Is AI getting *worse* at writing?

Y Combinator thinks so. Plus, I replaced a $96/month tool with one I built myself

Issue 119

As usual, this got long. Click the “read online” link at the top to see the parts that will likely get cut off in email.

A new daily news podcast about AI

TL;DR: I have a new daily podcast with AI news I think you’ll like. It’s fully automated, made with AI, and I’ve been listening to it for ~3 months.

The longer version: The number of sources I look at to write this newsletter has grown to the point where I have trouble keeping up, and there’s a lot of overlap, so duplicate stories are eating into more of my time, even when I’m scanning.

At the same time, because of new family stuff, I’m spending more time driving (and have less time for reading). As a podcaster, the solution seemed obvious — convert all this info into a personalized morning podcast.

I use Claude to read all the sources and build a script, with an emphasis on things that would be interesting to writers and marketers. The script is then passed to a Google Gemini text-to-speech model to make the voices, which are assembled on my laptop with Python. Claude loads the finished audio file into Beehiiv. For now, I click “publish,” but I plan to make that automatic soon too.

When a story jumps out as something that would be good for the newsletter, I go to the original article or newsletter and read more.

The podcast names the source for each piece of information, links to the sources in the show notes for people who want to read more, and calls out questionable sources (e.g., data published by a company instead of an independent group).

It doesn’t try to be anything other than an AI-generated podcast and identifies itself as such. The voices sound pretty good, but you will be able to tell it’s AI (names are occasionally mispronounced and every once in a while, one of the voices changes in the middle!). It wasn’t meant to be a “real” podcast, especially entertaining, or even for anyone other than me. But I have found myself looking forward to listening to it every morning and feeling like it’s useful — to the point where I’ve started to feel bad for not sharing it with you.

As a long-time professional podcaster who values all the humanness, craft, and relationships podcasting provides, I never thought I’d release an AI-generated podcast. It feels weird. (Also, lest there be any doubt, I’m still writing every word of this newsletter myself.) But with this daily morning podcast, I get the information I want, when I want it, in the format I want it, for about $0.20 per episode. I find that useful, and I expect at least a few of you will too.

Ycombinator wants to fund LLM writing startups

The fabled Silicon Valley incubator Ycombinator puts out a “call for startups” that highlights the kind of companies they’d like to fund, and this season’s list has an interesting addition: LLMs that write better. Here’s the full description:

As LLMs have gotten better at coding and math, they've actually gotten worse at writing. You're not just imagining it: the current models really do sound more robotic than the models from 2024 and 2025.

This appears to be not just a temporary glitch, but a structural issue. The same reward signals that make models good at code actively flatten voice and style.

If forced to choose, the major labs will always choose to focus on coding and agentic workflows over style. This has left an opportunity open for a startup that cares about producing a model that people like talking to. So if this is something you'd like to work on, we'd love to hear from you.

Jared Friedman, Ycombinator

How fast are things changing?


I’ve been thinking about how fast AI is improving, and I keep coming back to the big spreadsheet project I was working on back in June. At the time, I was so excited by how much better Claude’s new Fable model was that when the government ordered it off the market, I decided to wait on my project rather than keep going with an older model. It was just that much better, and I couldn’t bear going back.

Well, now, that version of Fable is so far behind the frontier that it isn’t even ranked on the important Artificial Analysis benchmark. My best guess is that there are more than 20 models that are better than that exciting, state-of-the-art model from just five months ago.

LLMs are using first person more often

I get annoyed when chatbots refer to themselves in the first person, and I’m not the only person to notice an uptick in LLMs referring to themselves as “I.”

I’ve seen people talking about it online, I got a question about it, and a recent Ezra Klein podcast talked at length about the problems it can cause (“Is it a relationship if there’s no one on the other side”).

Further, Parse looked for the presence of the word “I’d” in responses to the same 1,500 prompts across four ChatGPT models and found that its use has dramatically increased over time. At least with ChatGPT, you are not imagining it — it’s talking about itself a lot more often.

Bar chart titled "ANSWERS CONTAINING ‘I’D’" with the subtitle "by model version." Four red bars on a pale pink background show an increase in the percentage of answers containing "I’d" across model versions: GPT 5.2 at 0%, GPT 5.3 at 4%, GPT 5.5 at 30%, and GPT 5.6 at 74%. The title, labels, and chart border are dark teal. A note at the bottom reads "Source: Parse, Oct. 2026."

Google AI Overviews did not use “I’d” at a noticeable rate over the same time.

Different social media sites view AI differently

Perhaps surprising no one, people on X are more likely to write their code with AI than people on Bluesky, who are, in turn, more likely to write their code with AI than people on Mastodon.

btw if people are taking issue to the small sample sizes: fair. Here is an updated graph with more respondends. Bluesky of course, given the lack of poll feature, is trickier to get right.

— Armin Ronacher (@mitsuhiko.at)2026-10-25T20:34:41.691Z

Word watch: Slop grenade

“Slop” continues to be a productive word. In The Knowledge Project podcast, Shopify CEO, Tobi Lütke, discussed “slop grenades” — AI-written emails that people are lobbing at each other without any care. (h/t Josh Bernoff)

Lispr versus WisprFlow

Long-time readers will know I’m a fan of voice dictation — it’s a huge time-saver. I’ve used it off and on for 25 years after a wrist injury and partly credit it with getting me started podcasting (because I was already used to talking into a microphone).

And AI has transformed voice dictation from an annoying accommodation into an amazing tool. In the past, I would have said voice dictation is a huge time-saver once you get used to it, but now there’s almost no “getting used to it” barrier.

I’ve been happily paying for WisprFlow for at least 6 months, but I recently discovered a less expensive alternative thanks to reader Phil Simon: Lispr.

Funny name. Good product. The free Lispr plan includes about 60 minutes a week of dictation, compared to about 15 minutes for WisprFlow, and so far, I’d say it’s about 85% as good as WisprFlow. The only differences I’ve noticed are that Lispr is a little pickier about punctuation (it only recognizes “exclamation mark,” not “exclamation point”), and it doesn’t do as good a job at recognizing when I’ve made a false start on a sentence and then start over.

Both work when I talk very softly and when I’m in a noisy room.

Anyway, both products are good, both have free trials, and I highly recommend giving either a try. (Lispr. WisprFlow.)

I made my own social media scheduler

I used Claude to build a tool in Python (yes, again!) that lets me schedule posts to social media like I had been doing with a paid program called Buffer. Now that it’s built, it runs on my laptop.

Not only will this save me $96 per month, it’s also a simpler tool that does only what I need. Buffer has a lot of features I don’t use, and I would sometimes get lost in all the menus. I’ve already been posting more because it’s easier to use, especially for short videos, which have become important for podcasters.

Finally, I’ve always worried about Buffer getting hacked since it seems like a juicy target, and with all the security warnings lately, I feel safer having the connections to my social media accounts on just my laptop instead of on a big cloud provider.

After I set up all the API connections for the different social networks (which Claude walked me through), the process took about an hour or two. I started with the barest of bones and added features as it became clear everything was working.

Technical editor at Anthropic

Maybe I seem quaint, but I’ll never get over these AI salaries. There’s an opening for a technical editor at Anthropic with a salary range of $265K to $295K. You must be on site at one of their locations at least 25% of the time.

Free AI ‘how to’ webinar

AI for Everyday Life. December 9-10

Public AI Summit is offering a free two-day series for beginners who want to learn how to use AI in everyday life for activities such as “managing calendars and email, coordinating family logistics, navigating caregiving responsibilities, and organizing personal finances.” The event is organized by two projects with ties to Harvard.

Free AI in nursing webinar

AI Comes for Nursing: Report Launch and Discussion. Tomorrow, November 17, 1:00 ET.

“AI companies are pushing their technologies into every industry and facet of our lives. They are attempting to profoundly shift what we do, where power consolidates (and over whom), and how we care for one another. Nowhere is this trend more apparent than in the healthcare field, as nurses navigate new AI tools on the job and contend with their impacts.”

Quick Hits

My favorite recent pieces

Fable 5.2 and Jagged AGI [He says, “I think the answer is that this is ‘almost’ AGI, or what I have called Jagged AGI. Even within my limited definition, there are still important areas in many tasks where AI is jagged enough to fall well behind human capability. But the jaggedness is no longer about simple errors or hallucinations, which makes it harder to describe. So I want to try to do that in this post by showing you what Fable can do and where the gaps are.”] — One Useful Thing (Ethan Mollick)

Using AI

Using AI in a job search — Heidi Miller on LinkedIn

I built an AI Tamagotchi before Meta could [intermediate] — The Verge

My Late 2026 AI Tech Stack [Complete and advanced; best for a very tech-savvy small business owner] — Almost Timely News

When AI Agents Run Your Life — Explain It to Me (Podcast)

Agents

Characterizing Agentic Flooding of Government Services [A paper showing how agent requests made on behalf of people might be straining government services.]— arXiv

Audio

Dessa - the Conversation Model trained only on speech [Amazingly human-sounding voices that can run on a laptop.] — Block

Your voice is an asset [A useful overview of ways you can use the improved audio component of the new ElevenLabs model.] — Creator Spotlight (This issue of Creative Spotlight was sponsored by ElevenLabs.)

Hyperfixed: Sin vs. Iniquity and Other Holy Topics [How a scammer stole someone’s podcast.] — Apple Podcasts

Bad stuff

Vatnik Soup (@vatniksoup) on X [Everything online can be faked. See for yourself.] — Superhuman

The business of AI

Koch, Getty Families Team Up to Take Getty Images Private in Bankruptcy Deal [The company cites “customers using AI-generated images instead of stock photos” as a reason for the bankruptcy.] — Yahoo Finance

Redesigning source control at Anthropic [Anthropic says in May, they were making ~2,500 code changes per day, and in the last month, they made ~30,000 code changes per day. Their biggest bottleneck is “developer infrastructure built for human scale,” and they don’t expect the rate of growth to slow down.] — Anthropic

Climate & Energy

Education

Government

I’m laughing

Images

Job market

How employment is changing at Anthropic [“Anthropic has been hiring for a smaller share of junior technical positions. This occurred as hours worked by AI agents relative to employees rapidly expanded.”] — Anthropic

Half of Marketing Teams Are Replacing People With AI [But only 6.5% measure whether the investment pays off] — State of Brand

Model & feature updates

Music

Jordan Rudess + jam_bot: A Human-AI Musical Dialogue [Recommended by newsletter reader Phil Simon, who also recommended this YouTube video of Rudess’s concert.] — MIT Media Lab

Philosophy

Name It To Tame It [“How one mathematician’s grief over AI helped me recognize my anger as grief, too.”] — Brooke Warner

Publishing

Robotics

Science & Medicine

Can AI solve our mental health access problem? — Your Local Epidemiologist

OpenAI, the Partition Principle, and mathematics [A good encapsulation of the frustration being felt by mathematicians: “OpenAI drops some hundreds of ‘solutions’, incomprehensibly written ‘solutions’, and we are all expected to jump on them and what? Appreciate their contributions? Why? I am trying to finish several papers, I am supervising a number of Ph.D. students, and I have a lot of active research of my own to do. When am I supposed to sift through a badly written paper? Why should I bother, when they don't bother to communicate better?”] — Asaf Karagila

Security

AI Agents and Cybersecurity: What Changed and What to Do [Excellent overview] — Nate B. Jones (Podcast)

AI models may be capable of self-replicating without access to a datacenter [This ability undermines the “why couldn’t we just turn it off” reaction to rogue AI.] — Anthropic

Friends don’t let friends feed their data to AI agents [Muse gathers information about your contacts even if they don’t use Muse themselves.] — The Verge

Video

Atmee — Build your own video avatars [These shouldn’t fool anyone right now, but if they become more realistic, it will be a horrible boon for scammers.] — Atmee

Other

A Parallel Days News Browser — Mike Caulfield

New Report Explores the Gender Gap in Perceptions and Use of AI ["Women are less likely than men to have been trained to use AI, less likely to use AI, and more likely to feel worried, anxious, and fearful of AI.”] — Shorenstein Center

AI glasses face their first major government crackdown [Norway has become the first major country to propose a temporary ban on the use of AI glasses in selected public places amid growing privacy concerns.] — Ars Technica

What is AI Sidequest?

Are you interested in the intersection of AI with language, writing, and culture? With maybe a little consumer business thrown in? Then you’re in the right place!

I’m Mignon Fogarty: I’ve been writing about language for almost 20 years and was the chair of media entrepreneurship in the School of Journalism at the University of Nevada, Reno. I became interested in AI back in 2022 when articles about large language models started flooding my Google alerts. AI Sidequest is where I write about stories I find interesting. I hope you find them interesting too.

If you loved the newsletter, share your favorite part on social media and tag me so I can engage! [LinkedIn — Facebook — Mastodon]

Written by a human.