research / ai-human-interaction / reports
The twelve-paper Penwright Research Program plus six cross-cutting programs
Track A — twelve sub-papers across three tiers (foundational theory · measurement and mechanism · longitudinal empirical studies) drawing from a shared dataset generated by Penwright in production. Track B — six cross-cutting research programs the Penwright evidence may eventually seed.
AI–Human Interaction — research program
The program runs on two coordinated tracks, drawing from a single dataset generated by the Penwright system in production:
- The Penwright Research Program — twelve sub-papers across three tiers (foundational theory · measurement and mechanism · longitudinal empirical studies), centered on AI-augmented writing as the lead empirical case.
- Six cross-cutting research programs drawn from the broader HAI field map, where the methodological tradition is well-developed and the empirical work is thin.
Track A — The Penwright Research Program
You are not writing a paper. You are building a coordinated research system with multiple publishable outputs sharing a common dataset. — Penwright Sub-Paper Plan, v1.0
Tier 1 — Foundational (theory + framework)
Paper 1 — AI-Augmented Authorship and Human Skill Development. Establishes the problem (AI deskilling risk), the opportunity (developmental systems), and Penwright as a model. Reframes AI writing from generation to human capability development. Source domains: phenomenology of skill, attention theory, cognitive apprenticeship, feminist epistemology. Status: drafted in technical and public form (see penwright-paper-01-technical.md and penwright-paper-01-public.md).
Paper 2 — Corpus Control and Epistemic Agency in AI Systems. Formalizes the epistemic problem of LLMs and the role of corpus control as a design principle. Anchors: standpoint theory, epistemic injustice, AI-bias literature. Penwright tie-in: corpus selection, attribution visibility, source integration. Comparative data opportunity: writing outputs under default vs. curated corpora.
Paper 3 — Human–AI Writing as a Structured Cognitive Process. Replaces prompt-based models with structured authorship workflows. Defines the Authorship Packet Model as the unit of human-AI co-production. Anchors: transactive memory, translation theory, cognitive load theory. Penwright tie-in: packet builder, retrieval layer, structured drafting.
Tier 2 — Measurement and mechanism
Paper 4 — A Measurement Framework for AI-Augmented Writing Skill Development. Formalizes the Penwright Measurement Framework: six skill dimensions, six derived indices (Writing Quality, Independence, Integration, Metacognitive, Genre-Awareness, Authorial-Voice), three measurement layers, five-step learning loop. The first multi-dimensional framework for measuring writing skill in AI environments. Companion vision document at vela/docs/VISION-PENWRIGHT-MEASUREMENT.md.
Paper 5 — Dependency and Independence in AI-Assisted Writing. Quantifies when AI helps vs. harms. Identifies dependency thresholds. Hypotheses: early AI use helps; excessive reliance reduces independent capability. Data: corpus-reliance percentages, independent-writing performance under constraint mode. Penwright features used: constraint mode, independent-writing gate.
Paper 6 — Learning Loops in AI-Augmented Writing Systems. Evaluates whether structured feedback improves skill. Defines the learning-loop architecture as a falsifiable design construct. Penwright features used: teach moments, practice mode, reflection layer. Data: performance before/after intervention, pattern-recognition improvements.
Paper 7 — Genre-Specific Effects of AI Writing Systems. Compares nonfiction, memoir, and fiction. Pre-registered hypothesis: AI effects are genre-dependent — nonfiction risks shallow argument, memoir risks emotional flattening, fiction risks generic narrative. Data: genre-tagged sessions, skill progression by type. This paper is the load-bearing argument against treating "AI writing" as a single phenomenon.
Tier 3 — Empirical studies
Paper 8 — Longitudinal Effects of AI-Assisted Writing on Skill Development. The core validation study. Three groups: Penwright users · standard-AI users · no-AI controls. 3–6 month duration. Output: does Penwright create better writers, measured against the framework defined in Paper 4?
Paper 9 — Transfer of Writing Skill Outside AI Systems. Tests independence directly: blind writing tasks with no AI allowed. Capability-transfer rather than in-system performance is the key claim. The longitudinal test is "better writer with Penwright, than without it, in 6 months."
Paper 10 — Cognitive Load and Attention in AI Writing Systems. Studies how Penwright affects attention. Anchors: Stiegler's attention ecology, Citton's Ecology of Attention. Data: session duration, interruption patterns, flow-state indicators.
Paper 11 — Corpus Composition and Output Diversity. Tests the impact of corpus selection on writer output. Operationalizes epistemic-control claims from Paper 2.
Paper 12 — Human–AI Working Alliance in Writing Systems. Studies trust, reliance, and collaboration quality across writing sessions. Anchors: working-alliance theory (Bordin), psychoanalytic models of dependency and idealization. Cross-references Paper 5's dependency findings against alliance-quality measurements.
Cross-paper data strategy
All twelve papers draw from a single shared dataset:
- Penwright session telemetry (events, durations, packet structures)
- Writing outputs (drafts, final pieces, revision histories)
- Reflection logs and teach-moment responses
- Independent-writing gate samples (no-AI baseline tasks)
This is a single data-generating system feeding multiple papers — not twelve independent studies. The data discipline is itself a contribution.
Publication phases
| Phase | Window | Papers |
|---|---|---|
| 1 | 0–3 months | Paper 1 (drafted); Paper 4 (measurement); Paper 3 (authorship model) |
| 2 | 3–6 months | Paper 5 (dependency); Paper 6 (learning loops); Paper 7 (genre effects) |
| 3 | 6–12 months | Paper 8 (longitudinal); Paper 9 (transfer); Papers 10–12 (advanced) |
Product integration
| Paper | Penwright feature |
|---|---|
| Paper 1 — authorship model | Packet Builder |
| Paper 2 — corpus control | Corpus UI |
| Paper 4 — measurement | Adaptive Authorship Control Kernel (F-19) |
| Paper 6 — learning loops | Practice Mode |
| Paper 5 — dependency | Constraint Mode |
| Paper 7 — genre | Mode switching |
Track B — Six cross-cutting research programs
These six programs sit beyond the Penwright case. Each draws together a body of theoretical machinery the HAI mainstream has under-engaged with the empirical methods to test it. They are queued, not yet under formal study; the Penwright track produces the methodological proof-of-method first.
B1 — Longitudinal cognitive ethnography of professional AI use. A cohort of two to five years following lawyers, radiologists, or research scientists, instrumented and observed, tracking how skill, judgment, professional identity, and interpersonal dynamics shift. Methodological tradition: Hutchins's cognitive ethnography, Goodwin's interaction analysis. The cohort-level commitment is the hard part.
B2 — Comparative ontology of AI agents. Empirical mapping of the metaphors and conceptual schemes different populations (by culture, age, profession, religion) use to make sense of AI agents, and tracing which schemes lead to which behavioral and emotional outcomes. Brings cognitive linguistics, anthropology, and HCI into one frame.
B3 — Rupture-and-repair for human-AI interaction. Direct import from psychotherapy research. Studies what happens when AI interactions go wrong, how repair is attempted, what makes repair successful, and what the long-term effects of poorly-repaired ruptures are. Working-alliance theory operationalized.
B4 — Transactive memory for human-AI teams. Extends Wegner's framework empirically into teams that include AI participants. Specialization, coordination cost, adaptation when the AI member is removed or replaced.
B5 — Developmental cohort study of children with AI. Children growing up with AI from preschool through adolescence. Language acquisition, theory of mind, attachment, moral reasoning. Ethically and politically sensitive; exactly because of that, should be done with care rather than not done.
B6 — Sociology of AI epistemic authority. Empirical study of how communities of practice (medical, legal, scientific, journalistic) negotiate what kind of warrant AI outputs carry, and how those negotiations reshape the institutions of certified knowledge. Anchors: Collins & Evans on expertise, Knorr Cetina on epistemic cultures.
(See roadmap.md for the full field-mapping document the six programs are drawn from.)
How the two tracks relate
Track A is empirical — it ships papers with data. Track B is theoretical and methodological — it queues programs that the Penwright evidence base may eventually support. The Penwright dataset is rich enough to seed pieces of B3 (rupture-and-repair within Penwright sessions), B4 (transactive memory between writer and Penwright corpus), and the writing-specific arc of B1 (cognitive ethnography of professional writers using Penwright over time). The other arms wait their turn.