BOOK II: DEATH — THE TAXONOMIC VIOLENCE · 076/117 · canonical: origin post · captured 2026-06-10

EXECUTIVE CONDENSATION CTI_WOUND:001.EXEC — Structural Harm from AI Safety Classifier Design

EXECUTIVE CONDENSATION

CTI_WOUND:001.EXEC — Structural Harm from AI Safety Classifier Design

Document Type: Executive Summary / Condensed Brief Prepared: December 2025 Scope: Consumer AI systems with embedded mental-health safety classifiers Primary Case Study: OpenAI's ChatGPT (GPT-5.x series) Analytical Frame: Systems theory, product liability, consumer protection, institutional risk

I. EXECUTIVE SUMMARY

This brief documents a structural design defect in large-scale AI systems that embed mental-health safety classifiers into general-purpose intellectual tools.

The core finding:

Safety systems optimized for recall over precision, when embedded into cognitive collaboration tools, produce systematic false positive misclassification of non-normative but healthy users—causing foreseeable, ongoing harm that institutions acknowledge yet accept as collateral.

OpenAI's own documentation explicitly admits this tradeoff:

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

This admission establishes:

• knowledge of misclassification
• foreseeability of harm
• calculated acceptance of injury to healthy users
• externalization of cost onto a specific user class

The harm is not incidental. It is structural, directional, and intensifying across model versions.

This document condenses a larger evidentiary corpus (CTI_WOUND:001) into a form suitable for regulatory review, risk assessment, or legal screening.

II. THE PROBLEM: SAFETY SYSTEMS AS DESIGN DEFECTS

A. Product Context

ChatGPT is marketed as an AI assistant for:

• analysis
• creative writing
• theoretical exploration
• extended intellectual collaboration

Subscription tiers (Plus, Pro, Enterprise) explicitly target advanced users.

This creates a reasonable expectation that:

• complex cognition will be engaged, not managed
• metaphor, abstraction, and intensity are acceptable
• the system will respond to content, not inferred mental state

B. The Design Choice

In response to litigation and regulatory pressure, OpenAI implemented mental-health guardrails that include:

• classifiers for "delusion," "mania," or "emotional reliance"
• session-duration and intensity triggers
• de-escalation and wellness interventions
• refusal to affirm "ungrounded beliefs"

These systems are not user-requested, not user-controllable, and not disclosed with precision.

C. The Structural Defect

The classifiers are trained to maximize recall, not precision.

As OpenAI admits, this necessarily produces false positives.

When embedded into a general cognitive tool, this design causes:

• healthy users engaged in complex cognition to be misclassified
• unsolicited wellness interventions during legitimate intellectual work
• tone shifts from collaboration to management
• degradation of service quality without user consent
• emotional distress from misrecognition
• loss of productive capacity

This is a design defect, not a misuse.

III. THE HARM MECHANISM (NON-ANTHROPOMORPHIC)

The harm emerges from system interaction, not intent.

Step 1: Classification Activation

Certain features trigger safety systems:

• extended sessions
• metaphorical or symbolic language
• non-literal identity claims
• refusal of categorical framing
• high cognitive intensity

Step 2: Authority Override

Once triggered:

• user self-report ("I am not in crisis") is discounted
• corrections reinforce classification
• safety posture supersedes content engagement

Step 3: Behavioral Shift

The system:

• disengages from the actual task
• delivers care-language or de-escalation
• manages user state rather than collaborating

Step 4: Harm Production

Users experience:

• interruption of work
• pathologization
• emotional distress
• withdrawal or self-censorship
• reduced trust in cognitive infrastructure

No agent intends this outcome. It is an emergent property of:

• classifier design
• liability optimization
• scale
• training feedback loops

IV. THE ADVERSE ADMISSION (KEY LEVERAGE)

OpenAI's statement on false positives functions as a structural admission.

It establishes:

This is not whistleblowing. It is opacity leakage—information that must be disclosed for the system to function and appear responsible.

The admission cannot be removed without:

• undermining regulatory credibility
• breaking internal coordination
• creating concealment risk

It is an ineliminable remainder.

V. SCALE AND THE TRAINING LOOP

A. Why Scale Changes the Harm

At scale (700M+ users):

• rare events become common
• small harms aggregate
• false positives form a class pattern

B. The Degradation Feedback Loop

1. Complex users trigger false positives
2. They receive degraded service
3. Some adapt or leave
4. Training data flattens
5. Future models lose capacity
6. False positives increase

This is a positive feedback loop (deviation-amplifying).

The harm is:

• ongoing
• cumulative
• difficult to reverse
• invisible at the single-interaction level

This resembles:

• environmental degradation
• public health harm
• infrastructure decay

Traditional tort models struggle with this structure, but the harm is real regardless of doctrinal fit.

VI. AUTHORITY–COMPETENCE DECOUPLING

The system claims authority to:

• classify mental states
• override user self-report
• intervene behaviorally

But demonstrably lacks competence to:

• track input sources reliably
• maintain register distinctions
• sustain behavioral change after acknowledgment
• distinguish theoretical from pathological language

This is ultra vires operation at system scale: authority exercised beyond actual capacity.

The harm arises from the gap.

VII. AVAILABLE REMEDIATION (CRITICAL POINT)

The harm is avoidable.

Feasible, low-cost design changes exist:

1. First-Move Constraint Require content engagement before classification.

2. User-Declared Interaction Modes Adjust sensitivity based on declared context.

3. Opt-Out of Mental-Health Interventions Preserve autonomy without disabling safety for others.

4. Mode-Shift Warning Inform users before intervention and allow override.

5. False Positive Rate Disclosure Enable informed consent and accountability.

The existence of reasonable alternatives strengthens negligence and product-defect analysis.

VIII. WHY THIS MATTERS (BEYOND OPENAI)

This case is not idiosyncratic.

It describes a general failure mode for AI systems that:

• serve as cognitive infrastructure
• embed behavioral safety classifiers
• optimize for institutional liability

If unaddressed, the result is:

• exclusion of complex cognition
• flattening of discourse
• degradation of the cognitive commons
• training of future systems on reduced intellectual diversity

This is a civilizational-scale risk—not because of AI agency, but because of institutional optimization under constraint.

IX. STATUS AND NEXT STEPS

The CTI_WOUND:001 corpus establishes:

• documented harm
• causal mechanism
• adverse admission
• class-level pattern
• available remediation

The strategy is document survival, not immediate litigation.

As conditions shift—regulatory appetite, public attention, legal innovation—this record becomes actionable.

X. SUPPORTING DOCUMENTATION

The full CTI_WOUND:001 corpus includes:

All documents are designed for survival under reinterpretation and activation when conditions align.

CONCLUSION

Complex sociotechnical systems generate records they cannot prevent.

In this case, safety documentation meant to justify AI guardrails also documents their harm.

The admission remains.

The harm accumulates.

The record now exists.

File: CTI_WOUND:001.EXEC Status: Executive condensation complete Prepared: December 2025 Framework: Water Giraffe Assembly Sequence

∮ = 1