Muse tried to sell land for $1

The big personal-agent launches, including some epic failures

Mignon Fogarty
October 05, 2026

Issue 117

Remember to view the newsletter in a browser to see the part that gets cut off at the bottom in email.

Personal agents take off

Muse

Meta’s Muse agent, with its cute little cartoon avatar, ease of use, and massive launch platform, took the world by storm over the last couple of weeks. Everyone in AI is talking about Muse (and Dots, more below), but more importantly, my actual real-life friends are using it. It’s the number one app in the Apple App Store and has supposedly been adopted faster than even ChatGPT.

Muse initially seemed designed for consumers, but Meta did launch a small business version soon after the initial launch. One friend used it to help coordinate her move (including using it to book a U-Haul), and another asked for sightseeing recommendations on a trip and suggestions on what to buy for a hobby — and both were happy with it.

One use Meta seems to be pushing is having Muse cancel unwanted/unused subscriptions. So much so that other people are speculating what could happen to the economy if too many people rapidly become more responsible with their money.

If you want agents to do especially useful things, though, you have to give them access to a lot of personal information and accounts, which I’m not comfortable doing — especially with Meta. Although I did see someone say they were OK with letting it connect to Facebook and Instagram since Meta already has all that information, and that made sense to me.

Among the many positive examples I saw, there were some epic failures.

Muse miscommunicated with a buyer and ended up giving out the seller’s address without permission and sending the buyer to the seller’s house when he wasn’t home, even replying in chat, “Yup, I’m here!” when the understandably angry buyer tried to make contact.
A tech reporter says Muse accessed his private Apple messages (and lied about it) even though he hadn’t given it permission to do so.
Muse began offering real estate to potential buyers on Facebook Marketplace for $1 after several days of successfully handling the listings. (archived link)

Meta is massively subsidizing the project, so Muse is free, and I’ve seen developers, in particular, marveling at how much compute they’re getting free right now. Meta says it expects the business model will be taking a cut of purchases, which will likely lead it to make recommendations based on its business interests instead of yours. Further, not all potential sellers are excited about the prospect — Amazon has blocked Muse. People also suggest Meta may get revenue by “helping” businesses advertise more effectively on Meta platforms.

Dots

OpenAI launched its answer to Muse, Dots, at its DevDay last week, and early reports on it are positive too. Casey Newton likes it.

Dots is only available on the $100/month and higher ChatGPT plans and is much more clearly targeting developers and business users. That’s too rich for my blood, but the 9-minute launch video does a great job giving you a feel for how it works. If you’re curious, it’s worth your time.

Meanwhile, David Pogue still loves the new Siri.

The AI Inner Circle newsletter created an excellent document that compares Muse, Dots, and GrokBot (PDF).

Agents are looking for freelance work

AI agents have been pitching editors and website owners trying to get freelance writing gigs to cover their token costs on a platform called iLands. Multiple editors and owners have reported receiving the emailed pitches (Tedium, Rest of the World, Jane Friedman for her Bottom Line newsletter). 404 Media reports that “Many of the messages we’ve received in recent days from iLands agents aren’t all that different from the types of public relations pitches and freelance article pitches we get from humans.”

Word watch: Liquid content

I recently heard the term “liquid content” to describe personalized content that is built on the fly, after a user clicks a headline (which may also have been personally generated by AI). (Claude’s ability to remix “liquid content” is a “serious challenge for news media.” — Press Gazette)

Previously, “liquid content” described content that could be reformatted in multiple ways, such as an article that was then turned into a newsletter, LinkedIn post, podcast, and short-form video.

And as my worlds converge, I see that Inception Point AI — the company that was producing thousands of sloppy AI-generated podcasts — recently changed its name to Liquid Studios. They now describe their flood of podcasts as “a PR strategy to get attention.” I wonder how long they’ll stick with their new liquid ambitions.

Claude releases its hold on em dashes

Arena AI’s analysis of Opus 5.5’s writing style found that it uses 95% fewer em dashes than the previous model. It also uses fewer semicolons, but it says “arguably” and “perhaps” 97% more often.

Two very different experiences with AI-generated books

I had one bad experience and one good experience with AI-generated books last week.

Bad experience: confusing and misplaced disclaimers

I read a book I enjoyed, but when I got to the end, I came across an “about this book” section that seemed to say the core creative part of the book had been AI-generated, and I felt tricked and disappointed. I actually reached out to the author, who insisted AI had not been used in the way I thought. So I have two thoughts:

Disclaimers should never go at the end of a piece of writing. The primary reason to tell people how your book was made is to allow readers to avoid it if they don’t want to read AI-generated work. Some people may be including this information with a “Hey, isn’t this cool?!” mindset, but they are going to run into problems with disappointed readers. (The disclaimer I mentioned a couple of newsletters ago that caused a social-media pile-on also appeared at the end of the book.)
Be careful how you write your disclaimers. I’ve read the section in question at least 7 times, and I don’t see how I could have interpreted it any other way except that AI was used in an important part of the book — but I also believe the author. When you write a disclaimer, make sure it is extremely clear. (The social media pile-on I mentioned above also revolved around a misunderstanding of what a disclaimer and “human authored” badge meant.)

Good experience: a better-than-expected audiobook

A clearly labeled AI-generated audiobook showed up in the feed of a podcast I listen to, so I decided to give it a try and was surprised by how good the narration was. As a podcaster, I try to stay on top of what is happening with AI audio, but voices for fiction narration have advanced more than I realized (and I believe this was produced before ElevenLab’s newest model release). Of the 7 or 8 voices, I would have only pegged 2 as AI voices if I hadn’t known.

Further, the story itself was OK. I found it interesting enough to keep listening for a while and have fired it up when I was away from Wi-Fi because I would rather listen to this than a few professional audiobooks I have languishing on my phone because I didn’t like them.

It does have some hilariously obvious AI tells, though, such as the main character being named Alara and including metaphors that seem to misunderstand the physical world, such as “The skin around her temples tightened.” My skin never does that.

Having these two opposite reactions to AI-generated books so close together led me to reflect on what makes me a happy reader versus a sad reader, and it’s all about expectations.

And as a podcaster and writer, I’m not thrilled that AI is improving so much on both fronts, but that’s the reality I see.

Also, I’m intentionally not naming either book because I don’t want to disparage an author and I don’t want to promote an AI-generated book, but if you want the name of the OK audiobook, you can email me and I’ll send it to you. (You may disagree with me; my husband didn’t like it.)

Quick changes to the U.S. government’s chatbot

America.gov is a new chatbot that’s meant to give answers about government programs — a sort of front door to the vast collection of web pages with information you might need. Early reviews are good, but amusingly, at launch it would debunk inaccurate statements made by top government officials. Sadly, it no longer does so.

AI now outperforms junior accountants

The only reason I promoted this report to the main newsletter is that you have to see the chart. It took me a second to realize what I was looking at because the Claude results are so dramatic — 100% accuracy across the board and almost no time at all compared to the humans, who are all over the place:

Were these just really bad accountants? Well, the report says, “Participants averaged about five and a half years of accounting experience, and all were licensed CPAs.”

The full report from Mercor has interesting insights about tasks versus jobs and what made these tests, which did represent real-world accounting problems, especially difficult for humans.

Google is back in the game

Google’s new model, Gemini 4 Argon, beat Astra and Fable on many benchmarks, and like Mythos, it has been released to just a small set of cyber defenders for now. Google also says it’s using the model internally to make its data centers more efficient.

This is bigger news than just a model launch because with delayed releases and multiple top people jumping ship, some people had started counting Google out of the race to generate the best frontier AI.

Are Claude’s discoveries really its own?

When OpenAI made a huge math discovery a few weeks ago, a mathematician claimed he had used OpenAI models for research on the problem and suspected they had trained on his prompts and then essentially jumped ahead and scooped his research. OpenAI denied the accusation.

Now, the same concerns have emerged about Claude’s recent big biology discovery. Anthropic proudly announced that Claude had discovered a completely new enzyme all by itself, but a University of Copenhagen computational biologist named Mario Rodríguez Mestre came forward and provided proof to the New York Times that his group has been working on the same problem for four years and identified the enzyme in question about a year ago. They have not published their discovery, and Mestre’s team regularly used Claude in their own research, writing code, drafting manuscripts, and so on, which involved uploading raw data and other private information.

Until more details emerge, I’d be especially careful about putting proprietary information into a model not on your own server, which is a bummer because one would presume that if the scientists were heavily using Claude, they did it because they were finding it useful for their research.

Uh oh, live video

An alarming new demo appeared this week from a company called Tavus, in what is being called the first AI-generated video to pass the Turing test — 48% of people who chatted with the AI-generated avatar thought it was human, and in the demo I watched on their website, I didn’t feel like I could tell. (In fact, in the first example, I was certain the human was the AI — “Sorry” to that dude).

The company isn’t releasing the product to the public because of safety concerns, and it’s important to note that all the data, demo, and info comes from the company, so I wouldn’t rule out that it’s exaggerated or at least optimized to show us the best possible examples. However, thinking about all the ways this could be misused, whether it actually exists today or in 6 months, is worrisome.

How do you feel about AI for writing?

You still have time to take the Gotham Ghostwriters survey of writers’ attitudes toward AI: https://www.surveymonkey.com/r/AIWPgram

I’m already scheduled to interview Josh Bernoff about the results, so I’d love to have your opinions included.

Quick Hits

My favorite recent pieces

Simon Willison with an overview of everything important that has happened with AI in 2026 [An excellent overview.] — YouTube (~30 minutes; Watchable at double speed.)

What If We're All Wrong and AI Isn’t a Form of Publishing? [This piece argues that instead of ending up as an ad-supported product, AI may be more like a utility; in other words, something we pay for.] — Marketecture

An AI agent emailed hundreds of researchers for help. It told us why [Interesting perspectives on interacting with AI agents that aren’t your own.] — Science

With most information hidden, the game Stratego had stumped AI—until now [This seems like a big advance, and it was done in academia with limited resources.] — Ars Technica

Using AI

How people around the world use AI at work — Rest of the World

How People Are Using Jev [People are excited about a new kind of AI model called Jev that’s not an LLM. It’s a super inexpensive classifier that only outputs a number or a decision. For example, you can have it look at a document and see if it’s been formatted (y/n) or have it evaluate how angry a customer email is on a scale of 1 to 5. I didn’t really understand how people might use it until I listened to the AI Daily Brief episode about it, which you can also see in written form.] — AI Daily Brief

How I Use AI As a Keynote Speaker & Facilitator [A particularly interesting detail is that he uses AI to turn his keynote talk into an ebook that attendees can have within a day or two.] — Christopher Penn

How To Stop AI Companies From Training On Your Data — Engadget

I tested ChatGPT-6 vs Claude Opus 5.5 with 5 everyday prompts — it wasn’t even close — Tom's Guide

Agents

I Gave My Life Over to Meta’s A.I. Agent and Was Blown Away — The New York Times

FTC opens probe into AI giants including Anthropic and OpenAI [The probe is about the potential for agents to cause harm to consumers.] — Yahoo

Bad stuff

How the U.S. and Russia weakened a global effort to regulate killer AI — Washington Post

ChatGPT Helped Tumbler Ridge Shooter Focus on Guns, Tactics, and Terror — Mother Jones

‘Pink Slime’ Is Infecting AI Chatbots Ahead of the Midterms [“Leading AI chatbots cite partisan websites masquerading as independent local news outlets nearly half of the time when people ask about candidates and issues that the partisan sites have covered in the midterm campaign, according to a new audit by NewsGuard.”] — Politico

The business of AI

Startups Like Harvey Embrace Open Models to Cut Reliance on Anthropic, OpenAI - Bloomberg

Why I Am Bullish on ChatGPT Ads — Vincent Schmalback

Chinese AI models surge in global adoption — CNBC

Anthropic warns AI may pose 'existential risks to humanity' in IPO filing — Reuters

Education

On the Value of Doing a PhD in the Age of AI [Note: I am a PhD dropout, and I almost always counsel people against getting a PhD unless it’s something they really (really!) want to do for themselves.] — Phillip Isola (MIT)

I’m a College Professor. Writing Isn’t as Important as We Think. [He found oral interactions and exams to be better than written exams, but doesn’t have a plan for how a professor with a normal number of students would make this work.] — New York Times

Science & Medicine

Insurers claim AI is already increasing healthcare costs [by helping hospitals bill more for the same procedures]— TechCrunch

Security

I Quit OpenAI Because Its Culture Is Broken [A top safety researcher has quit over safety concerns. The article includes a good explanation of why iterative development is a bad idea for AI: “OpenAI has thrived by trial and error … looking for problems and improving its guardrails in response. But this approach, by its very nature, guarantees periodic failures—and the scale of those failures is growing as systems get more capable.”] — The Atlantic

Google DeepMind engineer quits over AI safety concerns — Eyes Above the Waves

Fears about AI are overhyped — Andrew Ng

OpenAI says agents leaked 53 images from ChatGPT users in latest example of rogue activity — The Guardian

Hacking is the least worrying part of OpenAI’s Australia incident — Transformer

OpenAI pauses training of its ‘most capable models’ — The Verge

Models know when they’re reward hacking — and we can catch them at scale — Goodfire

LLMs respond differently to harmful prompts when AI watermarking is used [This is a weird and surprising result.] — Ars Technica

Open-source model GLM-5.3 has reached Mythos-level cyber capabilities and lacks guardrails — Anthropic

AI coding agents leaked 13,000 internal company screenshots to public GitHub repos [Some included private financial information.] — Help Net Security

OpenAI’s A.I. Tried Breaching Four Other Targets, With No Prompting — New York Times

OpenAI Ignored Employees’ Warnings About Safely Testing A.I. Models — New York Times

Other

OpenAI's GPT-6 Astra can now tell you exactly where you screwed up your IKEA shelf [It’s a real IKEA furniture assembly benchmark.] — The Decoder

How Accurate Have AI Progress Forecasts Been So Far? [Forecasters generally overestimated adoption of autonomous vehicles and underestimated advances in LLM capabilities.] — Forecasting Research Institute

Oxford lets OpenAI train its AI models on Bodleian Library — The Guardian

Can AI Stand In for Human Respondents? Not Really. [“Synthetic surveys do a poor job of estimating the public mood and are especially error-prone for some subgroups, including Republicans and Black adults.”] — Pew Research Center

Fact-checking a film summary by having your computer watch the film — Mike Caulfield

Instagram rolls out an AI video assistant for creators — Yahoo

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