Skip to main content
Date: 2026-07-17 For: honest-scholar plugin (meta-spec §2.1) — the single most important guiding principle. Status: Verified-source digest. Migrates to the plugin’s resources/references/.
This is a curated, primary-source digest. It exists so the agency principle is recorded as a grounded rule, not an opinion. All identifiers were verified at compile time except where flagged.

The principle

The honest-scholar skills are assistants, not researchers. They keep the accounts of a research program, advise as a mentor, and discuss as a colleague, but they do not perform independent research and do not make material scientific decisions. Every material decision — is a hypothesis confirmed / refuted; is a result real; is a paper worth publishing; what the thesis claims; is it defensible — is the human researcher’s, recorded with a named human sign-off. You author; the skill drafts. Automation covers the mechanical and mnemonic, never the judgemental. The researcher is in the driving seat.
It is not an anti-AI stance. Every norm below permits AI assistance (editing, retrieval, bookkeeping, discussion) while forbidding AI authorship and AI decision-making. The principle draws that bright line and keeps the skills useful on the permitted side while structurally refusing to cross it.

Why it holds — grounded

1. Authorship & accountability: accountability is non-delegable and attaches only to humans

  • ICMJE — authorship requires final approval and accountability; AI tools cannot be authors because they cannot take responsibility, declare conflicts, or hold copyright; humans remain responsible for AI-produced content.
  • COPE position statement, “Authorship and AI tools” (13 Feb 2023) — AI cannot be listed as an author; use must be disclosed; authors are fully responsible.
  • Nature/Springer Nature — LLMs fail authorship criteria because “authorship carries accountability… which cannot be effectively applied to LLMs.”
  • Science/AAAS — Thorp, “ChatGPT is fun, but not an author,” Science 379(6630):313 (2023); undisclosed AI-generated text treated as plagiarism.
  • Elsevier / Wiley / T&F / SAGE — same consensus: only humans/legal persons can hold accountability.
  • ML venuesICML 2023: text produced entirely by LLMs prohibited (except as experimental object); polishing author text allowed; the human is “ultimately responsible.” NeurIPS: LLMs allowed as tools but must be described; responsibility stays with authors. ACL 2023: language/search assistance OK, but authors must read/verify every reference; errors and plagiarism are the author’s responsibility.
  • CRediT (ANSI/NISO Z39.104-2022) — 14 contributor roles attribute contribution to named people, not tools.

2. Human oversight / human-in-command: decision authority stays human

  • EU AI Act, Article 14 (human oversight) — high-risk AI must be designed for oversight by natural persons who can interpret output, override it, or decide not to use it, and must “remain aware of… over-relying on the output (automation bias).” Ties oversight directly to automation bias.
  • Santoni de Sio & van den Hoven (2018), “Meaningful Human Control over Autonomous Systems,” Frontiers in Robotics and AI 5:15 — grounds control in tracking (responsive to human reasons) and tracing (a human who understands and is answerable). “Recorded with a human sign-off” is the tracing condition made concrete.

3. Automation bias: why the skill must halt and ask, not decide

  • Parasuraman & Riley (1997), “Humans and Automation: Use, Misuse, Disuse, Abuse,” Human Factors 39(2):230–253 — canonical taxonomy; “misuse” = over-reliance causing monitoring failure and decision bias.
  • Skitka, Mosier & Burdick (1999), Int. J. Human-Computer Studies 51(5):991–1006 — empirical automation bias: people accept incorrect automated advice over available contradictory evidence.
  • Research-specific failure mode — trusting an LLM to decide (vs assist) injects fabricated citations; audits report high fabrication rates in LLM-generated references. (Some figures come from 2026 preprints — treat as illustrative; verify final-venue status before formal use.)

4. Research-integrity frameworks place accountability on the researcher

  • Singapore Statement on Research Integrity (2010) — four principles: honesty, accountability, professionalism, stewardship.
  • ALLEA, European Code of Conduct for Research Integrity (Revised 2023) — the EU’s primary integrity standard; the 2023 revision addresses digital/AI tools and places responsibility for good practice on researchers.

5. “Augment, not replace” — the positive framing

  • Engelbart (1962), “Augmenting Human Intellect: A Conceptual Framework” (SRI AFOSR-3223) — tools amplify human capability within a human-directed system.
  • Bush (1945), “As We May Think,” The Atlantic — the memex aids memory and association, not judgement.

Guardrails that follow (design-enforceable)

  1. Named human sign-off on every material decision (hypothesis verdict, result-is-real, publishable, thesis claim, defensibility). Record the human, not the skill. (Singapore/ALLEA accountability + ICMJE final-approval + meaningful-human-control “tracing”.)
  2. Skill halts at judgement points and asks. The human must be able to disregard the tool. (EU AI Act Art. 14; automation-bias literature.)
  3. Drafts are proposals, never authored text. The researcher reads, verifies, and adopts. (ICMJE/COPE/Science/Nature.)
  4. No unattended end-to-end paper/thesis generation. (Science = plagiarism; ICML = prohibited.)
  5. Verify every AI-surfaced reference before use. (ACL 2023; fabrication rates.)
  6. Disclose and log AI use (the mechanical/mnemonic assistance) — feeds the audit trail the skills already maintain. (COPE/Nature/ICML/NeurIPS.)
  7. Automate only the mechanical and mnemonic — bookkeeping, formatting, retrieval, roll-ups, reminders; never the judgemental.

The bright line (nuance to preserve)

Every policy cited permits AI assistance and forbids AI authorship / decision-making. The principle is that line, not a ban. One caveat: the “editing/polishing without disclosure” carve-out (Nature) sits close to the line — honest-scholar should default to disclosure to stay clearly on the safe side.

Sources

Verification notes

Verified this session: ICMJE, COPE, Nature, Science (DOI), CRediT/NISO, EU AI Act Art. 14, ICML/NeurIPS/ACL, Singapore Statement, ALLEA 2023, Parasuraman & Riley, Skitka et al., Santoni de Sio & van den Hoven, Engelbart. Not directly fetched: Elsevier policy page (corroborated), Bush 1945 (canonical). LLM citation-fabrication figures come partly from 2026 preprints — illustrative; confirm final-venue status before formal citation.