AI can now do far more than correct grammar. It can propose an argument, imitate a register, reorganize a chapter, supply transitions, and produce a conclusion that sounds as though it had been waiting inside the draft all along. Calling this “assistance” can understate what is happening.
Yet even when a machine contributes most of the visible sentences, there remains a last part that cannot be measured by word count. Call it the last five percent—not because the human always writes five percent, but because every finished text reaches a point where someone must decide: Yes, this is what I am saying.
Language can close too early
Writing is often a way of discovering that our first thought was not yet true enough. A sentence resists us. An example refuses to fit. A paragraph we loved turns out to depend on an assumption we do not accept. The difficulty is not noise around writing; sometimes it is the work itself.
AI is extraordinarily good at relieving this difficulty. It can finish the sentence, bridge the gap, and make two incompatible sections appear continuous. Sometimes that reveals the missing thought. Sometimes it conceals the fact that there is no bridge.
The danger is premature closure. Because the language sounds complete, the writer stops returning to the question. The text has an ending, but the inquiry ended somewhere else.
The last mile is not proofreading
Human review is often described as quality control: verify facts, remove clichés, correct tone. All of that matters. But the last mile begins after a draft has passed those tests.
Read the clean version and ask: Where does this text become most certain? Have I earned that certainty? Which paragraph protects me from having to admit something? What would a serious reader be right to resist? If the piece changes someone’s view or action, which change am I actually willing to have caused?
These are authorial questions. They do not ask whether the prose works. They ask what the prose is working for.
Originality is not an absence of influence
Some writers respond to AI by defending a fantasy of purity: every valuable phrase must emerge untouched from an isolated mind. But human writing has never worked that way. We inherit genres, rhythms, stories, arguments, and the remembered voices of other people. Originality does not mean having no sources.
It means that what has been received passes through a life capable of selection and refusal. An author does not merely combine influences; an author is answerable for the combination. This is why a heavily edited book can remain someone’s work, and why a text typed entirely by one person can still feel like no one is there.
AI does not abolish authorship. It makes the question of authorship harder to avoid. When generation is cheap, the scarce act is no longer producing acceptable sentences. It is deciding which possible sentence deserves to remain—and what the decision commits you to.
A practical last-five-percent pass
Once the draft is polished, make one pass that is not about polish.
- Mark the sentence that contains the piece’s real claim. Rewrite it without looking at the draft.
- Find the paragraph the model could have produced for almost anyone. Add the detail that makes the situation yours, or delete it.
- Identify the strongest objection. Do not merely “address” it; allow it to change the structure if it should.
- Read every “therefore,” “clearly,” and “must.” Decide who granted the transition.
- Keep one rough edge if smoothing it would erase an honest uncertainty.
This pass may alter only a few lines. It may also destroy the entire draft. Its value is not proportional to the number of edits. The point is to make the finished form pass once through the person whose name it carries.
Sometimes the right act is disclosure
Different forms of writing require different transparency. A routine internal update does not need a memoir about its production. Research, journalism, teaching, evaluation, and personal testimony may demand much more. The relevant question is not whether AI touched the text, but whether concealing its role would mislead the reader about evidence, expertise, experience, or responsibility.
Disclosure cannot solve everything. “Written with AI” is too broad to explain whether the system corrected commas or invented the argument. But the need to describe the collaboration can itself clarify where the writer still stands.
Sign the meaning, not the keystrokes
The last five percent is not a defense of inefficient labor. Let the machine remove repetition. Let it find ten structures while you would have found two. Let it show you a sentence better than the one you wrote.
Then ask whether you can sign not merely the words, but their direction. Can you explain why this claim is here? Can you hear the person the framing excludes? Can you defend the omission? Can you revise the piece if its effect reveals something you failed to see?
A signature does not certify that every word was generated by hand. It says that when language became action, someone was still there.