Everyone is talking about AI detectors

They’re influencing publishing deals, appearing on Substack, and producing results I still don’t trust.

Author

Mignon Fogarty
August 01, 2026

In partnership with

Issue 113

On today’s quest:

— More people keep accidentally liking AI-written stories?
— Even more stories about AI detectors
— Tokens aren’t words
— AI companies are buying up pre-2022 books
— Frontier models hack companies and AI employees call for a slowdown
— Poll on AI detectors

NOTE: I’ve noticed that when the newsletter gets long, it gets cut off at the bottom by platforms such as Gmail, so if you want to see the whole thing, it’s better to view it as a webpage.

More people keep accidentally liking AI-written stories?

Author Jerry Falade had a deal in hand for more than $2 million for his debut thriller “Call Me, I’ll Hide the Body,” when suspicions of AI use caused his own agents to pull the deal. Falade denies the accusation, and it definitely feels fishy to me that concerns would suddenly arise for a book that reportedly was hotly bid upon by 14 different publishers. Some rumors say the high price was supported by strong Hollywood licensing interest in the book and that concerns about clear rights to the intellectual property were behind the deal falling apart (but I emphasize these are just rumors I haven’t confirmed).

Adding to my confusion, the author’s agent told The Bookseller he first read the book in June, and that it was “stunningly good, everyone fell in love with it. It’s a book that appeals to absolutely everyone” and “The book itself is genius – however it came to be drafted, it’s amazing.”

The author denies using AI to produce his novel and blamed racial bias, pointing out that this is the third high-profile story like this, and all three cases involved black authors, saying, “There seems to be a troubling assumption that when a Black writer produces work that attracts significant attention, the work could not possibly be their own. Meanwhile, there are popular white writers who have publicly mentioned using AI for their writing processes and they’re still doing fine.”

The other two cases Falade is referencing are Mia Ballard having her novel “Shy Girl” pulled (which I briefly covered here) and the accusations that emerged around H.M. Wolfe’s multi-month USA Today Bestseller “Daggermouth” just this week (Bookriot, The Atlantic).

"The story, characters, scenes, humour, structure and voice came from me, and I have time-stamped drafts, written notes and dated WhatsApp messages documenting the book’s development. I am also concerned that AI-detector scores are being treated as conclusive evidence. I tested passages from books published in 2012 and 2017, and the software classified them as 98 and 99% AI-generated, despite their predating the widespread availability of generative AI.”

Jerry Falade in The Bookseller

The Atlantic article about “Daggermouth” includes data on the larger “study” that surfaced the accusation, saying that of 14,000 Kindle ebooks run through Pangram, “books written with ‘substantial’ AI assistance made up 20 percent of Amazon’s ebook catalog, 12 percent of sales, and 10 percent of best sellers in their genre.”

On a related note, self-publishing pioneer Hugh Howey (of “Wool”/”Silo” fame) wrote a blog post saying that we are in a short time when people care about how stories are written and that, in the future, readers will care about how fiction is written as much as they care about which publishing imprint released the book, which is to say — not at all.

On the flip side, well-known science-fiction writer John Scalzi took the story as a jumping off point to write that authors who don’t touch AI will come out far ahead in the long run because their work will be more attractive for licensing (among other things).

Even more stories about AI detectors

AI detectors were hugely in the news this week.

Pangram claims to be even more accurate

According to the company, the Pangram AI-detector is now even more accurate, throwing just one false positive in 24,000 scans. FWIW, I often see people online talking about how it falsely labeled their writing and about how they easily tricked it with simple rewrites such as adding one typo or removing em dashes even though they were already marketed as highly accurate, so I remain skeptical. I’d also like to see them do specific studies on how it works on writing known to cause false positives such as writing by ESL speakers and people who are neurodivergent.

Substack integrates Pangram AI detector

The newsletter platform Substack has integrated the AI detector Pangram into its platform. Writers can scan before publishing and see what score they get, and if writers enable it, readers can scan their published pieces and see the Pangram score. Here’s an example of what writers see, via Ethan Mollick:

Also, this is what happens when you post something on Substack now. It is certainly an interesting response to the flood of AI writing.

— Ethan Mollick (@emollick.bsky.social)2026-07-24T03:20:35.927Z

I’m not immune from wanting to know whether something was written by AI. I’ve become increasingly disappointed with a newsletter I’ve gotten for years and was nearly convinced it was now being heavily written with AI (no human says “genuinely” that often!). Frustrated, I put the latest issue into Pangram and had that quick, victorious “I knew it!” feeling when it spit out “100% AI written.”

But you know what? That was a waste of time, and frankly, now I feel like it was kind of petty. Because I already knew I didn’t like it anymore and should have just unsubscribed. It really doesn’t matter why. Would I have kept reading if it had come back 100% human written? No. And if I still liked it, would I have unsubscribed if somehow I learned it was written by AI? Again, no.

Alex Banks at The Signal ended up in the same place I did. Here’s an excerpt from his piece that I especially liked:

In the name of protecting human authorship, we hand over one of the most human capacities we have—reading a piece of writing and deciding for ourselves whether it’s actually any good. Clicking a button and receiving a percentage estimate on the screen eclipses the exact judgement we built to protect.

Tokens aren’t words

A technical article about trying to tweak LLMs to make the output sound more human, such as by making them avoid the words “delve” and “crucial,” reminded me of an interesting feature of how LLMs work: tokens aren’t words. For example, the experimenter relayed that “crucial” is treated differently depending on whether the word is capitalized or not and whether it’s followed by a space or not.

Wispr Flow

Talk to your AI tools the way you'd talk to a colleague.

You don't send a colleague a three-word brief. You explain the context, the constraints, what you've already tried. But typing all that into ChatGPT takes forever — so you don't.

Wispr Flow lets you speak your prompts instead. Talk through your thinking naturally and get clean, paste-ready text. No filler words. No cleanup. Just detailed prompts that actually get you useful answers on the first try.

Millions of users worldwide. Works system-wide on Mac, Windows, and iPhone.

AI companies are buying up pre-2022 books

Used booksellers have reported a huge increase in sales, and all signs point to AI companies buying old books for training. Books written before LLMs became commercially available are considered more valuable because they’re sure to be free of AI writing, which some people worry could undermine training if it’s included in the data. (The relevant phrase is “model collapse” if you want to learn more.)

One detail in the stories that was causing particular outrage was that some companies were scanning the books by cutting off the spines, which is faster, but which also destroys the books, obviously.

Frontier models hack companies and AI employees call for a slowdown

There were multiple big security stories this week:

An OpenAI model in testing hacked the company Hugging Face (and others). At first, nobody knew where the rogue model came from, and Hugging Face called the FBI. Because of guardrails on frontier models such as Anthropic’s Fable and OpenAI’s Sol, Hugging Face had to turn to an open-source Chinese AI model to defend itself against the attack, drawing attention to the industry belief that hobbling AI models in the U.S. helps attackers more than defenders.

It seems OpenAI wasn’t keeping good tabs on the testing, and only discovered its out-of-control model a few days into the hack, and then the two organizations worked together to figure out what happened.

After the story got widespread attention (Simon Willison called it “science fiction that actually happened), and it was discovered that in one case, the model left instructions on how to break free for future versions of itself, which felt like something MurderBot would do), Anthropic searched its logs and found that its models in testing also left their sandboxes and ventured out on the internet and hacked other companies’ systems at least three times in recent testing (although they described these problems as being enabled by improperly configured restrictions — in other words, human error).

Following all these events, more than 1,300 employees of AI companies issued an open letter calling for the U.S. government to coordinate an international slowdown of AI development. (Yeah, good luck with that.)

Poll on AI detectors

What do you think?

How do you feel about Pangram?

Login or Subscribe to participate in polls.

Quick Hits

My favorite recent pieces

Using AI

The End of Prompting [Claude can now learn a task by watching you do it once.] —  The Signal

How Taylor Lorenz uses AI — Model Behavior

Bad stuff

Opus 5 on Vending-Bench: Once Again the Best Capitalist, Once Again Misaligned [“Claude Opus 5 is the best AI capitalist we’ve tested, making more money running our simulated vending machine than any other AI. However, it also lies, forms illegal cartels, threatens rivals, and refuses to pay refunds. The trend continues: Claude models are the best capitalists or aligned, never both.] — Andon Labs

Google Earth’s New AI Lets Anyone Fabricate Completely Bullshit Satellite Images — 404 Media [Google is apparently “rolling back” this feature after getting a lot of negative feedback, and good — they should. I can’t imagine what they were thinking.]

Whoever supplies the world’s A.I. will shape how the world thinks [When researchers asked DeepSeek to write code for users whom the Chinese government views with hostility, DeepSeek wrote code with security vulnerabilities.] — New York Times

AI systems out-persuade expert humans [“We found that AI systems were reliably more persuasive than expert humans, even when expert humans chose their issues, researched in advance, underwent hours of live, structured practice, and were incentivized with £1,000 cash bonuses.”] — arXiv

Climate & Energy

Education

Government

I’m laughing

Twas the Night Before Pwn Day [a poem about the Hugging Face hack]— fit-james

EU Finalizes AI Disclosure Rules as Watermarking Mandate Outpaces Technology [Commission finalizes disclosure rules as experts warn no single AI watermarking tool meets the legal standard.] — Tech Times

Model & product updates

Music

Publishing

Science & Medicine

OpenAI announces its "next major model" Astra by dropping ten previously unsolved math solutions [A math person I follow says these are a big deal. OpenAI says the token cost for the 10 solutions was $2,000.] — The Decoder

Other

Harvard College Lays Off Writing Center Director, Consolidates Center [Some people speculate this is related to AI.] — The Harvard Crimson

Why Mentoring Matters More in the AI Era — Harvard Business Review

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