Iterative Scholarly Writing Partner

Workbook-grounded reconstruction. Academics repeatedly develop journal articles, chapters, proposals and reports through cycles of drafting and critique. Human editorial support is valuable but may not be available at the moment when a draft is taking shape, while a general-purpose text generator can too readily substitute its own argument or voice.

Proposed. Scholarly writing improves through iteration, yet sustained critical attention is difficult to maintain across structure, argument, evidence, genre and prose. When revision is compressed into a single request for improvement, plausible rewrites can conceal unsupported claims, flatten authorial voice and make it difficult to see which changes respond to which problem.

Proposed. Treat GenAI as an iterative scholarly writing partner whose role changes across clearly separated passes. First establish the purpose, audience, genre, evidence base and aspects of the text that must remain stable; then request diagnosis before revision, select which suggestions to accept, and verify the resulting claims and citations. Keep the author in control of the argument, interpretation and final language, and preserve a visible route from the source draft through critique to approved revision.

• A prioritised diagnosis of structural, argumentative and stylistic issues.
• A revised draft whose material changes can be traced to an author-approved purpose.
• Preserved authorial control over claims, interpretation, emphasis and voice.
• An explicit record of unresolved evidence, citation and genre questions requiring human judgement.

The author defines the contribution and voice, selects the scope of critique, accepts or rejects changes, checks every evidential claim and approves the final text. Editors, co-authors or supervisors may supply additional criteria. The system diagnoses, proposes and transforms; it does not determine scholarly validity or authorship.

Reconstructed risks

• The system rewrites rather than critiques, producing a polished text that no longer expresses the author’s reasoning.
• Fabricated references or evidential claims enter the draft during revision.
• Generic academic phrasing displaces disciplinary precision or a distinctive voice.
• Successive passes create circular revision, unnecessary homogenisation or loss of version history.
• Confidential manuscripts, reviewer material or participant information are disclosed to an unsuitable service.

• Minimum viable implementation: follow the documented four-stage procedure manually and retain a prioritised diagnosis of structural, argumentative and stylistic issues, together with review decisions.
• Robust implementation: add structured templates, version history, provenance, access controls, exception handling and explicit quality gates suited to iterative scholarly writing partner.
• Low-code implementation: connect approved services only where inputs and outputs remain inspectable, with a human approval gate before consequential use or publication.
• Local or privacy-preserving implementation: keep sensitive or unpublished artefacts in controlled storage and prefer local processing where policy, risk or participant expectations warrant it.
• Speculative implementation: explore additional automation only within constrained permissions, with logging, stop conditions, rollback and retained human authority.