How should cognitive labor delegated to AI be governed in writing and education?

The unit of governance is the delegated task

The paper's central argument is that "AI use" is too coarse a category to regulate. Writing is treated as a family of task–tool–context combinations, and three distinctions do the analytical work. First, what cognitive labor is delegated: suggesting wording, fixing grammar, or translating reshapes expression, whereas summarizing literature the author has not read, drafting arguments, or generating references reshapes content and reasoning. Second, how the writer engages with output: in critical collaboration the author iteratively prompts, selects, verifies, and revises while retaining control of claims and evidence; in uncritical substitution the output stands in for the author's own reasoning. Third, what is at stake in the setting — polishing a self-written Methods section or a student draft generally carries less risk than interpreting results or completing an assignment meant to test independent reasoning.

The same nominal use can therefore sit on either side of a governance line. The paper uses this to explain why duties and prohibitions alone are insufficient: without criteria tied to delegated labor, engagement, and stakes, institutions can over-regulate visible low-risk practices while neglecting higher-stakes problems that remain opaque.

Eleven policy families reviewed through August 2026

The empirical core is a policy mapping, not a survey or experiment. The author reviewed policies current through 11 August 2026 from three international editorial organizations and eight journal or publisher families selected for cross-title reach or influence, extending an earlier author/reviewer comparison to editors. Tables summarize author provisions and reviewer/editor provisions separately, and the supplementary material reproduces supporting policy passages verbatim, with provision-level mapping and evidence identifiers linking each summary to its source text.

The resulting dataset is described as 296 provisions and 259 evidence units. Because the unit is the policy provision rather than the individual journal or author, counts below should be read as counts of policy families or provisions, not of researchers or manuscripts. The paper also notes that university and classroom policies are more heterogeneous and often internal, which is why publishing rules serve as a tractable foothold rather than a complete picture of governance.

Policies converged on disclosure and human accountability

Early policies emphasized human-only authorship, human accountability, and disclosure, but varied in whether and how they restricted generated content. By August 2026, the pattern had shifted toward conditional permission differentiated by task and role: seven of the 11 policy families distinguished assistive, substantive (or generative), and methodological author uses, and 10 of the 11 included provisions for both reviewers and editors.

The three tiers carry different obligations. Assistive use — spelling, grammar, minor stylistic editing — is often exempt from disclosure. Substantive use, including outlining, drafting, substantial rewriting, literature summarization, and argument development, is generally permitted with disclosure, human review, and verification. Methodological use, where AI becomes part of the research method such as data classification or text mining, should be reported in Methods in enough detail to assess validity and reproducibility. Disclosure itself became more structured: authors are typically asked whether and how tools were used, where to report it (cover letter, Methods, Acknowledgements, or a dedicated statement), and sometimes what to supply, such as tool name, model version, dates, and prompts.

Confidentiality splits author and reviewer rules

Reviewer and editor provisions diverge sharply from author provisions. Of the ten policy families that gave guidance for reviewers, all restricted uploading manuscript-related content to AI tools, principally to protect confidentiality and authors' rights. Eight explicitly permitted bounded assistance under safeguards, typically for revising review language, organizing feedback, or non-confidential background searches.

Editorial judgment was guarded more tightly. Six of the 11 policy families expressly reserved editorial decisions or evaluative work for humans, or prohibited delegating core review tasks; four nevertheless allowed or reported internal use of AI-assisted manuscript screening, workflow checks, or reviewer identification and matching. The paper singles out Nature Portfolio's adoption, by 4 August 2026, of a cross-role framework organized by risk rather than by author task: assistive uses that introduce no new intellectual content and do not influence judgment are permitted; interpretive or evaluative uses require verification, disclosure, and demonstrable human oversight; opaque or unverifiable uses, uses that replace accountable human contribution, or those compromising confidentiality or integrity are prohibited.

Five domains link risks to safeguards

The policy map shows which risks current rules address, but the author argues it cannot by itself specify responsible oversight. The paper therefore interprets the patterns alongside scholarship on writing and major AI-governance frameworks, particularly the NIST AI Risk Management Framework and the EU High-Level Expert Group Ethics Guidelines, and groups consequences into five overlapping domains. Agency and development asks who performs the cognitive work and how that division affects skill and scholarly identity. Epistemic quality asks whether claims are accurate, evidence-supported, correctly sourced, and coherently argued. Fairness asks who has access to effective tools, whose work is judged credible, and who bears the costs of error and correction. Rights and governance asks whose material is used and under what authority, how consent, confidentiality, and intellectual property are protected, and how accountability is assigned. Economy and culture asks how cheap fluent text changes incentives, labor, infrastructure, and the informational ecology of scholarly communication. Transparency is not a separate domain but a set of cross-cutting mechanisms — disclosure, logging, and auditability — operating across all five.

The author is explicit that this framework is an interpretive synthesis rather than an empirically validated taxonomy, and that other partitions are possible, such as separating privacy from accountability or labor-market effects from epistemic ecology. That status matters for readers: the five domains organize risks and connect them to oversight mechanisms available to journals, universities, and supervisors, but they are not a measured classification of observed harms.

Development, equity, and incentives remain under-addressed

The mapping identifies a gap between what policies emphasize and what they leave thin. Current policies concentrate on transparency, authorship, verification, and confidentiality, but give much less attention to the developmental consequences of delegating cognitive labor, the distributional effects of unequal AI access and error burdens, and the long-run incentive shifts produced by cheap fluent text. On agency and development, policies speak mainly through authorship criteria and accountability and are increasingly explicit that substantive AI use must be disclosed, yet say little about process: how work should be evaluated in training contexts, where legitimate assistance ends and misrepresentation begins, or how to preserve room for slow acquisition of voice and style.

On epistemic quality, policies insist that authors verify AI-generated content and remain accountable for accuracy, but rarely specify how verification should be organized in actual workflows or how journals should respond when AI-induced errors are discovered. The paper names a structural problem it calls cheap fluency: when text becomes cheap to produce but expensive to verify, the costs of vigilance are pushed downstream to supervisors, reviewers, and readers, who may overweight polish relative to depth under time pressure. On rights and governance, policies are described as weaker on technical infrastructure such as secure deployments, logging, and auditability, and on legal-procedural frameworks such as redress mechanisms, training-data disputes, copyright over AI-assisted outputs, and persona impersonation in academic contexts.

Safeguards should follow the tier of delegation

The proposed safeguards are tiered rather than uniform. For agency and development, the paper suggests retaining authorship criteria and AI-use disclosure but supplementing them with short standardized statements describing how AI contributed; in teaching and supervision, evaluation should include draft histories, notes, oral defenses, and some AI-free in-class writing, with students trained to critique, debug, and improve model outputs. For research manuscripts, core claims, key inferential steps, and stated limitations should originate in observable human drafts such as tracked changes or annotated outlines, so the argumentative spine remains demonstrably the author's.

For epistemic quality, AI output should be treated as provisional: claims, data, and citations independently verified by human authors, with claim- and citation-checking built into standard visible workflows rather than left to ad hoc individual diligence, and unverified AI-generated citations or references prohibited. For rights and governance, the paper proposes consolidating existing COPE/ICMJE-style taxonomies into clear tiered categories — permitted uses exempt from disclosure, uses requiring disclosure, and prohibited uses — alongside institution-controlled models for sensitive work, data minimization, retention policies, security audits, red-teaming, and logging sufficient to support post-hoc reconstruction. In medical education specifically, the framework requires protected patient information and restricted assessment materials to enter only approved systems with defined retention safeguards, and proposes sequencing: establish foundational competence, then teach verification, then assess accountable human–AI collaboration.

Judge governance by reliability, equity, and efficiency

The paper's conclusion is conditional rather than a verdict on AI. AI assistance is described as neither inherently corrupting nor inherently emancipatory: it degrades scholarship when used as uncritical substitution, but can improve on the pre-AI baseline when tiered safeguards and clear role boundaries lower linguistic barriers, shift effort from surface polish to verification and synthesis, and make shallow or generic prose less valuable. The author states that some boundaries should not move absent a radical redefinition of authorship and accountability: human-only authorship and independent peer review, confidentiality safeguards against public or training-enabled tools, and human editorial judgment even where AI supports screening.

Within those constraints, the paper's decision rule is that governance should be judged by whether it improves the reliability, equity, and efficiency of scholarly communication relative to the pre-AI status quo, not by how much it suppresses AI involvement as such. The research agenda follows from the same logic: studies should distinguish tiers of use rather than treating AI use as a single category, measure equity consequences across languages, institutions, disciplines, and regions rather than assume them, and track how governance choices such as blanket bans, role-specific policies, and secure institutional tools affect reviewer burden, time to decision, and the signal-to-noise ratio in submissions. The closing formulation sets the threshold: under conditions of cheap fluency and expensive epistemics, scholarly writing is worth producing only when its claims are worth the cost of checking them.

Sources used in this answer