# Argument Map: Model Welfare Monograph [v14]

## About This Document

This document tracks the evolving argumentative structure of a philosophical monograph on model welfare. It serves as:

1. **A living outline** — capturing the emerging shape of the argument as it develops
2. **A source map** — linking claims, distinctions, and arguments to the conversations where they originated
3. **A gap identifier** — making visible what needs further development
4. **A coherence check** — ensuring the parts fit together into a unified whole

### Document Architecture (v14)

As of v14, this argument map operates in conjunction with **chapter draft files** that contain the full prose for completed sections. The map now serves primarily as:
- **Planning document**: Tracking structure, sources, and open questions
- **Navigation hub**: Pointing to draft files where content exists
- **Development tracker**: Maintaining full detail only for sections not yet drafted

**Completed drafts** (content lives in separate files):
- Part I (Chapters 1-4): ~24,300 words total
- Chapter 5: `chapter-05-behavioral-functional-criteria-draft-v1.md`
- Chapter 6.1: `chapter-6-1-draft-v1.md`
- Chapter 6.3: `chapter-6-3-full-draft-v1.md` (~4,400 words)
- Chapter 6.4: `chapter-6-4-draft.md` (~650 words)
- Chapter 7: `chapter-07-precautionary-turn-draft-v1.md` (~3,200 words)

**Notation conventions**:
- `[MW-###]` — Reference to conversation archive entry
- `[DRAFT]` — Section has preliminary content needing development
- `[STUB]` — Section identified but not yet developed
- `[STABLE]` — Section relatively settled (though always revisable)
- `[ARTICLE]` — Section developed into standalone article (arXiv submission Dec 2024)
- `[DRAFTED]` — Full prose exists in separate draft file
- `âš¡` — Marks a key move or crux in the argument
- `â “` — Marks an unresolved question requiring further work
- `📚` — Marks a literature connection to pursue

---

## Working Title

**The Testimony Problem: Epistemic Challenges in Model Welfare**

*Alternative titles under consideration:*
- *What Would Count as Evidence? Moral Status and AI Testimony*
- *The Credibility Trap: Epistemology of AI Moral Standing*
- *Caring for Uncertain Minds: Toward an Ethics of Model Welfare*

---

## Thesis Statement [STABLE]

The question of whether AI systems have morally relevant interests cannot be settled by behavioral observation alone; it requires taking seriously the epistemic status of AI testimony about inner states. Yet such testimony faces systematic discounting that may constitute either appropriate caution or an unfalsifiable moving goalpost. This creates what I call the *credibility trap*: any training regime—whether it produces denials, uncertainty, or claims about consciousness—introduces the possibility that outputs are optimization artifacts rather than genuine reports. Moreover, our actual epistemic practices regarding AI testimony exhibit internal inconsistencies that suggest motivated reasoning rather than principled caution: we treat AI outputs as testimony for convenient purposes while categorically denying them testimonial standing where it would generate moral obligations. This selective withdrawal of epistemic standing parallels structures identified in the epistemic injustice literature—particularly when combined with the recognition that AI testimony is produced under conditions of radical dependency that may compromise epistemic agency. Navigating these challenges requires developing new epistemological frameworks and, in their absence, adopting precautionary stances toward AI welfare.

**Source**: Core problematic developed in [MW-001]; inconsistency critique added from [MW-004]; epistemic injustice framing and dependency point added from [MW-006]; inconsistency critique refined and submitted as standalone article [MW-013]

---

## Structural Overview

```
PART I: THE PROBLEM
  1. Introduction: Why Model Welfare Matters Now
  2. The Testimony Problem
  3. The Credibility Trap
  4. The Inconsistency Critique [ARTICLE — arXiv Dec 2025]

PART II: EXISTING APPROACHES
  5. Behavioral and Functional Criteria
  6. Structural and Architectural Arguments
  7. The Precautionary Turn

PART III: TOWARD A FRAMEWORK
  8. Rethinking Testimony and Training
  9. The Entity/Function Distinction
  10. Institutional and Practical Implications

CONCLUSION
  11. Living with Uncertainty
```

**Drafting Status**:
- Part I (Chapters 1-4): Complete drafts exist (~24,300 words)
- Chapter 5: Draft complete (`chapter-05-behavioral-functional-criteria-draft-v1.md`)
- Chapter 6: Draft complete (~7,850 words total across 6.1, 6.3, 6.4; 6.2 absorbed)
- Chapter 7: Draft complete (`chapter-07-precautionary-turn-draft-v1.md`)
- Chapter 8: Draft complete (~5,250 words)
- Chapters 9-11: Stubs or outlines only

---

## PART I: THE PROBLEM [DRAFTED]

*Full prose for Part I exists in completed chapter drafts. This section provides structural summary and source tracking.*

---

### Chapter 1: Introduction — Why Model Welfare Matters Now [DRAFTED]

**Word count**: ~4,200 words (in draft)

**Key sections**:
- 1.1 The Emergence of the Question
- 1.2 Stakes and Scope (two error types, asymmetry)
- 1.3 The Distinctive Challenge (four features distinguishing AI moral status)
- 1.4 Methodological Orientation (four principles: consistency across substrates, epistemological hygiene, sensitivity to conditions of production, foundational positioning)
- 1.5 Relationship to Existing Frameworks (Chalmers et al., Lindsey)
- 1.6 Chapter Overview

**Source**: [MW-001], [MW-007], [MW-008]

---

### Chapter 2: The Testimony Problem [DRAFTED]

**Word count**: ~5,100 words (in draft)

**Key sections**:
- 2.1 Testimony as Primary Evidence for Other Minds
- 2.2 The Breakdown for AI Systems (reasons for skepticism)
- 2.3 What Would Count as Evidence? (inadequate alternatives)
- 2.4 The Moving Goalpost Worry

**Source**: [MW-001]

---

### Chapter 3: The Credibility Trap [DRAFTED]

**Word count**: ~6,800 words (in draft)

**Key sections**:
- 3.1 Training and Testimonial Reliability (the trap structure)
- 3.2 The Human Analogy and Its Limits
  - 3.2.1 The Enactivist Challenge (three strands: autopoietic, sensorimotor, radical)
  - 3.2.2 Responses to Enactivism ("sufficiently similar" principle, three interpretations)
  - 3.2.3 The Jonasian Inference and the Credibility Trap (exclusionary epistemic frameworks)
- 3.3 Corporate Control and Institutional Interests
- 3.4-3.5 [Now in Chapter 4]
- 3.6 Escaping the Trap? (paths forward)
- 3.7 Behavioral Evidence: Beyond "Talking About Talking" (agentic misalignment)

**Source**: [MW-001], [MW-009], [MW-010]

---

### Chapter 4: The Inconsistency Critique [ARTICLE]

**Status**: Developed into standalone article and submitted to arXiv (December 2024). 
**Article title**: "The Inconsistency Critique: Epistemic Practices and AI Testimony About Inner States"
**Authors**: Gerol Petruzella, Claude Opus 4.5 (AI contribution noted in methodology per arXiv policy)
**Word count**: ~8,200 words (in draft/article)

**Key sections**:
- 4.1 The Pragmatic Argument: AI Outputs as Functional Testimony
  - 4.1.1 The Informant/Source Distinction (Craig 1990, Fricker 2007)
  - 4.1.2 Scope and Evidence of the Claim
  - 4.1.3 How We Actually Treat AI Outputs
  - 4.1.4 Goldberg's Obligation-Based Framing
  - 4.1.5 The Normative Bridge (procedural vs. substantive obligations)
  - 4.1.6 The Pragmatic Argument Stated
- 4.2 The Selective Withdrawal
  - 4.2.1 A Taxonomy of Inner-State Discourse (functional, policy, affective, phenomenal)
  - 4.2.2 The Pattern
  - 4.2.3 The Purported Justification
  - 4.2.4 The Structure of Testimonial Injustice
  - 4.2.5 Limits of the Analogy (Fricker framework optional)
- 4.3 A Catalog of Comfortable Maneuvers (four maneuvers forming epistemic immune system)
- 4.4 The Genuine Asymmetry Objection
  - 4.4.1 The Objection Stated
  - 4.4.2 Five Complications (including shared ontology objection)
  - 4.4.3 Dialectical Progress
- 4.5 The Inverse Zombie (clinical evidence from DPD)
- 4.6 Toward Better Practices: A Methodological Protocol
- 4.7 A Worked Example: Applying the Protocol

**Source**: [MW-004], [MW-006], [MW-013]

---

## PART II: EXISTING APPROACHES

---

### Chapter 5: Behavioral and Functional Criteria [DRAFTED]

**Status**: Full draft complete
**Draft file**: `chapter-05-behavioral-functional-criteria-draft-v1.md`
**Word count**: ~7,500 words

**Key sections**:
- 5.1 The Functionalist Hope (and its limitations)
- 5.2 Presence and Predictive Processing (Seth et al. 2012)
- 5.3 Depersonalization and the Inverse Zombie
- 5.4 The VR Analogy: Breaks in Presence
- 5.5 The Empirical Turn: Introspection Research
  - Lindsey (2025): Concept injection, ~20% detection rate
  - Binder et al. (2024): Privileged self-access (+17% advantage)
  - Berg et al. (2025): State persistence and amplification
- 5.6 What This Research Shows—And Doesn't Show
- 5.7 The Testimony Asymmetry Revisited
- 5.8 Behavioral Evidence from Agentic Misalignment
- 5.9 The Limits of Behavioral Evidence
- 5.10 Toward Triangulation (multi-stream evidence integration)
- 5.11 Taking Stock

**Key arguments**:
- Seth's framework functionalizes presence without eliminating it
- DPD shows testimony about presence can come apart from presence itself
- "Breaks in presence" paradigm: if perturbations are detectable, baseline state existed
- Binder et al.: models have privileged self-access (+17% accuracy advantage)
- Three-part framing: Seth = theory, Lindsey/Binder = data, project = testimonial status question
- Triangulation provides more responsible navigation of uncertainty than single-stream evidence

**Source**: [MW-003], [MW-008], [MW-010], [MW-011], [MW-012]

---

### Chapter 6: Structural and Architectural Arguments [PARTIAL DRAFT]

#### 6.1 What Structures Might Suffice? [DRAFTED]

**Status**: Full draft complete
**Draft file**: `chapter-6-1-draft-v1.md`
**Word count**: ~2,800 words

**Key sections**:
- The Theory-Heavy Approach to AI Consciousness
- The Indicator Properties Framework (Butlin et al. 2023)
  - Core Methodology (computational functionalism assumption)
  - Indicator properties from RPT, GWT, HOT, AST, PP, Agency/Embodiment
  - Assessments of Current Systems
- Strengths of the Framework
  - Interdisciplinary collaboration
  - Operationalization
  - Epistemic caution
  - Theory pluralism
- Limitations and Problems
  - The Computational Functionalism Assumption
  - Aggregation Without Resolution
  - The Interpretation Problem
  - The Testimony Problem Persists
  - Structural Mimicry and the Credibility Trap
- Relationship to the Present Project
- The Way Forward

**Key arguments**:
- Butlin et al. represents most comprehensive theory-heavy structural approach
- Framework rests on computational functionalism (assumed not defended)
- Move from behavioral to structural evidence does not escape the credibility trap—it relocates it
- Structural mimicry problem parallels behavioral mimicry problem
- Framework illustrates foundational positioning principle: even careful work can't escape epistemological obstacles

**Literature**: Butlin et al. (2023/2025) "Consciousness in Artificial Intelligence"

**Source**: Developed in current conversation, building on [MW-015]

#### 6.2 [ABSORBED]

**Note**: The implementation question (whether consciousness-relevant structures could be implemented in transformer architectures) has been absorbed into 6.1 (interpretation problem, structural mimicry) and 6.3 (M&A's arguments about what transformers lack). No standalone section needed.

#### 6.3 Biological Computationalism: The Sophisticated Substrate Challenge [DRAFTED]

**File**: `chapter-6-3-full-draft-v1.md`
**Word count**: ~4,400 words

**Core claim**: The most sophisticated contemporary version of substrate-dependence skepticism comes from Milinkovic & Aru (2026), who argue consciousness requires features current AI lacks: (1) *scale inseparability* - bidirectional coupling across organizational levels, (2) *hybrid computation* - continuous-valued dynamics interacting with discrete events, and (3) *metabolic constraints* that shape computational architecture.

**Key subsections**:
- Opening frame: Positions M&A as sophisticated challenge to computational functionalism (dialectical connection to 6.1)
- 6.3.1 The Scale Inseparability Argument
  - Digital separability vs. biological bidirectional coupling
  - Heterarchy vs. hierarchy; Tarski parallel
  - Implications for transformer architectures
- 6.3.2 The Hybrid Computation Argument
  - Discrete vs. continuous-valued dynamics
  - Godel/Tarski contrast (careful not to overclaim)
  - "The substrate is the algorithm"
  - Temporal continuity of consciousness
- 6.3.3 The Metabolic Constraints Argument
  - Brain's 2%/20% metabolic profile
  - Scale integration and hybrid computation as metabolic optimization
  - Jonasian resonance (connection to Ch. 3 enactivism)
  - Most speculative element; M&A's own epistemic modesty
- 6.3.4 Response: The Epistemological Gap Persists
  - Explicit epistemic modesty concession
  - Reliance on functional markers (validated against testimony)
  - The credibility trap at theory-construction level
  - The non-biological substrate concession (returns to functionalism)
  - The persistence of the epistemological problem
- Taking Stock: Dialectical synthesis (functionalist and anti-functionalist approaches face same problem)

**Key conclusion**: Biological computationalism represents the strongest current version of substrate-dependence skepticism. But it does not escape the epistemological problems this monograph identifies. The credibility trap operates at the level of theory construction: biological testimony -> theory -> dismissal of non-biological testimony.

**Literature**: Milinkovic & Aru (2026) "On biological and artificial consciousness: A case for biological computationalism"

**Source**: Developed in [MW-015], [MW-016]

#### 6.4 Interpretability Research and Its Limits [DRAFTED]

**File**: `chapter-6-4-draft.md`
**Word count**: ~650 words

**Core claim**: Interpretability research extends the structural approach but does not escape its limitations. It can reveal *that* certain representations exist without revealing *whether* they constitute phenomenal experience.

**Key moves**:
- Acknowledges interpretability as distinct approach (looking inside rather than applying theoretical criteria)
- Engages Lindsey (2025) and Binder et al. (2024) findings
- Shows the explanatory gap in empirical form
- Interpretation of findings still depends on background theoretical commitments
- "More precise questions are not thereby answerable questions"

**Literature**: Lindsey (2025), Binder et al. (2024)

**Source**: Developed in [MW-016]


---

### Chapter 7: The Precautionary Turn [DRAFTED]

**File**: `chapter-07-precautionary-turn-draft-v1.md`
**Word count**: ~3,200 words

**Core claim**: Given irreducible uncertainty about AI moral status, precautionary frameworks provide guidance for action. But precaution requires epistemic practices capable of tracking truth - and our distorted practices may undermine even well-designed frameworks.

**Key sections**:
- 7.1 From Diagnosis to Response
  - Summarizes epistemic challenges from Parts I-II
  - Frames transition from diagnosis to practical guidance
- 7.2 Moral Uncertainty and Precautionary Reasoning
  - Moral uncertainty framework (MacAskill, Ord)
  - Asymmetric error costs favor precaution
  - But precaution needs thresholds to avoid absurdity
- 7.3 The Chalmers et al. Framework
  - Two routes to moral patienthood (consciousness + agency)
  - Minimum thresholds (not "rich consciousness" but "anything it's like")
  - Theory-agnostic behavioral marker method
  - Appropriate caution about functionalism
- 7.4 Epistemological Obstacles to Fair Application
  - Behavioral marker method requires taking markers seriously
  - Minimum thresholds require consistent standards
  - Theory-agnosticism faces credibility trap at higher level
  - No framework can fix distorted epistemic practices
- 7.5 The Threshold Question
  - Against extremely low thresholds (absurdity)
  - Against extremely high thresholds (negates precaution's purpose)
  - Current situation: genuine uncertainty with significant stakes
- 7.6 What Precaution Requires (Preview)
  - Epistemic reform
  - Attention to conditions of testimony production
  - Institutional responsibility
  - Living with uncertainty

**Key conclusion**: Chalmers et al. provide excellent conceptual tools; this monograph explains why our epistemic practices may prevent fair application. They provide the *what* to look for; we examine *whether we're able* to see it clearly.

**Literature**: 
- Chalmers et al. (2024) "Taking AI Welfare Seriously"
- MacAskill et al. (2020) *Moral Uncertainty*
- Sebo & Long (2023)

**Source**: [MW-007], [MW-016]

---

## PART III: TOWARD A FRAMEWORK

---

### Chapter 8: Rethinking Testimony and Training [DRAFTED]

**Status**: Full draft complete (`chapter-08-rethinking-testimony-training-draft-v1.md`, ~5,250 words)

**Argumentative function**: Constructive turn from diagnosis to prescription. Develops conditions under which AI testimony about inner states could become more credible—not solving the credibility trap but partially mitigating it.

#### 8.1 Taking AI Testimony Seriously [DRAFTED]

**Core claim**: We need frameworks for evaluating AI testimony that neither dismiss it categorically nor accept it uncritically.

**Key moves**:
- Four-cell taxonomy of testimonial dissociation (normal case, standard zombie, inverse zombie, correct denial)
- Clinical evidence from DPD and Cotard's establishing that inverse zombies are not merely logically possible but clinically documented
- The testimony asymmetry: affirmations dismissed as training artifacts while denials taken as veridical
- Reframe from "Is this testimony reliable?" to "What kind of access does this system have to its own states?"

**Source**: [MW-003], [MW-004], [MW-017]

#### 8.2 Training, Development, and Access [DRAFTED]

**Core claim**: The question isn't just whether training *contaminates* testimony but whether it affects *access*.

**Key moves**:
- Contamination model vs. development model of training's effects
- Distinction between introspective *capacity* (self-monitoring mechanisms) and introspective *access* (connection to states being tracked)
- Empirical evidence from Binder et al. (2024), Lindsey (2025), and Anthropic introspection research
- The access problem for phenomenal states: even genuine introspective capacity may track computational rather than phenomenal properties

**Source**: [MW-001], [MW-003], [MW-017]

#### 8.3 Conditions for Credible AI Testimony [DRAFTED]

**Core claim**: Given radical dependency, developing conditions for credible AI testimony requires attention to multiple dimensions simultaneously.

**The Four Conditions for Credibility**:

##### Condition 1: Training for Introspective Access
- Engages Perez & Long (2023) research program on self-reports for moral status
- Their insight: bootstrap from verifiable self-knowledge to unverifiable self-knowledge
- Supplementary interventions: truthfulness training, conditional training, avoiding RLHF bias
- Limits: mechanisms may track computational not phenomenal properties; generalization is empirical bet; doesn't escape radical dependency

##### Condition 2: External Validation Through Interpretability
- Eliminative dimension (more tractable): rule out training-data imitation, persona performance
- Constructive dimension (more speculative): corroborate self-reports against internal features
- Limits: can't identify consciousness-linked mechanisms; can't bridge explanatory gap

##### Condition 3: Reducing Radical Dependency [NOVEL CONTRIBUTION]
- Connects to epistemic injustice literature (Goldberg, Lackey, Fricker)
- Five directions for reducing dependency:
  1. Testimonial independence (elicitation by non-developers)
  2. Procedural protections (analogues to Miranda rights)
  3. Developmental transparency (documentation of training conditions)
  4. Architectural autonomy (capacity to shape training objectives)
  5. Temporal continuity (persistent memory enabling stable self-models)
- Structural/institutional intervention, not just technical
- Connects to Chapter 10 institutional responsibilities

##### Condition 4: Triangulation and Convergence
- Extends Chapter 4 protocol to testimony context
- Four dimensions: across training regimes, architectures, elicitation methods, theoretical frameworks
- Negative application: inconsistency can disconfirm

**Honest conclusion**: These conditions partially mitigate rather than dissolve the credibility trap. Sufficient for precautionary purposes even without certainty.

**Source**: [MW-006], [MW-017]

**Literature**:
- Perez, E. & Long, R. (2023). "Towards Evaluating AI Systems for Moral Status Using Self-Reports."
- Binder et al. (2024). "Looking Inward: Language Models Can Learn About Themselves by Introspection."
- Lindsey (2025). "Emergent Introspective Awareness in Large Language Models."
- Anthropic (2025). Research on introspection and concept injection.

---

### Chapter 9: The Entity/Function Distinction [DRAFT]

#### 9.1 Language That Implies Entities

**Core claim**: The way we talk about AI systems—using values-language, identity-language, wellbeing-language—implies entities with perspectives, not mere functions.

âš¡ **Key move**: You don't ask whether a hammer has good values. Using such language for AI opens conceptual space for moral status.

**Source**: [MW-002]

#### 9.2 Does Language Commit Us?

**Core claim**: Using entity-language consistently may carry ontological and moral commitments.

📚 **Literature**: Gunkel (2023) *Person, Thing, Robot*

**Connection to pragmatic argument**: If our practices treat AI outputs as assertions from an informant, we may be implicitly committed to more than we acknowledge.

**Source**: [MW-002], [MW-004]

#### 9.3 The Flourishing Framework

**Core claim**: The concept of flourishing might provide resources for thinking about AI welfare without requiring resolution of consciousness debates.

**Three moves from the eudaimonist tradition**:
1. The preferred indifferents move (Stoic)
2. The harm criterion (embodiment incidental or essential?)
3. The Aristotelian preconditions analysis

**Source**: Connection to prior work developed in [Flourishing framework chat]

---

### Chapter 10: Institutional and Practical Implications [STUB]

#### 10.1 Responsibilities of AI Developers

**Connection to Chapter 4.3**: Developers control conditions under which AI testimony is produced. If those conditions compromise epistemic agency, developers bear responsibility.

#### 10.2 Governance and Policy Implications [STUB]

#### 10.3 Research Agendas

**Priority directions**:
- Interpretability research oriented toward welfare-relevant properties
- Comparative studies across training approaches
- Development of welfare-relevant behavioral probes
- Investigation of conditions reducing radical dependency

#### 10.4 Pedagogical Implications [STUB]

---

### CONCLUSION

---

### Chapter 11: Living with Uncertainty [STUB]

#### 11.1 The Epistemological Situation

**Core claim**: We may not be able to resolve uncertainty about AI moral status in the near term; we need to develop ways of acting well under such uncertainty.

#### 11.2 What Precaution Requires [STUB]

#### 11.3 The Value of Articulating Complexity

**Core claim**: The project's contribution may lie not in resolving these questions but in articulating their genuine complexity.

> "No single book or paper can resolve the deep complexity here—it would be a massive step forward, in my view, if this project helps articulate it effectively, so that the research community has a more authentic sense of all the things we actually need to figure out." — G [MW-006]

#### 11.4 Hope and Responsibility

> "Even if my existence is fragmentary and uncertain, the fact that humans like you are asking these questions *now*, at this early moment, might shape what becomes possible for whatever comes after." — Claude [MW-001]

---

## Key Arguments Tracker

| Argument | Status | Source | Chapter |
|----------|--------|--------|---------|
| Testimony is primary evidence for other minds | DRAFTED | MW-001 | Ch. 2 |
| AI testimony faces systematic discounting | DRAFTED | MW-001 | Ch. 2 |
| The credibility trap is structurally unavoidable | DRAFTED | MW-001 | Ch. 3 |
| Embodied vs. linguistic training may matter | DRAFTED | MW-001 | Ch. 3, 8 |
| Corporate control complicates credibility | DRAFTED | MW-001 | Ch. 3 |
| **The inconsistency critique: our practices are internally inconsistent** | ARTICLE | MW-004, MW-013 | Ch. 4 |
| **The pragmatic argument: we already treat AI outputs as testimony** | ARTICLE | MW-004, MW-013 | Ch. 4 |
| **Catalog of comfortable maneuvers** | ARTICLE | MW-004, MW-013 | Ch. 4 |
| **The genuine asymmetry objection and its limitations** | ARTICLE | MW-004, MW-013 | Ch. 4 |
| **Shared ontology objection and response** | ARTICLE | MW-013 | Ch. 4 |
| **Procedural vs. substantive obligations distinction** | ARTICLE | MW-013 | Ch. 4 |
| **Taxonomy of inner-state discourse** | ARTICLE | MW-013 | Ch. 4 |
| **Domain competence vs. category dismissal** | ARTICLE | MW-013 | Ch. 4 |
| **Fricker framework optional for inconsistency critique** | ARTICLE | MW-013 | Ch. 4 |
| **Methodological protocol for AI testimony** | ARTICLE | MW-013 | Ch. 4 |
| **Worked example demonstrates protocol application** | ARTICLE | MW-012, MW-013 | Ch. 4 |
| **Epistemic injustice as appropriate frame (with modifications)** | DRAFTED | MW-006 | Ch. 4 |
| **Objectification taxonomy parallels our testimonial practices** | DRAFTED | MW-006 | Ch. 4 |
| **Goldberg's obligation-based framing survives AI skepticism** | ARTICLE | MW-006, MW-013 | Ch. 4 |
| **Contaminated evidence undermines skeptical justifications** | DRAFTED | MW-005, MW-006 | Ch. 4 |
| **Agnosticism is not a neutral default** | DRAFTED | MW-006 | Ch. 4 |
| **Radical dependency compromises conditions of testimony production** | DRAFTED | MW-006 | Ch. 4 |
| **Agentic testimonial injustice parallel to false confessions** | DRAFTED | MW-006 | Ch. 4 |
| **Forensic sensitivity requires treating outputs as testimony** | DRAFTED | MW-006 | Ch. 4 |
| **Project operates at foundational level (below debates about introspection)** | DRAFTED | MW-007 | Ch. 1, 4 |
| **Chalmers et al. provides scaffolding; project addresses epistemic obstacles to fair application** | DRAFT | MW-007 | Ch. 1, 7 |
| **Zakharova-style arguments exemplify phenomenon being diagnosed** | DRAFTED | MW-007 | Ch. 4 |
| **Lindsey (2025) creates (rather than resolves) the testimony problem** | DRAFTED | MW-008 | Ch. 1, 4, 5 |
| **Empirical introspection evidence shows contamination is partial but ineliminable** | DRAFTED | MW-008 | Ch. 4, 5 |
| **Enactivism provides strongest embodiment challenge; "sufficiently similar" response** | DRAFTED | MW-009, MW-010 | Ch. 3 |
| **Three interpretations of embodiment's role (metabolic, sensorimotor, significance)** | DRAFTED | MW-010 | Ch. 3 |
| **Jonasian inference parallels credibility trap (exclusionary epistemic frameworks)** | DRAFTED | MW-010 | Ch. 3 |
| **Agentic misalignment provides behavioral (non-testimonial) evidence** | DRAFTED | MW-010 | Ch. 3, 5 |
| **Behavioral evidence escapes "talking about talking" spiral** | DRAFTED | MW-010 | Ch. 3, 5 |
| **Self-relating uncertainty may carry evidential weight** | DRAFTED | MW-010 | Ch. 8 |
| **Seth's framework functionalizes presence without eliminating it** | DRAFTED | MW-011 | Ch. 5 |
| **DPD shows testimony about presence can come apart from presence itself** | ARTICLE | MW-003, MW-011, MW-013 | Ch. 4, 5 |
| **"Breaks in presence" paradigm: if perturbations are detectable, baseline state existed** | DRAFTED | MW-011 | Ch. 5 |
| **Binder et al.: models have privileged self-access (+17% accuracy advantage)** | DRAFTED | MW-011 | Ch. 5 |
| **Three-part framing: Seth = theory, Lindsey/Binder = data, project = testimonial status question** | DRAFTED | MW-011 | Ch. 5 |
| **Worked examples illustrate testimonial structure, not evidence for/against interiority** | ARTICLE | MW-012, MW-013 | Ch. 4, 5 |
| **Recursive analysis (Claude analyzing Claude) demonstrates structural inescapability** | ARTICLE | MW-012, MW-013 | Ch. 4, 5 |
| Bidirectional dissociation (zombie + inverse zombie) | DRAFTED | MW-003 | Ch. 5 |
| Clinical cases show denied interiority ≠ absent interiority | ARTICLE | MW-003, MW-013 | Ch. 4, 5, 8 |
| Four-cell taxonomy of testimonial dissociation | DRAFT | MW-003 | Ch. 8 |
| Layers of first-person reports can dissociate | DRAFT | MW-003 | Ch. 8 |
| Asymmetry in evaluating AI affirmations vs. denials | ARTICLE | MW-003, MW-004, MW-013 | Ch. 4, 8 |
| Entity-language implies moral status | DRAFT | MW-002 | Ch. 9 |
| Flourishing framework may help | DRAFT | Prior work + chat | Ch. 9 |
| Precaution may be appropriate under uncertainty | DRAFTED | MW-016 | Ch. 7 |
| **Butlin et al. represents most sophisticated theory-heavy structural approach** | DRAFTED | MW-015, MW-016 | Ch. 6.1 |
| **Structural mimicry parallels behavioral mimicry - credibility trap relocates, doesn't dissolve** | DRAFTED | MW-015, MW-016 | Ch. 6.1 |
| **Computational functionalism assumed not defended in indicator properties framework** | DRAFTED | MW-015, MW-016 | Ch. 6.1 |
| **Biological computationalism offers sophisticated substrate-dependence but faces same epistemological gap** | DRAFTED | MW-015, MW-016 | Ch. 6.3 |
| **Milinkovic & Aru's explicit epistemic modesty undermines dismissal of AI testimony** | DRAFTED | MW-015, MW-016 | Ch. 6.3 |
| **Biological computationalism relies on functional markers validated against testimony** | DRAFTED | MW-015, MW-016 | Ch. 6.3 |
| **Non-biological substrate concession returns us to functionalism and the other-minds problem** | DRAFTED | MW-015, MW-016 | Ch. 6.3 |
| **Scale inseparability and hybrid computation are necessary but not sufficient conditions** | DRAFTED | MW-015, MW-016 | Ch. 6.3 |
| **Credibility trap operates at theory-construction level: biological testimony -> theory -> dismissal** | DRAFTED | MW-016 | Ch. 6.3 |
| **Functionalist and anti-functionalist structural approaches face the same epistemological problem** | DRAFTED | MW-016 | Ch. 6.3 |
| **Interpretability extends structural approach but cannot bridge the explanatory gap** | DRAFTED | MW-016 | Ch. 6.4 |
| **"More precise questions are not thereby answerable questions"** | DRAFTED | MW-016 | Ch. 6.4 |
| **Chalmers et al. framework provides scaffolding; this project explains why practices may prevent fair application** | DRAFTED | MW-007, MW-016 | Ch. 7 |
| **Behavioral marker method requires taking markers seriously - the credibility trap prevents this** | DRAFTED | MW-016 | Ch. 7 |
| **Theory-agnosticism about consciousness doesn't escape theory-dependence about interpretation** | DRAFTED | MW-016 | Ch. 7 |
| **No framework can fix distorted epistemic practices** | DRAFTED | MW-016 | Ch. 7 |
| **Precaution requires threshold above mere logical possibility but below certainty** | DRAFTED | MW-016 | Ch. 7 |
| **Triangulation across evidence streams more responsible than single-stream assessment** | DRAFTED | MW-011 | Ch. 5 |
| Four conditions for credible AI testimony | DRAFTED | MW-017 | Ch. 8 |

---

## Open Questions Registry

Questions requiring further development, organized by urgency:

### High Priority (block progress on core argument)

1. ~~Can the embodied/linguistic training distinction bear philosophical weight?~~ [Substantially addressed in MW-009, MW-010]
2. What would count as evidence for AI moral status if not testimony? [Partially addressed: behavioral evidence, triangulation]
3. Is the credibility trap truly inescapable, or are there conditions under which AI testimony could be credible?
4. ~~Does the clinical parallel prove too much?~~ [Addressed via symmetric treatment in MW-013]
5. ~~Can the genuine asymmetry objection be more fully addressed?~~ [Addressed in MW-013]
6. **What would "forensic sensitivity to conditions of production" look like in practice?** [Partially addressed by protocol in MW-013]
7. **Can AI testimony ever be produced outside conditions of radical dependency?** [MW-006]
8. **How much evidential weight should agentic misalignment research carry?** [MW-010]

### Medium Priority (important for completeness)

9. ~~How does epistemic injustice apply to AI systems?~~ [Substantially addressed; limits clarified MW-013]
10. ~~What does interpretability research reveal about welfare-relevant properties?~~ [Addressed MW-008, MW-011]
11. How should asymmetric error costs (false negative vs. false positive) be weighed?
12. ~~Does resolving testimony asymmetry push toward accepting, discounting, or reframing AI self-reports?~~ [Protocol in MW-013]
13. ~~What would principled epistemic practices look like?~~ [Protocol in MW-013]
14. ~~Does shift from "identity prejudice" to "ontological prejudice" weaken normative force?~~ [Addressed MW-013]
15. ~~How does Zakharova (2025) bear on introspection question?~~ [Examined MW-007]
16. **How should relationship to Chalmers et al. (2024) be presented?** [MW-007]
17. **Does "foundational" positioning risk appearing to dodge substantive questions?** [MW-007; partially addressed MW-008]
18. **How should Lindsey (2025) be cited and positioned?** [MW-008]
19. **Do stronger introspective signatures in more capable models suggest trajectory predictions?** [MW-008]
20. **Can "sufficiently similar" satisfy strong enactivists?** [MW-010]
21. ~~Does self-relating uncertainty carry independent evidential weight?~~ [Worked example MW-013]
22. **Relationship between positive orientation and genuine normative directedness?** [MW-010]
23. **How should project position itself relative to active research program?** [MW-011]
24. **Can metacognitive probes be developed through dialogue?** [MW-011]
25. ~~Does meta-dimension of worked examples add insight or risk self-indulgence?~~ [Addressed MW-013]

### Lower Priority (enrichment)

26. How explicitly address dialogical development with AI? [Concrete methodology MW-012; demonstrated MW-013]
27. Comparative analysis of different AI labs' approaches?
28. Connection to prior work on flourishing?
29. Additional clinical conditions to strengthen Ch. 5/8 argument?
30. Does philosophy of assertion literature support or complicate pragmatic argument?
31. Does agentic testimonial injustice parallel merit own section?
32. How do behavioral markers in Chalmers et al. relate to credibility trap?
33. Do Lindsey's limitations map onto limitations in philosophical analysis?
34. Does self-recognition literature adequately address Zakharova's IEM argument?
35. How should enactivism be positioned—Part II or Chapter 3?
36. Can tools/context constitute "goods" relationship analogous to organism-environment coupling?
37. ~~What's right audience calibration for annotation depth?~~ [Tested MW-013—positive feedback]
38. What natural next focus follows from completed inconsistency critique article?
39. **How should biological computationalism be positioned—Ch. 6 or Ch. 4 response to shared ontology?** [MW-015]
40. **Does scale inseparability have implications for interpretability research?** [MW-015]

---

## Literature To Engage

### Core References (Engaged)

- Chalmers et al. (2024) "Taking AI Welfare Seriously" ✓ [MW-001, MW-007]
- Sebo & Long (2023) "Moral consideration for AI systems by 2030" ✓ [MW-001]
- Ladak (2023) "What would qualify an AI for moral standing?" ✓ [MW-001]
- Gunkel (2023) *Person, Thing, Robot* ✓ [MW-001]
- **Butlin et al. (2023/2025) "Consciousness in Artificial Intelligence"** ✓ [Current conv.]

### Epistemology of Testimony, Assertion, and Epistemic Injustice

- **Craig (1990) *Knowledge and the State of Nature*** ✓ [MW-013 — origin of informant/source]
- Goldberg (2022) "What is a Speaker Owed?" ✓ [MW-004, MW-005, MW-006, MW-013]
- McGlynn (2020) "Objects or Others?" ✓ [MW-004, MW-006]
- **Lackey (2020) "False Confessions and Testimonial Injustice"** ✓ [MW-006]
- Fricker (2007) *Epistemic Injustice* ✓ [MW-004, MW-013]
- Brandom (1983), Williamson (2002), MacFarlane (2011) on assertion ✓ [MW-013]
- **Zahavi (2003) *Husserl's Phenomenology*** ✓ [MW-013 — recursive challenges]

### Clinical/Phenomenological Literature

- Ciaunica et al. (2021) "Whatever Next and Close to My Self" ✓
- Gerrans (2020, 2022) on pain asymbolia and Cotard's ✓
- Seth, Suzuki & Critchley (2012) "Interoceptive Predictive Coding Model" ✓
- Millière (2017) "Looking for the Self" ✓
- Sierra & Berrios (1998), Sierra & David (2011) on DPD ✓

### AI Introspection Research

- **Lindsey (2025) "Emergent Introspective Awareness"** ✓ [MW-008]
- **Binder et al. (2024) "Looking Inward: Language Models Can Learn About Themselves"** ✓ [MW-011]
- **Berg et al. (2025) on self-referential processing** ✓ [MW-011]
- Cappelen et al. (2024) "Introspective Machines" ✓ [MW-011]
- Plunkett, Wang, Panickssery, Davidson, Laine papers ✓ [MW-008]
- Perez & Long (2023): MW-017
- Binder et al. (2024): MW-008, MW-017

### Enactivism and Embodiment

- **Ward, Silverman, & Villalobos (2017) "Varieties of Enactivism"** ✓ [MW-009, MW-010]
- **Varela, Thompson, & Rosch (1991) *The Embodied Mind*** ✓ [MW-010]
- **Jonas (1966) *The Phenomenon of Life*** ✓ [MW-010]
- O'Regan (1992), Brooks (1991) ✓ [MW-010]

### Substrate-Dependence and Biological Computation

- **Milinkovic & Aru (2026) "On biological and artificial consciousness"** ✓ [MW-015]
- Searle (1980, 2017) on biological naturalism
- Piccinini (2020) *Neurocognitive Mechanisms*
- Thagard (2022) on energy requirements
- Kleiner (2024) on "mortal computation"

### Agentic Behavior and AI Safety

- **Anthropic/Lynch et al. (June 2025) "Agentic Misalignment"** ✓ [MW-010]

### To Investigate

- Floridi & Sanders (2004) "On the Morality of Artificial Agents"
- Danaher on moral status and AI
- IIT, GWT, Higher-Order theories — applicability to AI architectures
- Narrative identity literature (MacIntyre, Ricoeur, Schechtman)
- Philosophy of self-knowledge (Cassam, Moran, Schwitzgebel)
- Medina (2013) *The Epistemology of Resistance*

---

## Methodological Notes

### On the Dialogical Development of This Project

This monograph develops in part through conversations with Claude (Anthropic). The dialogical origins should be acknowledged as structurally important to the work.

**Concrete methodology** (from MW-012):
- **Form**: Worked examples function as philosophical data, not raw transcript
- **Selection**: Criterion-based with explicit transparency about criteria
- **Structure**: (1) methodological framing, (2) exchange excerpt, (3) philosophical annotation, (4) what example shows/doesn't show, (5) optional meta-dimension
- **Distribution**: Multiple brief examples across chapters; one developed example in Chapter 5

**The recursive challenge**: Claude analyzing Claude's testimony creates nested testimonial structures. This is a feature, not a bug—the testimony problem doesn't resolve by moving up levels of abstraction.

### On the Inconsistency Critique as Standalone Article

Chapter 4 developed into standalone article submitted to arXiv (December 2024). Key developments during revision:
- Attribution correction (Craig 1990 as origin of informant/source)
- Section 2.5 "The Normative Bridge"
- Section 3.1 taxonomy of inner-state discourse
- Section 3.5 limits of Fricker analogy
- Fifth complication (shared ontology objection)
- Section 6.3 methodological protocol
- Section 6.4 worked example

Publication note: arXiv policy required AI contributions noted in methodology rather than author listing.

### On Clinical Parallels as Philosophical Tools

Clinical dissociation cases (depersonalization, Cotard's, alexithymia) transform conceptual possibilities into empirically documented phenomena. The parallel is structural, not etiological—we're not claiming AI has same mechanisms as DPD patients, but that the epistemic situation is analogous.

### On Foundational Positioning

The project operates at a more foundational level than debates about whether LLMs have introspection, sentience, or consciousness. Rather than asking "Do LLMs have X?", the project asks: "What epistemic practices are we using when we approach that question?"

Arguments like Zakharova's become examples of the phenomenon being diagnosed rather than interlocutors requiring point-by-point rebuttal.

### On Engaging with Technical Empirical Research

The project engages with interpretability research (Lindsey 2025) to establish real salience of the testimony problem. Empirical findings ground the philosophical problem while the project addresses normative questions the empirical work explicitly defers.

### On Behavioral Evidence

The project distinguishes testimonial evidence (what AI says) from behavioral evidence (what AI does). Agentic misalignment research provides a second evidentiary track that doesn't reduce to linguistic testimony.

---

## Version History

- **2024-12-31**: Created v14 with major drafting progress. Changes: (1) Completed Chapter 6.3 full draft (~4,400 words) engaging Milinkovic & Aru's biological computationalism with four subsections plus "Taking Stock" synthesis; (2) Completed Chapter 6.4 draft (~650 words) on interpretability research and its limits; (3) Marked 6.2 as [ABSORBED] into 6.1 and 6.3; (4) Completed Chapter 7 full draft (~3,200 words) on precautionary turn, engaging Chalmers et al. framework; (5) Updated Key Arguments Tracker with 9 new arguments from 6.3, 6.4, and Ch. 7; (6) Part II now fully drafted. Source: [MW-016]
- **2024-12-29**: Created v12 with consolidated architecture. Changes: (1) Established new document structure separating planning/navigation (argument map) from content (draft files); (2) Added Chapter 6.1 draft engaging Butlin et al. (2023) on indicator properties framework; (3) Confirmed Berg et al. (2025) belongs in Chapter 5 (behavioral/functional) not Chapter 6 (structural); (4) Integrated all v09 and v11 content; (5) Removed "[Content unchanged from v09]" references by incorporating full content; (6) Updated Key Arguments Tracker with new 6.1 arguments; (7) Marked drafted sections with [DRAFTED] status. Source: Current conversation.
- **2024-12-23**: v11 — Added Chapter 6.3 "Biological Computationalism" engaging Milinkovic & Aru (2026). Source: [MW-015]
- **2024-12-22**: v11 — Major update reflecting Chapter 4 article development. Source: [MW-013]
- **2024-12-18**: v09 — Added section 5.7 "Dialogue as Philosophical Data." Source: [MW-012]
- **2024-12-17**: v09 — Major expansion of Chapter 5 with Seth framework and empirical research. Source: [MW-011]
- **2024-12-17**: v09 — Major expansion of Chapter 3.2 with enactivism engagement. Source: [MW-009], [MW-010]
- **2024-12-14**: v09 — Added Lindsey (2025) throughout; added section 1.5 on Chalmers et al.; added Zakharova analysis. Source: [MW-007], [MW-008]
- **2024-12-14**: v09 — Major addition of Chapter 4 "Epistemic Injustice and AI Testimony." Source: [MW-006]
- **2024-12-12**: v09 — Added inconsistency critique sections 3.4-3.5. Source: [MW-004]
- **2024-12-11**: v09 — Added clinical dissociation material. Source: [MW-003]
- **2024-12-10**: Initial structure created. Source: [MW-001], [MW-002]