The complaint I hear most about AI-assisted writing is that everything comes out sounding the same.
It does, and the reason is straightforward. If you ask a model to write the article, you get the average of every article ever written on that subject, which is by definition characterless.
But that is a description of one way of using it, not of the tool.
The version that works, and the version I use daily, treats it as a research assistant and a structural editor, and keeps the actual sentences human.
Here is the workflow, and the specific line I do not cross.
Quick Info
The Dividing Line
Here is the rule I work to, stated as plainly as I can.
Delegate the thinking about the content. Never delegate the sentences.
Research, structure, gap analysis, objection finding, fact organising, outline pressure-testing. All delegable, all improved by delegation.
Phrasing, rhythm, opinion, humour, anecdote, the specific way you order a sentence. Not delegable, because that is what makes it yours and it is also the only part a reader remembers.
Everything below is an application of that one line.
Step One: Interrogate the Topic Before Writing
This is the highest-value use and almost nobody does it.
Before writing anything, I ask a series of questions designed to find what I do not know I am missing.
Half an hour of this produces a better article than three hours of writing from what I already had in my head. Not because the answers are authoritative, but because they surface the gaps I would otherwise have shipped.
Step Two: Structure, Then Argue With It
I ask for an outline, then I disagree with it, which is the productive part.
The first outline is always the conventional one, which tells you exactly what everyone else's article looks like. That is useful negative information.
So I ask what an unconventional structure for this would be, what could be cut, and what the article would look like if the third section were the opening. Then I build my own from the pieces that survive.
The outcome is a structure I chose rather than inherited, which is most of what makes an article feel like it has a point of view.
Step Three: Write It Yourself, Properly
This is the step people try to skip and it is the one that produces the voice.
Write from the outline, in your own words, at speed, without going back to the assistant. Bad sentences are fine at this stage. Generic sentences are not, because they are the ones that will survive editing unnoticed.
Two things that help keep it yours.
Step Four: Use It as an Editor, Not a Rewriter
Once the draft exists, the assistant becomes useful again, with careful instructions.
What I ask for:
What I never ask for: rewrite this to be more engaging. That instruction produces the exact flattened prose everyone complains about, because engaging to a model means conventional.
Diagnose with the assistant. Fix by hand.
Fact-Checking Is Not Optional
Every statistic, name, date and citation that comes out of an assistant has to be verified against a primary source before publication. Every one.
Confident, plausible, entirely invented references remain a normal failure mode, and a fabricated statistic in your article is your error, not the tool's.
My rule is that a number without a link to its source does not go in. It has cost me a few good-sounding statistics and it has never cost me a correction.
More on this in my fact-checking article.
Teaching It Your Voice, and Its Limits
You can improve how well an assistant matches your register by giving it examples of your writing and a description of your rules. It genuinely helps for short pieces.
What I keep in a saved instruction: British spelling, no em dashes, first person, short paragraphs, no metaphors about journeys, never open with the word imagine.
But be clear about what this achieves. It gets you a closer imitation of your surface habits. It does not get you your judgement, your experience or your opinions, and those are the parts readers come back for.
Which is why the workflow above keeps the writing human even when the imitation is decent. The imitation is good enough to fool a skim and not good enough to be worth reading twice.
Frequently Asked Questions
Google's guidance targets unhelpful, low-quality content regardless of how it was produced. Content researched with AI and written and reviewed by a knowledgeable human is not the problem.
The sentences, the opinions, the anecdotes and the final fact-check. Those are where the voice and the accountability live.
Finding what your draft is missing. Sub-questions, objections, gaps in the argument and unclear passages.
You can get a reasonable imitation of your surface style with examples and explicit rules. It will not reproduce your judgement or experience.
Follow every number to a primary source before publishing. Fabricated but plausible citations remain common, and the error becomes yours on publication.
Yes, but mostly in research and revision rather than drafting. The drafting stays roughly as slow as it always was, which is the point.
Before You Go
The short version: use it to work out what to say, then say it yourself.
The prompts I use for the interrogation and editing steps are in my resources section, along with the saved instruction I keep for voice.
If you want a content process built that keeps your voice intact, get in touch.
Delegate the thinking, keep the sentences, Tariq
I'm a marketing consultant, entrepreneur and content creator. I help businesses grow through practical marketing, websites, SEO, content and AI.
More About Tariq →