How I Use AI to Make My Newsletter, and What I Won't Let It Do.

In January 2025, I was working on a piece about the fourth anniversary of January 6th and all that had changed in that short period of time. I worked the way I usually do. I dumped everything in my head into Claude, along with a first draft, and started pulling threads.

At one point, the robot handed back an edited line that was sharper than anything I’d written that morning.

I was worried it might be plagiarized, so I Googled it. I asked the model if it had taken it from somewhere. Nothing came back. As far as I could tell, it hadn’t been lifted from anyone.

I left the draft open in my browser tabs for two days (I also got pulled into other things when Facebook announced a bunch of big changes ahead of Trump’s inauguration). The news cycle moved, the piece went stale, and I never published it, so I never had to make the call. The question stayed anyway. If the words came out of the model and I fed it every idea underneath them, whose sentence is that?

Those two days produced my first ethics and disclosure statement instead of an answer.

I’d been looking for a way to stop relitigating this at the top of every post. Adding a line to each piece saying “I used AI” would have been true and useless. Studies show people start ignoring consistent labels like that. The nuanced version of what I was doing needed more room: which tools, for what, where I stop, what I verify, what I won’t put into a model at all. I’ve updated it many times since, mostly because I keep adding tools and changing my approach.

We’re living through a moment where AI, politics, media, and technology are all crashing into each other at once. Anchor Change is where I connect the dots, share what I’m noticing, and help people panic responsibly about what comes next. Subscribe for grounded analysis and strategic insight from someone who’s been inside the rooms where these decisions get made.

Subscribe now

What Substack shipped

Last Tuesday, Substack announced an integration with Pangram that lets readers scan any post, note, reply, or comment over 100 words and see an estimate of how much was written by hand and how much with AI assistance.

The scanning tool is what’s getting the most attention right now. I get it; I’m worried about this making it easier for people to make knee-jerk judgments about how I use these tools. That said, the other part of the release is what I’d point you to. Writers can now run Pangram on their own drafts before publishing, turn scanning off post by post, and add a “How I make this” statement that readers see when they run a scan.

My read: the scan is what everyone is arguing about this week. The “How I make this” field is what will matter in six months.

A detection score measures the surface of the prose. It can’t tell you who had the idea. It can’t tell you who noticed the same pattern showing up in four countries, who went back and checked that the number in paragraph six came from a survey of 68 people and rewrote the claim accordingly, or who decided halfway through that the argument was wrong and started over.

So let me save you the scan. A lot of my writing is going to come back AI-assisted. Probably most of it. That’s accurate, and I’d rather tell you what it actually means than argue about the percentage.

What’s mine and what isn’t

The ideas are mine. The connections are mine. The analysis, the read, the call on what matters and what doesn’t — all mine. I’m not asking a model to do any of that, and it isn’t good at it anyway.

Here’s what the tools are actually doing.

I start with a brain dump, usually by talking. I use Wispr Flow, and I do a lot of it in the car, because that’s where I have my best ideas and it’s faster for me to talk than type. Then I ask the model to interview me. Being asked questions pulls more out of me than staring at a blank document does, and it surfaces the thing I actually think two or three questions in.

With data, I go through the numbers myself first and find what’s interesting. Then I ask what I missed. Sometimes it catches something. Every number it gives back gets verified against the source before it goes anywhere near a draft.

With any stat, I ask the model where it came from and then check it, or I Google it myself. If I can’t source it, it doesn’t run. I load in the links and reports I want a piece to draw from rather than letting it fill gaps on its own.

And I’ve put real work into training these tools on my own writing, creating skills, instructions, and markdown files so what comes back sounds like me instead of sounding like everyone.

The trade I’ve made is straightforward. I’m a solo operation. The tools let me produce at a volume I couldn’t otherwise, and in exchange I’ve accepted that my work will register as machine-assisted. Other people won’t, and I have a lot of respect for the writers who treat the drafting itself as the craft and want every sentence to be theirs. That’s a legitimate choice. It just isn’t the one I made.

The part I’d watch

Pangram says its false positive rate is 0.01%. I want that to be true. The track record of this category makes me cautious about it. In fact, multiple notes I’ve posted since the announcement where I used a mix of AI and my writing have come back as fully human-written, whereas others I’ve done the same way have come back 100% AI. Detection tools have repeatedly flagged writing by non-native English speakers at higher rates than everyone else.

This is an impossible tradeoff with pain on either side. Readers genuinely deserve some way to know whether a person was on the other end of what they’re reading. And a writer who gets wrongly flagged has no appeal in the moment, no way to prove a negative, and the accusation is what people remember. Substack knows this — they’ve said as much, and they already loosened the opt-out after the first round of feedback. (Though I would prefer they offered this at the publication level AND the post level. I also would hope that writers would be willing to be transparent about the fact they turned the tool off and why.)

There’s a second-order effect worth watching. If the score becomes the scoreboard, the rational move for a heavy AI user is to sand the prose down until it passes. Detection rewards concealment. Disclosure rewards being straight with people. Substack shipped both in the same week, and which one the culture actually picks up is still open.

If you want to write your own, the hard part isn’t the statement. In talking to people about this, the wording is almost never where they get stuck. What’s missing is a decision about their own values and guardrails. You can’t disclose a practice you haven’t examined.

That’s what I built the AI Transparency Workbook for. It walks through where AI actually touches your work, where your lines are, and what the people reading you deserve to know — and you come out of it with a statement you can publish.

Norms around this are still forming, and the argument happening on Substack right now is one we need to have. We all choose who we read and how we spend our time, and that choice only means something if the people making the work are straight about how they make it.

I still don’t know whose sentence that was in January 2025. I’d rather hand you the whole process and let you decide than have a percentage decide for you.

If this helped you connect some dots, consider sharing it with someone else trying to make sense of where AI, technology, politics, and public trust are all heading next. Word of mouth is still how most people find me, and I’m grateful for every share.

Share