I Use AI for Every Piece I Publish. Here Is Exactly How the Work Stays Mine.

I publish almost every week in professional technology publications, on LinkedIn, and on my McMillanAI website. I have never been much of a writer. I find the process painful, particularly the mechanics. For more than two decades, I built and deployed technology inside a large financial institution without writing publicly about it. I could have pushed through. I chose not to, and much of what I learned stayed inside the firm.

What changed is that the mechanics stopped being the obstacle. I use AI at every stage of my writing, and I say so. My standard is straightforward: does the piece communicate a true and useful idea clearly? I am responsible for its accuracy and meaning, whether I supplied every sentence or not.

Writing can expose what you do not understand

The effort of drafting often exposes incomplete thinking. A tangled sentence can reveal a gap in your argument. The difficulty of getting the words right can help you figure out what you actually mean.

A model can remove that difficulty without requiring you to understand the subject. The paragraph reads clearly whether or not the thinking behind it is sound. The reader never saw the struggle. The writer loses a useful check.

That is the risk I try to manage. You can publish something that reads well but says little, and the process may not warn you. You have to put that discipline back deliberately.

The pattern

Everything I do follows the same pattern: I originate. The model challenges and proposes. I decide. That sequence repeats throughout a thirteen-step process.

I start by writing out what I want to say, the assumptions behind it, and the arc of the argument. The model challenges and expands that thinking, then reviews it against my editorial rules. I decide what survives. I build the structure, the model critiques it, and I decide what to change. It then drafts from the material we have developed.

I edit that draft on paper, away from the screen, feed my changes back for a content edit, and edit it again myself. I draft in Claude and use Gemini and GPT for additional reviews of readability and factual accuracy. In my experience, GPT is the most useful of the three for catching factual errors. Using different models provides another check, but it does not replace going back to sources. I assess their feedback, apply what is useful, and finish with a proofread.

I own the first and third steps: setting the direction and deciding what is true, what matters, and what belongs. The model contributes in between. That gives readers two questions to ask: what did the writer originate, and what did the writer decide?

A decision means more than approving what appeared on the screen. You should be able to name what you turned down and why. The harder test is whether you rejected something good because it was not what you meant.

For a long piece, this process takes three to four hours of my focused time over several days. It probably takes longer than drafting it myself. Speed is not what I am getting. A LinkedIn post gets a shorter version, sometimes twenty minutes in total: I originate, the model drafts, and I decide what stays. The pattern remains the same. The amount of review should reflect the risk of being wrong.

The step that does the most work

In the challenge stage, I ask the model to play three roles. One challenges my thinking. One expands it. One acts as an editor applying my rules. My instruction is roughly this: act as a coach and editor, push back when warranted, let me own the decisions, and ask when you are unsure.

The challenge is useful even when the model is wrong. Rejecting it forces me to explain why. That explanation is often sharper than my original claim and frequently ends up in the piece. The difficulty has moved from drafting to deciding.

The central argument in this piece came from that exchange. My original claim was that AI makes writing better or worse depending on the effort you put in. The model pointed out that this was true of almost everything and explained little. It proposed a different thesis: writing can expose gaps in understanding, and a model can remove the difficulty without resolving the gaps. I agreed and rebuilt the piece around it. The problem I was circling was mine. The explanation was not. The decision to adopt it was mine.

I also overruled the model. It drafted a sharp section cataloging the weaknesses in my approach. I cut it because it treated my process as a universal prescription rather than a description of what works for me.

One objection deserved an answer: the difficulty of drafting is hard to avoid, while a review process is easy to skip. On a rushed week, I can cut steps, and nothing prevents that. What I preserve is the basic pattern. Even a short post starts with my idea and ends with my decision.

Why I keep a written set of rules

Models often fall into a recognizable style. They reach for a metaphor where a plain fact would do. They state what something is not before saying what it is. They talk around a point rather than naming it.

A draft of my piece on the economics of running AI models included the heading "The decision on your desk." It dramatized a choice without naming it. I replaced it with "Rent the model, or run it yourself." That states the decision in seven words.

My rules address imprecision, regardless of who produced it. They are linked below.

Where this breaks

The decision step fails when I accept a sentence without examining it.

A model can produce a fluent, well-structured sentence that is completely wrong. A confident tone is not evidence of accuracy. My review works only if I make a judgment rather than accept what appears.

In a piece I wrote about Nvidia, a draft stated that the company had tried to buy Hugging Face for $500 million. The reporting described something different: a proposed $500 million investment at a $7 billion valuation, which Hugging Face rejected.

I caught the error by returning to the reporting and comparing it with the draft. The shorter version read well, but it changed the transaction. Reading for style alone would have taken me past it.

Revision creates a different risk: a series of reasonable edits can leave the argument inconsistent. In this draft, I wrote that a model would never tell me to abandon a piece, shortly after describing how one had dismantled my central claim. I had blurred the distinction between challenging an argument and judging whether a piece should run. Reading the whole draft after each round helps catch those shifts.

Useful criticism is not the same as independent editorial judgment. A model has no reputation at stake in what I publish. A human editor is accountable for deciding whether a piece deserves to run, not just how to improve it. That is one reason my important work still goes to human editors. With the mechanics largely handled, they can give more attention to whether the argument holds up.

Much of what I publish does not receive that outside check. Two imperfect safeguards help. I start with subjects grounded in work I spent twenty years doing, so I have a basis for judging the idea before a model gets involved. I also sleep on drafts I am unsure about. Something that looked finished at eleven at night can read differently in the morning. Distance does not replace an outside opinion, but it can make weaknesses easier to see.

The same brief, run without me

While working on this piece, I gave the same opening brief to a model and asked it to write the essay on its own, in one pass. It produced a readable piece and reached roughly my conclusion: use the model to think, rather than simply to generate text.

The differences were in the details. Its version named no publication, company, or person. It described my process as twelve steps rather than thirteen. It quoted no sentence from a real draft. It also used em dashes, which my editorial rules prohibit.

Specifics matter because they give readers something to check. The factual error I caught, the sentence I cut, and the hours the process took all came from my experience. The model could not supply details I had not given it. The comparison showed what my continued input and review contributed.

The more serious problem came at the end. Its version described how it had argued with a model and rewritten its central argument three times. None of that happened. It had been generated in a single pass.

The distinction was not prose quality. Both versions read well. They differed in whether they described real work and told the truth about how they were made.

Why I publish the transcript

My personal ethics policy commits me to disclosing how I use AI. In practice, that means giving readers enough information to judge the model's contribution for themselves.

A percentage would imply precision I cannot defend. After repeated revisions, most sentences have passed between me and the model several times. No single number captures that honestly.

The full transcript for this piece is linked below. It shows the decisions: what the model offered, what I accepted, what I rejected, and why.

I expect this kind of disclosure to become more common among people whose credibility depends on published work.

This is my approach

This process works for me, someone who never built a career as a writer. Other people will choose differently, and they should. Some will conclude I am not doing the real work. Others will conclude the opposite. I am content to leave that judgment to them.

My last ten articles are linked below. Read them and decide whether the ideas were worth your time.

A year ago, there would have been nothing to read.