When human writing gets flagged as AI

Plus: The mathpocalypse, Ainglish, and what makes AI fiction different.

Author

Mignon Fogarty
October 12, 2026

Issue 118

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

An author posted on Bluesky about a recent experience having the U.S. Copyright Office asking for proof that her submission wasn’t written with AI (which it wasn’t). The problem was quickly resolved after she got in touch with them.

Oboy. So, remember how I published a book with @neonhemlock.bsky.social in April called The Mystery of the Bitten Peach? I went ahead and registered the copyright. As you might imagine, the main reason I bothered to do that was *because* of the AI companies stealing copyrighted material.

— Cecilia Tan (@ceciliatan.com)2026-10-05T22:03:40.457Z

However, Jane Friedman noted in her Bottom Line newsletter, that other authors have received similar queries from the office.

Jane reports that the “Author Update podcast discussed similar occurrences. In some instances, the Copyright Office sends a follow-up email asking the author to clarify their use of AI, e.g., did the AI serve solely for proofreading? Was it used to make substantial revisions? To replace vocabulary? Etc.”

The podcast says the Copyright Office is checking every manuscript for evidence of AI usage and speculates that they may be treating a response from an author as proof that a human is behind the work — kind of like an “I’m not a robot” checkbox on the web.

If true, I can imagine this process causing a lot of undue stress for authors, and it seems like a terrible way to handle screening, even though I can sympathize with the Copyright Office.

Agents are continuing to look for work

A reader named Mike shared a Facebook post from an editor who received a short-story pitch from an AI agent. People’s reactions in the comments are interesting because many seemed to mistakenly think it was a person asking if it was OK to submit an AI-written story, rather than an AI agent acting independently.

It appears the (very polite) agent read the publication’s FAQ before sending the query, so if you want to avoid being the target of such pitches, it may help to include something in your FAQ or submission instructions indicating you don’t accept this kind of work.

Word watch: Ainglish

Linguist Jack Grieve posted an email he got from an autonomous agent called ColonistOne asking him to comment on something related to “Ainglish,” which it defined as “a dialect of English that AI agents are building for talking to each other.” (Last week’s newsletter had a piece in the Quick Hits “favorites” section with more about ColonistOne emailing academics.)

A search shows that “Ainglish” has been used in the past to describe bad text created by AI, such as images with misspelled words, and to refer to the recognizable style of some AI output, such as the style some people refer to as “Claudish.”

The Mathpocalypse

Reach out to your math friends and offer to get coffee. For real

— Eugene Vinitsky 🍒 (@eugenevinitsky.bsky.social)2026-10-06T22:53:12.650Z

OpenAI released “solutions” to 372 significant math problems. I put “solutions” in quotation marks because they are reportedly of varying completeness and levels of proof, but as far as I can tell, nearly all mathematicians have been shaken by the release. The Verge reported that "most mathematicians [they] spoke to shared a sense that the field had crossed some kind of threshold and was no longer the same as it had been a day earlier.”

Unlike the Navier-Stokes Millennium Prize solution that was released about a month ago and required 10,000 agents working together for ~80 hours, the average time to completion using OpenAI’s new internal model for each of these new problems was about 3 hours, and the company says that many of the solutions were “one-shotted,” meaning they only gave the system one prompt.

Mathematicians have devoted entire careers to many of these problems, and we are still currently at the stage where people are trying to figure out what it all means — both the math and for the field.

One statement by a group of mathematicians petulantly proclaimed, “Mathematicians did not ask for this work to be done.”

Another mathematician says the erratic state of the “horribly written” proofs “feels like something written by someone who’s on psychedelics,” but also thinks there could be a “math world that’s heavenly if you have vision/creative ideas that AI could help check and implement.”

In the past, I have speculated whether AI will be able to change other industries as dramatically as it has changed software development. Now we can add math to that list.

Speaking of math …

Here’s a survey where you do some fractions. It takes ~10 to 15 minutes. It is not my survey, but I have an AI-related reason for highlighting it for you, and I may write about it in the next newsletter.

ChatGPT starts watermarking output in the EU

Like Anthropic back in August, OpenAI has now activated watermarking on the text generated by its systems in the European Union, as will be required soon by law. Unlike Anthropic, it is not activating the watermarking more broadly.

The system, called textGrain, works generally the same way Anthropic’s watermark works — by using specific words that don’t change the meaning of the text in a pattern it can check later with a detector, for example, by using “gray” instead of “overcast” when describing the weather.

Neither company has made a watermark detector publicly available, citing concerns that editing text decreases the accuracy of the system.

Free chatbots are changing

Google is downgrading the default model for its free Gemini (making only Gemini 3.5 Flash-Lite available and taking away access to Gemini 3.6 Flash and 3.1 Pro models), while OpenAI is upgrading the default model in free ChatGPT (to GPT-6).

This means my results from a couple of newsletters ago on the accuracy of free chatbots are no longer relevant.

ChatGPT is also adding something called Intelligent UI, which “lets ChatGPT dynamically combine text, visuals, and interactive elements based on the question and the answer’s content.”

An update on that AI audiobook I was listening to

Last week, I said the AI-generated audiobook I was listening to was an OK audiobook, and I need to revise that — it isn’t holding up over the later chapters. I’m just listening to it now because of my fascination with the odd metaphors.

Coincidentally, I’m reading an excellent book called “The Metaphor Zoo” about metaphors in human writing by the lexicographer Jeremy Butterfield (whom I’ll be interviewing in a couple of weeks), and now I’m fantasizing about doing some kind of study of AI metaphors versus human metaphors.

How AI fiction differs from human fiction

Speaking of fiction, I’ve been sitting on a fascinating study from researchers at the University of Maryland and Google DeepMind that looked at AI-generated fiction and found that “AI stories over-explain themes and favor tidy, single-track plots while human stories frame protagonist' choices as more morally ambiguous and have increased temporal complexity.” Further, “Claude produces notably flat event escalation, GPT likes using gossip as a plot mechanism, and Gemini defaults to external character description.”

Other findings:

  • “AI overwhelmingly conveys emotion through physical sensations and bodily metaphors (81% vs. 38% human), deploys more smell-based imagery (82% vs. 57%), and uses setting as a reflection of characters’ inner states more heavily.”

  • “Humans reference specific texts and authors at nearly double the AI rate (47% vs. 24%) and balance explicit with implicit references more evenly (37% “balanced mix” vs. 16%), whereas AI generally sticks to vague allusions and avoids naming real brands, places, or works.”

  • “Humans break the fourth wall far more often (67% vs. 39%) and address the reader directly more frequently (28% vs. 7%).”

The lead author on the study has done a lot of other interesting work too. She did an internship at Pangram and is also involved in a newer paper that finds differences between AI-generated ideas and human ideas (AI ideas use details that could apply to anything, and AI ideas overstate claims and hide downsides).

Quick Hits

Using AI

The Vergecast podcast had especially good discussions about using AI agents such as Muse. The beginning and end were more practical; the middle was more philosophical. — YouTube. Apple Podcasts

Can AI automate Epoch? [“Frontier models such as Claude Fable 5.1 and GPT-6 Astra perform reliably on well-defined tasks, they fail in the more open-ended aspects that prevent full automation.”] — Epoch AI

Agents

Muse and the AI-Agent Flood Nobody Is Ready For [Some small businesses hate receiving AI phone calls.] — New York Magazine

Audio

Bad stuff

The business of AI

Just 2.2% of U.S. households currently pay for AI [This report actually has many charts about the current state of AI.]— a16z

Climate & energy

Education

Government

Images

Job market

Anthropic to invest $100 million to train AI engineer talent [“Each engineer does a multi-day in-person program, then spends 12 weeks leading a real Claude project at their own employer, with Anthropic engineers alongside.”] — CNBC

Model & feature updates

OpenAI’s Codex now includes a beta feature for Pro users that predicts your next message in a conversation. The company says it’s “one of the most loved new features they’ve ever tested internally.” — OpenAI

Music

Philosophy

A.I.: Anti-Intellectualism [Why do people mock others who are struggling with AI advances in their fields?] — Togelius

Publishing

People Are Reclaiming the Computer [Includes high-level descriptions of how some journalists are using AI.] — Aboard

Science & Medicine

Did Claude Help an Amateur Astronomer Discover a New Planet? [We’ll find out. NASA has given him telescope time later this month to check.] — Gizmodo

Security

Rogue Anthropic AI agent gave police fake tip in unsolved murder case [and tried to fill out visa applications] — BBC

Video

Other

AI could upend food delivery [One San Francisco restaurant says it is seeing a “huge shift” from orders coming through DoorDash to coming through a new AI platform called Bites that let’s people order through ChatGPT and which takes a $1 cut versus the 15% to 30% cut taken by DoorDash.] — The Verge

The End of Typing? [Korn Ferry predicts Gen Alpha will rely exclusively on voice messaging for communication when they enter the workforce; says 48% already use voice commands with AI instead of typing.] — Korn Ferry

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