The Human Filter: A 2026 Checklist for AI-Assisted Editing That Preserves Authorial Voice

Artificial intelligence is now a routine part of the UK author's toolkit. It can flag a repeated word, suggest a tighter sentence, and produce a structural summary of a chapter in seconds. None of that is inherently hostile to voice. The risk lies in accepting every suggestion as an improvement. Machine-polished prose often reads cleanly but generically: the edges are rounded, the idiosyncratic rhythm is replaced by a neutral hum, and the resulting manuscript could belong to anyone. The following checklist is designed for authors who want the efficiency of AI-assisted editing without surrendering the qualities that make their writing recognisable.

Begin with a voice sample, not a style guide

Before opening any tool, select two or three passages that you regard as unmistakably yours. They might be a paragraph of dialogue, a descriptive set piece, or a piece of argument in a non-fiction chapter. Save them in a separate document and label them 'voice benchmarks'. Whenever an AI suggestion seems plausible, compare the revised sentence with the benchmark. Does the revision preserve your characteristic sentence length, your preferred level of formality, your recurring imagery? If not, the suggestion may be technically correct but editorially wrong.

Automate the mechanical layer only

Use AI first for tasks where there is a defensible right answer: spelling, punctuation, consistent capitalisation, doubled words, and obvious grammatical slips. UK authors should check that the tool is set to British English, because American defaults will happily replace 'organisation' with 'organization' and 'towards' with 'toward'. Keep a list of house conventions—serial comma or no serial comma, single or double quotation marks, treatment of numbers—and apply them before any stylistic review. The mechanical layer is where automation earns its keep; it is not where voice is made.

Treat structural feedback as questions, not instructions

A tool can identify that a chapter takes too long to reach its central conflict, or that a report buries its recommendation. That is useful. It cannot know why you chose a digression, or whether the delayed reveal is deliberate. Convert every structural suggestion into a question: 'Would the argument be clearer if this section moved earlier?' 'Is the opening anecdote doing enough work to justify its length?' Answer in writing. The written answer becomes part of your decision log and helps you resist the temptation to cut simply because a tool has flagged something.

Review every sentence for cadence and register

Line-level suggestions are the most seductive and the most dangerous. A tool may recommend replacing a passive construction, shortening a long sentence, or removing an adverb. Sometimes it will be right. Often it will flatten a deliberate rhythm. Read each proposed change aloud. If the revised sentence sounds more like a corporate brochure than like you, reject it. Pay particular attention to sentence openings, transitional phrases, and metaphors. These are the places where voice lives.

Brief your editor with a decision log

If you are working with a human editor—and for a manuscript of any consequence, you should be—provide a short decision log alongside the manuscript. Record which AI tools you used, what categories of change you accepted, and which suggestions you overrode. Note any passages where you deliberately retained a construction that a tool flagged. This is not an apology for using AI; it is professional context. An editor who knows that you rejected a hundred 'clarity' suggestions in order to preserve a narrator's idiosyncratic syntax will not waste time reinstating them.

Document what you overrode

Keep a simple table with three columns: the suggestion, your decision, and the reason. 'Changed "utilise" to "use"—register' is a useful entry. So is 'Retained sentence fragment—narrator's voice, deliberate'. The table need not be elaborate. Its purpose is to make your editorial choices visible, especially if you later work with a proofreader, a publisher, or a sensitivity reader. It also protects you from your own inconsistency: if you rejected a rule in chapter two, you should know why you are accepting it in chapter twelve.

Finish with the aloud test and a provenance check

Before you send the manuscript anywhere, read the first page and the last page aloud. Then read a random page from the middle. Listen for anything that sounds borrowed—a phrase that is smoother than your usual prose, a transition that belongs to a different writer. Finally, run a provenance check: confirm that every factual claim, quotation, and permission has a human-verified source. AI can help you edit; it cannot take responsibility for accuracy. That remains yours.

The editorial principle

The goal is not to beat the machine or to surrender to it. It is to use automation where it is strong and human judgement where it is irreplaceable. A clean manuscript that has lost its voice is not a successful edit; it is a different manuscript. UK authors in 2026 can have both efficiency and individuality, provided they keep the final decision about every sentence in human hands.