Key takeaways
AI can move people into reviewing work after important decisions have already been made. Preserving a human influence takes more than a final approval.
My articles, podcast, and latest book involve substantial AI-generated material, along with repeated human decisions about what the work should say and whether it succeeds.
The useful efficiency gain is having more capacity to develop worthwhile work while retaining the time and authority to question it. Oh, and I spend more time with the people in my life as a result of the reduced hours spent working.
If you’re not careful, you can ruin a board with a miter saw real quick.
Setting the wrong angle can let you make a beautifully clean cut that leaves you with two pieces that won’t fit together. Sure, the motor has saved you effort, but whether it’s helped your project is another matter.
I’ve been thinking about that distinction a lot as I work on this blog, my podcast, and my forthcoming book, The Cosmically Conscious Corporation. Generative AI plays a substantial role in all three. It helps me develop ideas, locate research, build structures, draft prose, and revise it. Sometimes it supplies much of the language in a first draft.
I understand why that last sentence might change how someone sees the work. If a machine produced the initial paragraphs, what, exactly, did the writer do?
It’s a reasonable question, but the answer requires looking at more than the paragraphs.
A draft that looks finished can settle things too soon
In “An AI-Powered Centaur Makes This Podcast,” I explored the distinction between centaurism and reverse centaurism: an arrangement where a person directs a machine’s capabilities, and one where the person increasingly serves a process the machine directs.
A human can remain involved throughout either arrangement. What matters is the influence that person can exercise.
The question of what a human’s actual contribution looks like becomes harder when the machine can produce an entire draft before one has worked through the problem. This can lead to someone evaluating an argument whose structure they haven’t really considered while looking at something that already appears assembled.
There’s research behind this concern. Jana Retkowsky, Ella Hafermalz, and Marleen Huysman studied creative professionals using Midjourney in advertising. Their work describes process collapse: activities such as generating ideas, visualizing them, and developing a finished result becoming compressed into an interaction with AI. A university account of the year-long study at RTL Nederland notes that colleagues sometimes entered later, leaving less opportunity for shared discussion and course correction.
The researchers also describe early aesthetic closure. An image can look so accomplished that it anchors expectations and constrains further development of the idea.
In that arrangement, one creative professional may be working with AI from the beginning while other people lose opportunities to shape the project. Asking whether a human was involved misses whose judgment entered when.
I see a version of that risk in writing. A coherent opening, a few confident headings, and a satisfying conclusion can make an argument seem settled while its assumptions are still up for grabs. Editing what’s there can feel easier than asking whether those sections belong in the article at all.
That doesn’t make a complete first draft inherently bad, but it does make its apparent completeness something I have to resist in my own work.
Before the draft, and after it
My process often starts in conversation.
I bring a question, an observation, something I’ve read, or a frustration I haven’t figured out how to articulate. Sometimes I talk it through aloud. The conversation wanders; I qualify something, push back on an interpretation, realize two things are connected, or discover that the question I started with isn’t quite the one I want to pursue.
ChatGPT contributes to that development. It proposes connections and counterarguments. It can find a structure in a mess of thoughts, and sometimes its suggestion encourages me to reconsider. Its influence begins before prose generation, yes, which is one reason describing it as a glorified spellchecker would be dishonest.
Eventually, though, we have enough direction to draft. For this article, I proposed the woodworking metaphor and asked ChatGPT to draw on our actual conversations and revision discussions about the Cosmically Conscious Corporation book. The generated text starts from material and an assignment I’ve supplied, and it contributes its own organization and language.
I’ve developed separate instructions for articles and podcast scripts because reading and listening place different demands on an audience. Those instructions include developmental editing, prose refinement, and line editing. They ask the system to test the reasoning, identify repetition, preserve useful qualifications, and make one section lead into the next. The voice guidance draws on my own writing, although reproducing familiar rhythms doesn’t establish that a passage says something I mean.
These are passes performed by AI. Calling one a developmental edit doesn’t turn it into an independent human editor’s assessment, and a message announcing that the checks are complete doesn’t establish that the writing works.
I still read it.
I notice where my attention drifts. I ask why a section needs another example, what a pronoun refers to, or why a chapter promises an answer it hasn’t supplied. Sometimes the problem is a sentence. Sometimes it’s a whole stretch of argument that needs to be reconsidered.
The instructions I provide have themselves changed in response. When line edits to parts of the book made the reading experience worse, I asked for stronger checks on transitions and progression, including comparison with the preceding draft. The label line edited had been less reassuring than it sounded, and was ultimately abandoned after even a second AI-guided pass failed to pass muster.
As a result, the division of labor has changed. In my writing sessions, I now spend less time constructing a first version sentence by sentence and more time developing, diagnosing, and revising it. That redistribution is part of what makes the method useful to me.
That said, I don’t replace every generated sentence to demonstrate that I’ve touched it. A paragraph that expresses the point well may survive. Another may need to go even if every sentence is individually competent. If I spent my time rewording everything for the sake of rewording everything, I’d be keeping myself busy, yes, but that wouldn’t mean I’d made the piece better.
It’s also worth noting that I came to an understanding of how to use these tools after years of writing and revising books and working with advanced AI workflows in my day job as a software engineer. Those experiences are part of how I judge the output. Both give me something to test a proposed fix against besides the system’s own assurance that it’s an improvement.
The chapter that couldn’t tell its own story
One example came during revisions to Chapter 13 of The Cosmically Conscious Corporation.
The book follows three fictional robotics companies with different guiding philosophies. In this particular chapter, an employee at one company questions why a small wording change to maintenance instructions needs approval from a full committee.
As a result, the company tests a smaller review process. This then guides the reader through an easy-to-follow process: a concern is raised, a proposed remedy proposed, a test reveals a weakness, and a revision is made.
The AI-generated version I was reading, however, made the reader work backward. We first encountered the resulting arrangement before the initial proposal and the mistake that explained the change.
My response in the revision discussion was fairly direct: “Why are we jumping around in the chronology?”
I asked for the initial proposal first, followed by the mistake it permitted and the response to what the company had learned. We revised the chapter to follow that progression.
This wasn’t a preference about where to put an adjective. The order affected whether readers could understand how the company remained open to revising its own remedy. The chapter’s example needed to demonstrate the philosophy the book was asking people to consider.
There was a related problem in how we described the employee’s contribution. Earlier wording treated it as a complaint. I questioned that characterization: the employee was offering an observation or suggestion about a process that wasn’t serving its purpose particularly well.
The distinction mattered in a book whose principles include “Difference is information, not merely friction.” Calling the contribution a complaint risked teaching one attitude while the surrounding argument advocated another.
AI helped generate and revise that chapter, sure, but it also produced deeper structural content I had to challenge, and some corrections took more than one exchange. The help from AI was substantial, then, but so was the need to keep reading and revising critically.
The podcast changes the work again
Turning an article into a podcast gives me another opportunity to find problems I missed on the page.
An argument that works on paper (so to speak) may become difficult to follow when someone can’t pause to reread a paragraph. A heading might separate two sections visually without explaining how the thought moves between them. A sentence can be perfectly grammatical and still be a nuisance to say aloud.
The script workflow therefore asks for more than a reading of the article. It includes adaptation for listening, meaningful transitions, and attention to pacing.
None of this guarantees that the finished episode does what I want it to do.
Music makes this especially apparent. AI can suggest backing tracks and where to place them, but I have plenty of times rejected suggestions because a track felt too optimistic for the passage beneath it. A pleasant piece of music can be the wrong piece of music, just as an elegant sentence can make the wrong argument.
In either case, I still have to decide whether the elements belong together and listen to what happens when they do.
This is where the efficiency becomes useful to me. I can move between an idea, an article, an adaptation, and an audio production without rebuilding every instruction and production convention from scratch. More of my limited creative time can go toward developing the material and evaluating the result.
Oh, and with the time saved, I can also spend more hours with my wife, our pets, our friends, and myself.
Despite all of these guardrails, some passes still create work I don’t want. Those costs matter when I judge whether the process helps. So does the value of getting a substantial possibility in front of me that I can test and reshape.
Who gets to change the plan?
The miter-saw comparison helps explain part of this.
A loose blade, a working saw, a person learning basic cuts, and an experienced woodworker all tell us something different about what’s possible. Looking only at the cutting mechanism misses the surrounding knowledge that makes a finished project useful.
Learning angled cuts expands what someone can build, just like learning about materials, proportion, assembly, and finish. There are many levels of competence between operating a tool and practicing a craft, and that’s the distinction I want to make visible in AI use.
The analogy has an important limit, though: a saw doesn’t normally propose the design of your furniture. AI can propose the shape of the article, the argument it makes, and the criteria by which you might judge it. Its suggestions can set the direction while appearing merely helpful.
That means my ability to reject a draft has to extend beyond rejecting its wording. I need to be able to question its premise, change its structure, and abandon the project if it no longer seems worth doing.
The same question matters at a different scale inside the companies building these systems. In a September disclosure, Anthropic rated 26 percent of its AI research and development as AI-led as of August 2026. Its definition still included human supervision; the company reported no fully autonomous work in the measured categories, and I should note these are the company’s own measurements, with methodological limitations, rather than an independent verdict.
That distinction leaves me asking what human supervision actually permits. Someone may be empowered to investigate and redirect the work. Someone else may be expected to sign off on more output than they can meaningfully examine.
I control my publication’s process in ways an employee working under an imposed production target may not. My experience is a case for what a person can do with that freedom, with limited relevance to a workplace that withholds it.
Nor does showing care in my writing settle objections about training data, compensation, environmental costs, or the institutions deploying AI. Those objections can remain even when a creator exercises care; no woodworking metaphor can answer them.
What I do hope the metaphor complicates is the assumption that generated language tells us everything about the process that produced a piece.
Room to change the work
Some people want to make every cut by hand. There can be satisfaction, learning, and meaning in doing that. They don’t need my permission, and I don’t think they owe anyone an efficiency upgrade.
For my part, I want to explore more questions and bring more of them into work I can share. AI helps me do that, including by making contributions I wouldn’t necessarily have produced independently.
The risk is allowing that capacity to become a demand: more drafts, more episodes, more material to approve because the system can keep producing it. If review becomes a hurried formality, the tool has changed the terms of the work.
In my own process, I haven’t waited until the draft exists to become involved. I return to the work in different ways, including after AI has done substantial drafting. Other people may enter at that later point for the first time. They need a real opportunity to change the work too, before its consequential decisions become untouchable.
In Chapter 13 of The Cosmically Conscious Corporation, I needed to take apart a passage and put the events back in an order that let readers understand them. Whatever time the process saved elsewhere, it still needed to make room for that work.
That’s the efficiency I want: enough help making the pieces that I can give proper attention to how they fit.
Tell me in the comments
Which parts of your creative process would you want to keep doing yourself, even if a tool could do them faster?
When you’ve used or reviewed AI-generated work, what has helped you meaningfully change its direction?




Setting the wrong angle on a miter saw and getting two beautiful pieces that won't fit together is such a good picture of a finished looking draft. Looking busy is not the same thing as thinking, and a coherent opening with a few confident headings can look very busy. Your point about whose judgment entered when is the one I'd want every team using AI to argue about. Spending more time with your people because of the saved hours is the best return on AI I've heard of lol