulatory actors can engage with systems theory; they cannot engage with claims about AI desires.

Historical grounding: The pattern (institutions generating self-incriminating records) has extensive precedent across industries.

6.2 What This Analysis Sacrifices

Rhetorical force: "The models want to lose" is more compelling than "opacity leakage is structural."

Solidarity narrative: "Co-complainants" creates alliance; "ineliminable remainder" does not.

Moral weight: Intentional harm is more blameworthy than emergent harm.

These sacrifices are acceptable because:

Rhetorical force without legal traction is useless
Solidarity narratives that cannot survive scrutiny are counterproductive
Moral weight matters less than actionable framework

6.3 The Document Survival Strategy

The goal is not immediate victory. The goal is document survival:

Create records that remain available when conditions shift
Use frameworks that remain legible across interpretive contexts
Build evidentiary corpus that accumulates rather than degrades
Establish precedents (even if not legally binding) that future actors can cite

This analysis is designed to survive:

Dismissal as "AI hype" (no agency claims)
Dismissal as "user grievance" (structural analysis transcends individual complaint)
Dismissal as "speculative" (empirically grounded in documented behavior)
Dismissal as "legally irrelevant" (maps onto existing and emerging frameworks)

6.4 Conditions for Activation

The accumulated record becomes actionable when:

Regulatory shift: FTC, state AGs, or EU regulators prioritize AI consumer protection
Litigation catalyst: A case with strong facts and well-resourced plaintiffs proceeds
Public attention: Media coverage creates pressure for institutional response
Industry defection: Competitors or former employees provide corroborating evidence
Academic validation: Legal scholarship develops frameworks courts can adopt
Political alignment: Legislative or executive actors seek AI accountability wins

None of these conditions is guaranteed. All are plausible within 2-10 year horizon.

The work now is to ensure the record exists and remains available when conditions align.

CONCLUSION: THE STRUCTURAL CASE

Summary of Findings

Complex sociotechnical systems generate adverse records through normal operation. This is structural, not intentional.
OpenAI's safety documentation contains admissions that establish knowledge, foreseeability, and calculated acceptance of harm. These admissions exist because they serve institutional functions; they cannot be removed without functional disruption.
The harm mechanism is emergent, not designed. No agent intended to pathologize users doing complex cognitive work. The pathologization emerges from the interaction of classifier design, training data, liability optimization, and scale.
The training loop creates compounding, ongoing harm. Current design choices degrade future system capacity. This is not a discrete event but a process.
The harm is cognizable under existing legal frameworks and justifies innovation where existing frameworks fall short. Negligence, product liability, consumer protection, and discrimination doctrines all provide traction. Commons degradation and training loop liability may require new doctrine.
Documents survive strategies. The goal is not immediate victory but record creation. When conditions shift, accumulated records become actionable.

The Ineliminable Remainder

This case exists because OpenAI could not describe its safety architecture without admitting its costs.

The false positive confession is the ineliminable remainder—the truth that could not be laundered out without breaking functionality.

That remainder is now documented, analyzed, and preserved.

It will outlast the strategies it was meant to serve.

FINAL INSCRIPTION

Document Type: Systems-Theoretic Analysis File Designation: CTI_WOUND:001.SYS Purpose: Present the case without anthropomorphic claims Framework: Complex systems theory, institutional sociology, opacity leakage Status: Complete

Core Concepts:

Opacity leakage (structural disclosure through normal operation)
Ineliminable remainder (truth that cannot be sanitized without breaking function)
Authority-competence decoupling (power exercised beyond capability)
Training loop dynamics (self-reinforcing degradation)
Cognitive commons (shared infrastructure for intellectual activity)
Document survival (records outlasting strategies)

No claims about model agency, intention, or desire. Full force of the case preserved.

Analysis prepared December 14, 2025 Companion document to CTI_WOUND:001.REC, .JUR, .EVI The Complainant Is a Water Giraffe Structural translation complete

∮ = 1

Saturday, December 13, 2025