Get your fine, handcrafted artisanal essays right here

I use AI a lot in the writing process—but not for drafting

I don’t use AI to draft my writing. These are fine handcrafted artisanal essays, every word placed individually with loving care.

I use AI a lot for researching, occasionally for brainstorming, and increasingly for fact-checking. But not for outlining, drafting, or directly editing.

This isn’t some principled objection to AI use, or a revulsion at the idea of letting technology intrude on my craft. It has simply never occurred to me to do it.

Why? I just don’t think AI writing is good enough today. It’s good at research reports or letters to Congressmen or anything else generic. It isn’t compelling enough to make a good blog post. And it certainly doesn’t have my voice or unique point of view. (I don’t think any AI today has a real point of view.)

Here’s an example. In revising “The Progress Agenda” for publication as a chapter in my forthcoming book, I wanted to add a paragraph on immigration.

Here was my process. First, in my notes, I wrote down several points I might want to make about immigration, off the top of my head. From that, I wrote a simple structure for the paragraph:

1. immigration is good/important

2. but it is difficult/time-consuming/uncertain

3. so we should make it easier (need examples)

At that point I needed data and examples to substantiate these points, which I didn’t have off the top of my head. This is where I turned to AI. I started with the promprt “What are some examples or data points to substantiate the idea that immigration, particularly high-skilled immigration, is important to the US?” I had a conversation with it from there, checking out the sources it pointed me to and asking followup questions until I felt I understood each point well. You can read the whole thing here.

Once I had enough material, I went back and summarized what I felt were the most relevant points in my notes, with sources. Then I drafted the paragraph.

Here’s what I ended up with (at the end of a section calling for across-the-board regulatory reform):

Progress in all of these areas depends on talent, and in the US, a lot of that talent comes through immigration. Immigrants are about 14% of the population, but 43% of doctorate-level scientists & engineers. Immigrants or their children co-founded nearly half of the Fortune 500 list as of 2025, including Amazon, Apple, Google, and NVIDIA, and more than half of US billion-dollar startups, including Stripe, Instacart, and Notion. But we restrict our supply of talent through a slow and burdensome immigration process. The H-1B visa used by many STEM workers is capped at a level far below demand: in 2026, there were almost 339,000 applicants, but only 35% of them were selected. The selections are done by lottery. Once here, the path to permanent residency can be epic, particularly for some nationalities: as of April 2026, some Indians even in the second-highest employment-based preference category (professionals with advanced degrees or “persons of exceptional ability”) have been waiting around 12 years for their green cards. And despite the exceptional success of immigrant entrepreneurs, the US has no startup visa. For the sake of national prosperity and security, the US should have more visas and green cards, including an entrepreneur visa; the ones we have should be better targeted to need, not assigned randomly; and processing should be made more efficient, to eliminate decade-long backlogs.

That paragraph isn’t going to win me a Pulitzer, but it’s solid and it does its part in the overall essay.

As an experiment, after I wrote this, and without showing it to ChatGPT, I prompted it:

Given all this discussion, please draft a paragraph on immigration, arguing that 1. immigration is good/important, 2. but it is difficult/time-consuming/uncertain, 3. so we should make it easier (examples). I want to insert it into this essay, near the end of the first major section, on law/regulation: https://newsletter.rootsofprogress.org/p/the-progress-agenda

Here’s what it wrote:

Immigration is another area where law and regulation need reform. For over a century, the United States has benefited enormously from immigrants, who have contributed disproportionately to science, engineering, entrepreneurship, and economic growth. Many of our most important technology companies were founded or co-founded by immigrants, and immigrants are heavily overrepresented among PhDs, inventors, and startup founders. Yet our system for admitting and retaining high-skilled talent is slow, cumbersome, and irrational. We educate foreign students in our universities, then subject them to visa lotteries, years-long backlogs, and administrative obstacles if they want to stay and build here. The result is that we needlessly turn away scientists, engineers, and founders, or trap them in long periods of uncertainty, even as we complain of labor shortages and worry about national competitiveness. We should make it far easier for talented people to come to America and remain here: expand and streamline visas for high-skilled workers and entrepreneurs, staple green cards to advanced STEM degrees earned at US universities, recapture unused employment-based green cards, reduce or eliminate per-country caps that create absurd backlogs, and generally treat the attraction of global talent as a national advantage rather than a problem to be managed.

I don’t like this nearly as much. It’s more bland and less interesting. The main difference that leaps out at me is that my version wove in many more specific stats and examples. The ChatGPT version stays abstract.

I’m sure I could prompt or train a model to write more like me. It would be an interesting exercise to see how many rounds of editorial feedback you’d need to give in order to get it to produce a paragraph much closer to mine (maybe only one or two?), or whether you could come up with a general style guide or set of skills that would get it to produce better writing the first time. (In the general case, it would have to be much more than “include specific numbers and examples,” because this paragraph is not one that shows off my unique voice.)

I am unmotivated to tinker with this. Frankly, I’m kind of a diva about my writing. I don’t co-author with other people (I kind of dread the idea). I almost never allow guest posts on my blog. I don’t even let other people write social media posts for me. I flatter myself that my voice is inimitable. When someone sends me a draft of something that is supposed to come from me, more often than not I just rewrite it completely. (In contrast, I’m very happy to have AI write code for me.)

There’s also the fact that when I publish any writing under my own name, I feel that I am taking on a sacred responsibility. Any writing in my name has to be correct, to the best of my knowledge and ability. It has to not only say true things, it has to select the most important things to include, and take the most interesting and illuminating angles on them. When I sign a piece, I’m not just taking responsibility for errors, I’m taking responsibility for the entire worldview and perspective that gave rise to the piece. And I simply can’t do that unless I wrote it.

And the writing process does more than just produce a piece of writing. I learn a lot from researching, outlining, drafting, and revising, and the knowledge and worldview that is built up in my head through that process is part of the outcome. What I learn helps me decide what to research and write about next, serves as examples in future essays, and provides the substance of interviews that I give. Even if I could crank out perfect essays using AI, I couldn’t be a public intellectual that way.

If I did have a way to produce truly excellent writing with AI, I might publish it, but not with my name as the author. My role then would not be author but editor. This is, indeed, the future I predicted for writers, and it might be a direction I go in the future. But if I do, you’ll know. It’ll be right there in the byline.