Paper 1 (technical) — Human–AI Authorship, Skill Development, and Measurement

Provenance: verbatim copy of HUMAN–AI AUTHORSHIP, SKILL DEVELOPMENT, AND MEASUREMENT.txt, drafted 2026-05-03. Tier-1 foundational paper of the Penwright Research Program (see program.md and penwright-sub-paper-plan.md). A technical literature-review draft positioning Penwright against the existing HAI, educational-psychology, and epistemic-justice literatures. Companion public-audience version at penwright-paper-01-public.md.

Status: outline draft. Full prose draft pending — this file is the structural skeleton + claim inventory the prose draft will be built against.

---

📄 HUMAN–AI AUTHORSHIP, SKILL DEVELOPMENT, AND MEASUREMENT OUTLINE DRAFT

A Technical Literature Review for AI-Augmented Writing Systems Outline Draft

1. Introduction

The rapid integration of large language models (LLMs) into writing workflows has produced a fundamental tension:

AI systems can dramatically improve output quality, but may simultaneously degrade the underlying human capability to write, reason, and structure thought.

This tension is not merely practical—it is epistemological, cognitive, and developmental.

Most current AI writing systems:

optimize for output fluency rely on opaque training distributions position the human as editor rather than author

This raises three unresolved problems:

Skill Degradation (Deskilling Risk) Epistemic Opacity (Black-box knowledge sourcing) Measurement Absence (No framework for human improvement)

This paper synthesizes multiple research traditions to address:

How can AI systems be designed to improve human authorship capability, and how can that improvement be measured rigorously over time?

2. Conceptual Foundations 2.1 Phenomenology of Skill and AI-Augmented Expertise

The Dreyfus model (novice → expert) describes expertise as:

progressing from rule-based reasoning to intuitive, embodied performance

Hubert Dreyfus argued that:

expertise cannot be reduced to symbolic manipulation it emerges through situated, embodied engagement

Later work (e.g., Maurice Merleau-Ponty) introduces:

body schema → tools become extensions of perception Implication for AI Writing

LLMs risk:

interrupting the progression toward intuitive expertise replacing “coping” with “selection”

However, if properly integrated:

AI can become part of the extended cognitive system Key Tension Path Outcome AI replaces writing Deskilling AI scaffolds writing Skill development

👉 Penwright explicitly targets the second path.

2.2 Phenomenology of Attention

Work by Bernard Stiegler and Yves Citton reframes attention as:

a cultural and technical achievement, not just a cognitive resource

AI systems:

do not just offload cognition they restructure attentional ecology AI Writing Risk fragmentation of attention loss of sustained reasoning capacity collapse of deep work into reactive editing Penwright Alignment

Penwright’s:

Deep Writing Mode structured packet assembly

→ restore intentional attention structures

2.3 Transactive Memory and Cognitive Offloading

Transactive memory theory (Wegner):

knowledge is distributed across systems

Modern extension:

AI becomes a cognitive partner Risk over-reliance leads to: reduced recall reduced synthesis ability Opportunity

If structured:

AI can support: retrieval integration exploration

👉 Penwright’s corpus control + retrieval system is a controlled form of transactive memory.

3. Epistemology and Power 3.1 Standpoint Theory and Situated Knowledge

Donna Haraway:

“All knowledge is situated.”

Sandra Harding:

“Strong objectivity requires acknowledging perspective.”

AI Problem

LLMs:

appear neutral are actually: trained on historically dominant voices shaped by power structures Consequence

AI writing systems:

reinforce dominant narratives suppress marginal or emerging voices Penwright Innovation

Corpus Control Layer

Writers:

choose sources control influence override default distributions

👉 This is a direct operationalization of standpoint epistemology

3.2 Epistemic Injustice

Miranda Fricker:

testimonial injustice hermeneutical injustice AI Risk minority perspectives underrepresented certain experiences become “unspeakable” Penwright Contribution

By allowing:

custom corpus inclusion experience × emotion tagging

→ expands what can be articulated and retrieved

4. Developmental and Learning Theory 4.1 Cognitive Apprenticeship

Rooted in:

Lev Vygotsky (Zone of Proximal Development)

Learning occurs when:

support is provided just beyond current ability AI Application

LLMs can act as:

“more knowledgeable other”

BUT:

only if they scaffold—not replace—performance Penwright Alignment Practice Mode Constraint challenges pattern-based teaching

→ align with ZPD scaffolding

4.2 Skill Acquisition vs Automation

Educational research shows:

Approach Outcome Automation-first Reduced skill Scaffolded learning Increased skill Key Insight

The presence of AI is not the issue—the structure of interaction is

5. Relational Models of Human–AI Interaction 5.1 Working Alliance Theory

Edward Bordin:

alliance = goals + tasks + bond Application to AI

Human–AI systems require:

aligned goals transparent processes trust calibration Penwright Contribution authorship packet → shared goal structure explicit constraints → aligned tasks 5.2 Psychoanalytic Perspectives

Key constructs:

projection idealization dependency AI Risk

Users may:

over-trust outputs defer judgment collapse agency Penwright Safeguard non-compliant AI stance counterposition panels attribution visibility 6. Improvisation and Collaborative Cognition 6.1 Improv Theory

Keith Johnstone:

“yes-and” offer acceptance co-creation AI Analogy

Human–AI writing:

resembles collaborative improvisation Penwright Extension

But Penwright adds:

structured constraint deliberate divergence

👉 Not just “yes-and,” but:

“yes—and understand why”

7. Translation Theory 7.1 AI as Translator of Domains

Translation theory (Lawrence Venuti):

domestication vs foreignization fidelity vs transformation AI Behavior

LLMs:

translate: ideas → language domains → accessible forms Risk flattening nuance false equivalence Penwright Approach explicit source attribution multi-source integration counterposition

→ preserves difference, not just similarity

8. Niche Construction and Co-Evolution 8.1 Core Idea

Kevin Laland:

organisms shape environments environments shape organisms AI Implication

AI:

changes writing environment reshapes cognition itself Penwright Position

Penwright is:

a deliberately constructed cognitive niche

Designed to:

reinforce skill avoid degradation loops 9. Institutional and Economic Perspectives 9.1 Transaction Costs and AI

Oliver Williamson:

technologies reduce coordination costs AI Writing Effect writing becomes cheap quality becomes commoditized Penwright Differentiation shifts value from: output → capability 10. Measurement Gap (Critical)

Across all literature:

👉 There is no unified framework for measuring:

human skill development in AI-augmented systems longitudinally Existing Limitations HCI → short-term usability Education → controlled environments AI → output evaluation Missing Piece

A system that measures:

writing process skill acquisition independence over time 11. Penwright’s Contribution (Positioning) 11.1 Novel Integration

Penwright combines:

phenomenology of skill cognitive apprenticeship epistemic control transactive memory relational AI theory 11.2 Key Innovations

1. Authorship Packet

Structured input replacing prompts

2. Corpus Control

Epistemic agency for the writer

3. Measurement Framework

Multi-dimensional skill tracking

4. Learning Loop

Write → reflect → practice → improve

11.3 Research Contribution

Penwright introduces:

A measurable model of AI-augmented human skill development

12. Open Research Questions
13. Skill Transfer

Does improvement persist outside the system?

2. Optimal Assistance Level

How much AI support maximizes learning?

3. Dependency Thresholds

When does assistance become harmful?

4. Genre Differences

Do learning dynamics differ across:

nonfiction memoir fiction? 13. Conclusion

AI writing systems are at a crossroads:

optimize for output → risk cognitive decline optimize for development → enable human advancement

This review suggests:

The future of AI writing is not automation, but augmentation with accountability

Penwright represents:

a shift from generation → development from black-box AI → controlled epistemology from output quality → human capability