The Authorial Capacity Constraint

Abstract

Generative AI can improve the quality and sophistication of an intellectual artefact without proportionately improving the human user's understanding, metacognitive calibration or capacity to defend the claims it contains. This paper uses Authorial Capacity Constraint as a local organising label for that configuration while acknowledging close antecedents in AI-assisted metacognition, verification gaps, claim accountability, epistemic dependence, contributorship and computational opacity. The contribution claimed is synthesis and research architecture, not discovery of the underlying phenomenon. The paper separates artefact validity, human comprehension and epistemic accountability; recommends standardized contribution provenance plus a distinct claim-level authority record; treats low authority as a possible trigger for a provenance-preserving epistemic handoff; and connects the resulting verification and uptake burden to the Economics of Epistemic Abundance. A deeper prior-art review also replaces the earlier “Discovery Attribution Problem” with a Distributed Discovery Provenance programme grounded in existing computational-discovery, collective-discovery and scientific-priority literatures. The practical question is how to record what humans and machines actually contributed without mistaking local surprise for novelty, contribution for authority, or machine-readable provenance for a settled philosophy of knowledge.

1 · Intellectual lineage

The central disturbance is real, but it is not ours alone.

Hardwig's epistemic dependence shows that modern knowledge already relies on expertise an individual cannot personally reproduce. Humphreys' computational-opacity work shows that reliable computational processes can exceed a human's capacity to survey every epistemically relevant step.

Scientific publishing has long recognised that authorship compresses heterogeneous contributions. BMJ contributorship and guarantorship reforms separated “who did what” from responsibility, and CRediT later standardized fourteen contributor roles.

The closest AI-era antecedents are more direct. Fernandes and colleagues report improved LLM-assisted reasoning performance without comparable improvement in metacognitive accuracy. Maier separates assisted performance, understanding and self-evaluation. O'Keefe explicitly describes a “Verification Gap” between artefact quality and demonstrated understanding. Van Zoonen, Morgan-Thomas and Tursunbayeva propose claim accountability, requiring a named human positioned to reconstruct and defend scholarly claims.

The research question is what follows when those antecedents are combined with standards-compatible contribution provenance, claim-sensitive authority, explicit handoff and the economics of a rapidly expanding supply of candidate knowledge.

2 · The practical constraint

Production capacity and stewardship capacity can diverge.

Authorial Capacity Constraint: in machine-assisted knowledge production, the sophistication or apparent quality of an artefact can exceed the human curator's capacity to understand, interrogate, verify and appropriately qualify its claims.

The label is useful only if it helps separate three dimensions that are often collapsed:

• Artefact validity: is the claim, proof, analysis or result actually warranted?
• Human comprehension: what can the named human reconstruct, explain, transfer and diagnose?
• Epistemic accountability: which person, system or institution can responsibly answer for the claim, and on what basis?

A theorem may be correct while the human presenting it cannot follow the proof. A researcher may deeply understand an argument that is nevertheless wrong. Perfect provenance can tell us where a claim came from without making it true. The three dimensions therefore require separate evidence.

3 · Calibration rather than caricature

Dunning–Kruger is background, not the foundation.

The popular Dunning–Kruger caricature is too blunt to ground this work. The stronger foundation is the more modest empirical problem of metacognitive calibration: assisted performance can improve without an equivalent increase in accurate self-assessment of what the user understands or can do independently.

The relevant inference is correspondingly narrow:

4 · Defensible stewardship

Use claim accountability as a foundation, not a competitor.

This is closely aligned with claim-accountability and guarantorship traditions. The possible extension is a machine-readable research object in which different claims carry different authority bases rather than one undifferentiated authorial halo.

Distributed cognition is ordinary science. The failure is not dependence itself. It is hidden dependence or authority inferred from polish rather than warrant.

5 · Contribution without another taxonomy

Use CRediT first. Add only what it cannot express.

CRediT already provides standardized roles such as Conceptualization, Methodology, Investigation, Formal analysis, Validation, Software and Writing. Agalmic Research should not duplicate them.

Fine-grained process descriptors may be retained only when they preserve information materially relevant to the human–AI research process, such as distinguishing an initial problem framer from a later conceptual developer or recording selection among machine-generated candidates.

The experimental addition is therefore not “contributors have different roles.” It is a separate, claim-sensitive epistemic-authority record.

6 · Distributed discovery provenance

The machine-discovery question has already been asked.

Computational scientific discovery substantially predates generative AI. Bradshaw, Langley and Simon's BACON work simulated important discovery processes in the early 1980s. Langley later analysed multiple stages at which humans influence computational discovery systems and explicitly recommended human–computer cooperation.

Clark and Khosrowi (2022) is closer to the present concern. They argue that AI helps expose weaknesses in agent-centred accounts of discovery and propose a collective-centred view in which different contributors can perform different parts of a discovery. They directly consider the intuition that human recognition of significance might preserve discoverer status, while cautioning against treating that intuition as a complete criterion.

Scientific-priority scholarship supplies another correction. Gross treats discovery partly as retrospective social judgment. Vale and Hyman distinguish disclosure from later validation. Rubin and Schneider analyse how priority rules allocate credit and privilege.

Accordingly, the project no longer claims an original “Discovery Attribution Problem.” It adopts an operational programme:

1. problem framing / search direction;
2. candidate generation;
3. selection or significance recognition;
4. validation;
5. integration with prior knowledge;
6. realization.

These are proposed provenance events for empirical study, not a novel philosophical decomposition.

AlphaDev demonstrates why the machine contribution cannot always be described as mere retrieval: it discovered previously unknown sorting routines later integrated into LLVM. At the same time, the UK DABUS judgment shows that legal inventor status is a separate institutional question and may remain restricted to natural persons.

7 · Epistemic handoff

Low authority can trigger routing rather than erasure.

The deeper handoff review found extensive prior art in absorptive capacity, transfer stickiness, knowledge brokerage, innovation intermediaries, boundary spanning and transactive-memory systems. Generic routing is therefore not claimed as new.

The narrower local construct is:

This allows a non-expert originator's contribution to survive expert validation without turning origination into false authority or validation into retroactive origination.

8 · Economics of epistemic abundance

The uptake burden has strong economic ancestors.

Simon shows that information abundance creates attention scarcity. Cohen and Levinthal's absorptive capacity concerns recognition, assimilation and application of external knowledge. Weitzman's recombinant growth highlights processing constraints in a vast idea space. Teece shows rents moving toward complementary assets. Recent AI-era work explicitly models cheap cognitive capability coexisting with scarce expert judgment.

Accordingly, the project uses Economics of Epistemic Abundance for the narrower allocation question and retains Displaced Scarcity Hypothesis only as an organising synthesis, not a law.

The deeper scarcity review suggests that any formal contribution must distinguish technical bottlenecks, marginal/shadow values, rent capture, complementarity, substitutability and induced demand. Otherwise “scarcity moves” is merely standard production theory in new clothing.

9 · Standards-first provenance

The Memory Palace should compose existing standards.

W3C PROV already models entities, activities, agents, derivation and association. RO-Crate packages research objects in JSON-LD. CRediT standardizes scholarly contribution roles. The Memory Palace is therefore an experimental application profile and extension, not a new provenance standard.

Local metadata should remain small and research-specific: claim-level authority status, explicit handoff, search branch state, roads not taken, search boundaries and reversal conditions.

The repository now includes an experimental RO-Crate and W3C PROV-compatible JSON-LD. “Compatible” is intentionally weaker than claiming formal conformance or expert validation.

10 · Research programme

The surviving claims should become measurable.

1. How large is the gap between AI-assisted artefact quality and demonstrated human understanding across domains?
2. Which tests measure defensibility without merely measuring rhetoric, memory or familiarity?
3. Can claim-level authority metadata predict expert judgments about appropriate reliance?
4. When can formal verification or independent reproduction compensate for limited human comprehension?
5. Do provenance-preserving handoffs improve validation of useful outsider contributions?
6. How does machine-generated candidate abundance affect organizational and field-level absorptive capacity?
7. How should framing, generation, recognition, validation, integration and realization be credited in distributed discovery?
8. Can W3C PROV, RO-Crate and CRediT represent enough of the process that local extensions stay small?
9. Can a formal Displaced Scarcity model generate predictions beyond established complementarity and bottleneck theory?

11 · Deep future

What if reliable knowledge outruns human-scale comprehension?

The speculative Human-Scale Epistemic Horizon remains only a project label over an established family of computational-opacity problems. Mathematics is a useful test case because formal verification can make correctness and human surveyability come apart cleanly.

If opacity became normal at a knowledge frontier, defensibility might migrate from personally reconstructing every inference to defending the verification architecture, assumptions, independent checks and known failure modes. This is future work, not a forecast.

12 · Reflexive application

Agalmic Research is exposed to the exact failure it studies.

Large language models make independent rediscovery deceptively easy: a coherent machine-assisted synthesis can feel unprecedented before the relevant literature has been retrieved. The project's response is now procedural. Concepts receive lineage review before promotion; mature standards are reused where possible; and close antecedents narrow or rename claims.

The human curator initiated the central concern about limited capacity to defend machine-assisted work, insisted on transparency, introduced the separation between idea initiators and those with epistemic authority, and raised the latest question about machine generation versus human direction and significance recognition. ChatGPT materially contributed literature search, terminology, synthesis, criticism and drafting.

Those contributions are provenance facts. They are not evidence that either participant has established the paper as novel or true.

13 · Limitations

What v0.4 still does not establish.

The literature review is deepening but remains non-systematic. Scopus, Web of Science, EconLit, PhilPapers, ProQuest, patent databases, archival books and non-English scholarship have not all been exhausted.

No validated metric of defensibility, comprehension, claim-level authority, handoff quality or uptake capacity is supplied. The authority layer may prove cumbersome, gameable or redundant. The philosophy of machine discovery is not resolved here.

The Displaced Scarcity Hypothesis is not a law. The Human-Scale Epistemic Horizon is not a claim to have discovered opacity. Standards interoperability has not yet been reviewed by W3C PROV, RO-Crate or scholarly-metadata experts.

AI-assistance & contribution declaration

How this version was produced.

Human contribution: Conceptualization, search direction, selection and publication decisions. The human curator initiated the concern about authorial defensibility, introduced the separation between idea initiation and epistemic authority, and required extensive correction against prior art.

Machine contribution: OpenAI ChatGPT materially contributed Investigation/literature search, synthesis, terminology, criticism, original drafting and revision.

Authority: neither contribution statement establishes independent domain expertise or peer validation. Further handoff remains appropriate to scholars in metacognition, epistemology, philosophy of science, scholarly communication, economics/innovation studies and provenance standards.