(PARTIAL-)CROSSPOST: TAYLOR LORENZ: Model Behavior
The subheadline: "Kicking off an honest series about writers, thinkers, creators, & AI—how a tech journalist works with AI..." Taylor Lorenz has 20 AI tools & fortifies the hill "human-written"...
Taylor Lorenz doesn’t perform outrage about AI; she itemizes her use of it in SubStack’s new “Model Behavior” series. In her production pipeline AI aggregates research, organizes notes, drafts titles across five models, and runs on a script she wrote with Claude Code. She estimates her income has dropped about 25% because of AI. It is the most concrete portrait I have yet found of how a working journalist integrates these tools…
As Nilay Patel and David Pierce of The Verge <https://podcasts.apple.com/it/podcast/you-cant-ignore-google-zero-anymore/> noted yesterday, SubStack will not label or provide tools to exclude Neo-Nazi-adjacent writings from its algorithmic feed to you, but is building tools so that you can exclude AI-slop. And this is a very interesting development—a very interesting place to draw a line: <https://braddelong.substack.com/p/substackpangram-is-scanning-the-substack>
SubStack appears to be doing three things:
giving readers very easy access to Pangram’s tools for classifying work as AI-slop,
giving writers very easy one-click access to how Pangram will judge their work, and
attempting to start a less unreasonable discussion about, or perhaps vibe-register thermometer check around, the different hopes and fears of its readers and writers with respect to AI-tools and AI-slop.
And the first second installment of this third effort made its way into the harbor yesterday:
(PARTIAL-)CROSSPOST: TAYLOR LORENZ: Model Behavior
<https://www.usermag.co/> <https://on.substack.com/p/model-behavior-taylor-lorenz>
Kicking off an honest series about writers, thinkers, creators, and AI—how a tech journalist works with AI
Jul 29, 2026
We [SubStack] want to have an honest conversation about how publishers are feeling about and using AI…. We think it’s important to ask top writers, creators, and thinkers about their work, their businesses, and how they intersect with where AI stands today…. We’re introducing Model Behavior…. We’ll ask candid questions about the tech industry, AI pet peeves, how it’s impacting their sense of identity and livelihoods. We’ll also ask each guest to give us a behind-the-scenes look at their creative process…. We start off this week with Taylor Lorenz, the founder of User Mag.
Do you find yourself altering your writing style to sound “less AI”? Oh my gosh. A while ago, I wrote a piece for Vanity Fair where I used the phrase “deeply human.” That’s actually a phrase I use a lot. I use adverbs a lot. And somebody was like, “This is AI.” I actually had to show the edit history to prove it was my own bad writing.
How much do you spend on AI tools every month? Probably about $300 a month.
When does AI use cross over from “tool” into “something I’d owe my readers a disclosure about”? I think the more transparency the better, and I believe in answering readers’ questions….
Would you read a newsletter where the copy was 100% AI-generated Absolutely, I love several newsletters that are AI-generated…. I find a lot of value from those as well.
What are your AI pet peeves? Is there anything that irks you about how you see your peers using AI? I’m in a lot of journalist groups. Journalists are using AI constantly. I think there are a lot of people who are very performative about AI online. They’ll spend all day shitting on AI 24/7, and then they’ll come into the group chat saying, “By the way, does anyone know how to use this new Claude tool?” I think that’s a disservice to readers, especially if you’re a tech writer….
Has AI changed how secure you feel in your own value as a writer? I’m actually a terrible writer…. What LLMs cannot do—yet—is uncover things the way actual journalists can. AI has caused a great deal of instability in my career. I certainly question my ability to maintain a living as a journalist…. I would say I’ve lost about a quarter of my income [because of AI].
Have you changed your editorial strategy to counter that? I am in the process of trying to figure that out. The way I’ve tried to adapt so far is to try to leverage these tools to be more productive. But I don’t know how to counter it…
Do you think the way the tech industry has rolled out AI has helped or hindered its adoption? It’s a little bit of both…. There are millions of people getting utility from AI products…. At the same time, the people who run these tech companies are self-interested. They’re often bad actors…. People need to remember that AI is not just OpenAI and Anthropic, that large language models are a new technology, and personally, I use DeepSeek probably more than I use those other tools. So that’s another issue a lot of these people have—they think AI, they think Sam Altman.
What responsibility do platforms and AI companies have to writers and creators? I think platforms should do everything they can to support creators. There are platforms that are predatory, like Meta…. I also think how we should fund journalism is an open question, and I don’t think it should be on the tech companies necessarily to fix—their job is to build these products, not save journalism….
Any advice for someone who wants to start in journalism in the midst of all of this? I would tell them that the business model for journalism is in flux right now, so if you want to be a journalist, number one, build a direct relationship with an audience. That’s the most valuable thing you can do. And don’t shy away from new technology….
We asked Taylor Lorenz to take us through a day of research and writing, from a podcast episode to a newsletter, to give us some insight into her process and the decision points along the way.
PODCAST…. RESEARCH
Aggregate research: when I’m getting up to speed on something I don’t know well, like security and LLMs for this week’s episode on the China open source model debate, I’ll take posts from someone I trust (in this instance Alex Stamos) and ask AI to pull together everything he’s said on the topic.
Search Gemini, and run queries on Grok to check for anything I missed.
Paste what I’ve gathered into Claude and ask it to organize the information.
Run similar searches across models to track the differences in their responses and understand each one’s capabilities better.
RECORD: Schedule a Riverside conversation with my guest for that week. Record the conversation itself, human talking to human, no AI. I don’t write questions ahead of time; it makes the conversation feel less natural.
EDIT
Sometimes do a first cut myself in Adobe Premiere, sometimes send it straight to my human podcast editor.
My editor levels the audio, cuts any flubs, and puts together a draft of the episode, using AI tools in his process, especially for audio processing, while doing the cut manually in Premiere.
TRANSCRIBE: Save the edited video to my computer. A script I built with Claude Code takes over from there, copies the video into the correct folder, and automatically transcribes the file, generating a transcript.
TITLE AND DESCRIPTION: Run the transcript through five different models—Claude, Gemini, Grok, DeepSeek, and ChatGPT—with the same prompt.1 Pick whichever headline and description I like best and tweak it myself from there.2
THUMBNAIL
Use the thumbnail ideas generated for me, plus my own ideas, to build a thumbnail in Canva. This takes me over an hour at least; I’m not great at digital design. Canva’s AI tools let me auto-expand and transform images until I get something coherent.
Upload the draft video to VidIQ, which analyzes it and suggests tweaks to the thumbnail and alternate titles.
Go back into Canva and make VidIQ’s suggested changes. I usually go through about three rounds of this until I have a working thumbnail and title.
PUBLISH: Tweak the SEO-optimized description with my own wording, or reword anything that doesn’t read well. Upload the video to YouTube, then upload separate copies to Patreon and Spotify.
CLIP
Check Spotify first, since their auto-clipper generally picks the best moments.
Re-create those clips from scratch in CapCut, since Spotify’s short-form clips can’t be downloaded, reworking the aspect ratio and adding stock footage from Pexels.
Run each piece of short-form content through VidIQ to optimize the title and hashtags.
I’ve also been experimenting with OpusClip for generating short-form content.
This entire process takes multiple days. I’m effectively a one-woman operation, and I publish two video episodes a week.
NEWSLETTER
Because producing video and podcast content takes up 95% of my time, I don’t have as much left over to write separate original content for my newsletter in every edition. My newsletter functions as a complement to the podcast, and it’s the way I fund the podcast and video content, since I have zero sponsors. When I do write articles or longer pieces, they’re 100% human-written. Usually I’m reporting on a phenomenon or trend I’ve noticed, and to write it I conduct interviews, pull research, and do classic low-tech reporting. I’ve been thinking a lot about agents lately. I’d love to build some kind of agentic system that skims my articles and automatically sends emails from my address reaching out for comments to everyone necessary, but I haven’t built that yet.
BUILD THE TOP OF THE NEWSLETTER: Take the SEO-optimized YouTube video description and put it at the top of the newsletter, sometimes tweaking it or adding an opening line or two. Embed the YouTube video below that. This stays above the paywall, and it’s built this way to keep that top section optimized for search.
WHAT I’M READING: A list of headlines and links to articles I’m reading, mostly deep dives and long reads. Below the paywall, no AI involved.
THE NEWS RUNDOWN: Curate about 30 to 40 news items I’m following that week, always 100% human-curated. Hand it to Claude to sort into categories, something readers specifically asked for. Claude still isn’t great at this, so it takes some cleanup afterward.
Every item in the rundown is a single sentence with a link back to the original source—something that caught my interest or that I thought was notable or funny. Sometimes that’s a tweet I’m copying, and honestly, I don’t always know if the tweet itself was written by AI. I didn’t write it, so I can’t say for sure. Sometimes it’s a headline; sometimes I write the sentence myself. The most important thing is that I always link back to the original source, so readers can see for themselves where the information came from…
1 MY PROMPT: The transcript for a YouTube video is attached. Your job is to act as a YouTube growth strategist who specializes in long-form news, politics, and technology videos. Before generating anything, do the following:
Read the entire transcript.
Identify the video’s:
Central thesis
Biggest revelation
Most emotionally compelling moment
Most controversial claim
Highest-curiosity unanswered question
Most surprising quote or fact
Search YouTube for the highest-performing recent videos covering similar topics and analyze:
Common title patterns
Curiosity gaps
Emotional framing
Word choice
Length
Search current news headlines and discussions across the web related to this topic to identify trending language, names, and keywords people are actively searching for.
Optimize every recommendation using those findings rather than simply describing the video. Base recommendations on current trends, but do not invent facts or promise things the video does not deliver.
Next, generate 10 YouTube titles.
Requirements:
Under 60 characters
No periods
No colons
Every title should feel different
Rank them from strongest CTR (click-through rate) potential to weakest
Each title should optimize for one of these approaches:
Curiosity
Shock
Conflict
Fear
Stakes
Revelation
Narrative
Mystery
Authority
Search intent
Avoid generic or corporate wording.
Make them sound like titles currently outperforming in this niche.
Do not repeat the same structure multiple times.
For each title include:
Estimated CTR potential (1-10)
Generate 10 thumbnail concepts.
For each concept include:
Exact image composition
Camera framing
Facial expression
Background
Lighting
Color palette
Objects shown
Where text should appear
Exact thumbnail text (1-4 words)
Rules:
No split-screen thumbnails
No clutter
One dominant visual idea
Text must be readable on mobile
Thumbnail should create curiosity that the title completes, not duplicate it.
SEO Description
Write a YouTube description that:
Begins with a compelling two-sentence hook
Naturally incorporates 5+ trending search terms relevant to the topic
Explains what viewers will learn
Includes keywords without keyword stuffing
Ends with a strong call to subscribe
Generate:
30 high-value YouTube search keywords
20 long-tail keyword phrases
10 hashtags
Assume the objective is maximizing long-term YouTube performance in addition to maximizing search traffic. Prioritize click-through rate and viewer curiosity while remaining completely truthful to the contents of the transcript. Prefer titles that create an “information gap” over descriptive summaries…
<https://www.usermag.co/> <https://on.substack.com/p/model-behavior-taylor-lorenz>
Brad DeLong here: What do I think of all this? I do not know yet. I just have a bunch of questions. Eleven of them, in fact:
