I use AI regularly for prompting and research. I commissioned a review of selected research because I wanted clearer guidance for authors: where can these tools help, what needs checking, and how can we explain their contribution to a book?
The useful questions are practical. What did you ask the tool to do? What did it supply? What did you accept, change or reject? And what obligations apply when you submit or publish the result?
Research gives us a better basis for those decisions. It also shows why a confident answer, a tidy reference list or a detector score needs closer examination.
Asking for questions about a scene, finding research leads, correcting punctuation and generating a chapter involve different contributions. Calling all of them “AI-written” tells us very little about the process.
Possible uses include testing whether a character’s motivation is clear, comparing chapter structures or looking for a continuity problem. Assess the suggestion against your intention. Does it preserve the character, rhythm, cultural detail and meaning? You make the decision to accept, adapt or reject it.
Doshi and Hauser’s short-story experiment found that access to AI story ideas improved readers’ ratings of the resulting stories. It also made the AI-assisted stories more similar to one another. Both findings matter. Read the study and author manuscript.
This was a short writing task, not a test of complete novels or years of professional practice. It supports a specific concern about variety, rather than a claim that AI inevitably destroys an author’s voice.
For your book: examine a revision’s effect on your meaning and style, not just whether it sounds smoother. A deliberate awkwardness, uncertainty or unusual phrase may be doing useful work.
A study published in April 2026 examined twelve AI-generated mental-health literature reviews. All citations could be traced to identifiable publications. Yet only 145 of the 333 citation-claims were judged accurate; the others contained varying degrees of error or unsupported material. Some bibliographic details were also wrong. Read Linardon and colleagues’ study.
The researchers tested the model they described as ChatGPT-5 thinking, using one prompting approach in December 2025. Where a passage made several claims, its lowest claim score was retained. This is not an error rate for every AI tool or for authors’ books. It demonstrates a narrower point: finding a real publication does not establish that the answer represents it accurately.
For your book: open the original source behind an important factual statement. Check the passage, its context and its qualifications. Keep a note of where the support appears. For fiction, that might be a historical detail; for non-fiction, it may be evidence central to your argument.
In a July 2026 evaluation, Epoch AI tested three detectors on 495 human-written passages. Pangram and GPTZero recorded no false positives in that sample. Originality.ai incorrectly flagged 19 passages. All three performed strongly on AI text generated from basic prompts, but missed between 30 and 53 of 297 passages generated to imitate an author’s style. Read Jaeho Lee’s evaluation and methods.
These were roughly 500-word passages, with specified detector versions and settings. Zero errors in one sample does not establish zero risk elsewhere. Nor do these tests tell us how often publishers rely on detectors or penalise authors.
For your book: treat a result as information to examine alongside the relevant policy, your drafts and an account of the writing process. If your work is questioned, ask which rule is at issue and what evidence supports the assessment. Honest compliance is the aim.
Before submitting, read the agency, publisher or platform policy and your agreement. Ask for clarification where the wording does not clearly cover your intended use. Describe that use accurately.
Keep your starting drafts, significant revisions, source notes and a brief account of material AI assistance. Before uploading someone else’s unpublished work or sensitive information, establish your authority and understand the service’s confidentiality and data-use terms.
Lewis Silkin’s October 2026 briefing recommends meaningful human authorship, documenting the creative process and checking AI outputs for third-party IP infringement. These are useful risk-management principles; records alone do not guarantee copyright or acceptance. The application of legal rules to a particular book needs case-specific assessment. Read the Lewis Silkin briefing.
This is a proposed editorial checklist, not a workflow whose effectiveness has been experimentally established. Evidence on sustained book writing and long-term skills remains limited. We can acknowledge those gaps and still make careful decisions now.
Use assistance deliberately. Keep your judgement active. Make your account of the work accurate and your evidence traceable.
For the way we apply those principles in our own work, read How Casover Works With AI. You can also discuss questions about your writing process in the free Casover author community.
Read the fuller evidence behind this Insight: a 41-page companion covering author craft, factual reliability, detector findings and publishing decisions, with source links and clear limits.
Download AI and Authors: Research and Evidence (PDF, 41 pages, 382 KB)
Research edition 0.2, 6 October 2026. Selected-source checks and unresolved questions are identified in the document.
Commissioned by Steve Castledine, Draft2Book LLP. Research discovery used Perplexity Deep Research; selected original-source checks, synthesis and preparation used OpenAI Codex under Steve’s direction. The four sources above were revisited on 6 October 2026. This article offers evidence-informed editorial guidance, rather than a legal determination or a systematic review of all research. No endorsement by the cited organisations is implied.