The headslap that saved me from tedious work

How one change turned a week-long task into a 39-hour background job.

Author

Mignon Fogarty
July 10, 2026

Issue 112

On today’s quest:

— It’s hot model summer
— Tip: get more from AI by changing your thinking
— How many people are using AI to generate fiction?
— People keep accidentally liking AI-generated stories
— Clarify ‘hallucination’
— AI did a good job summarizing bills
— Copilot uses stereotypes to categorize identical data

It’s hot model summer

Fable 5 is back (and it’s as good as I remembered), OpenAI just launched ChatGPT-5.6 (a Fable-level model), and other groups are launching powerful models, too.

Plus, rumors are that both Anthropic and OpenAI will have newer, even more advanced models within four weeks — maybe even by the end of July. However, it’s likely the U.S. government will be reviewing all advanced models from U.S. companies before they are released now, so it’s not clear whether the models will just be “ready” or actually available.

Also, since we’re talking (peripherally) about national security, China is making noises that it might keep its powerful open-source models for itself, which is interesting because we had a couple of weeks of heavy chatter about moving to open-source models to protect yourself from the vagaries of U.S. government restrictions — but now it seems that may not be as helpful as people hoped.

Tip: get more from AI by changing your thinking

I was thinking about a project wrong, and understanding my mistake may help you too.

Background: I use AWS Glacier Deep Archive as a backup of last resort because it’s incredibly cheap — $1 per terabyte per month — but you can’t access your files on demand, and the user interface is terrible. It’s the backup you’ll use if your house burns down and all your other backups fail.

Problem: I uploaded files in spurts and lost track of what I had backed up and what I hadn’t, and because of the previously mentioned terrible interface and the sheer number of files, it’s hard to tell what’s going on just by looking.

My First Idea: It occurred to me that maybe Claude Cowork could figure out what I still needed to back up. And yes! Claude quickly made inventory lists of the files on AWS and on my hard drive, compared them, identified what was missing, and gave me a list of folders and files that still needed to be backed up. Hooray!

Implementing the plan: I began going down the list finding the files and folders that still needed to be updated, dragging them over to AWS, and updating the file list as I went. I was doing the project in batches, and it was distracting, annoying, and probably would have taken me a week.

Isn’t This What Agents Are For? Then, while reading about agents, a question hit me: why did I let Cowork give me a to-do list? Isn’t it supposed to be the other way around now?

So I asked Cowork if it could do this project for me, and yes, of course it could <headslap>. In fact, it’s not even really an AI project. It wrote a Python script for my computer to do it from the Terminal. (Past readers may not find this surprising. The solution to my problems always seems to be a Python script.) It ran for 39 hours in the background, and now all my files are backed up.

Why I need to change my thinking: Because of the way I asked, Cowork didn’t tell me the easy way to do the project — I asked if it could help me find and track the files. I didn’t ask if it could do the work.

It’s hard to break out of a lifetime of thinking of computers as having limited capabilities and needing specific instructions.

Remember to start by explaining your problem and then asking an agent system like Claude Cowork or ChatGPT Work if it can solve the big picture problem (I need the files I haven’t backed up added to AWS), not the small problem (what are the files I haven’t backed up yet).

Just how many people are using AI to generate fiction?

The banner headline you may have heard — that more than 30% of all AI chats are for generating fiction — is inflated by a few outliers. (Only 2% of users accounted for 80% of all fiction conversations!) But the more realistic number from a new study is still quite high.

The study looked at interactions in a corpus called WildChat built of chats from users who traded free access to models in exchange for allowing their data to be collected and found that 7% of all users were generating fiction — with a big chunk of that being fanfiction and erotica.

People creating fiction in WildChat also often engaged in repetitive behavior, asking for many iterations of the same story. The researchers argue that such repetition isn’t completely new because some people are known to listen to the same audiobooks on repeat, watch the same movies over and over, repeatedly visit the same cities, and so on. The researchers ponder whether a new kind of reader is emerging who now enjoys exploring stories on their own, through “on-demand, personalized, and repetitive cultural forms.”

People keep accidentally liking AI-generated stories

The steady trickle of anecdotes in which editors unknowingly accept AI-generated stories and award panels unknowingly choose AI-generated stories as winners continues. This week, the volunteer editors of a very small press who were creating an anthology published a painstaking description of the process they went through to try to determine whether two of their choices were written by AI after they were alerted by other publications who had concerns about specific authors — new authors without any previous online presence. The editors had chosen 20 stories out of 606 submissions for their anthology, and two of those choices turned out to be written by AI. Further analysis revealed that three long-listed pieces were also likely written by AI.

The editors went through some of the most extensive verification I’ve seen — asking for evidence of a writing process (such as notes or earlier drafts), having email exchanges discussing background (such as motivations for the work), and doing face-to-face interviews to discuss the work in depth. They rightly bemoan the lost time spent on this process and their previous naivety in feeling confident they would always be able to just tell when something wasn’t written by a human.

They wrote, “We used to think AI-generated fiction would always be obvious, and we were not prepared. Over the coming years, agents, editors, and slush readers at every level are going to need to educate themselves on how AI writes.” I will add that how AI writes continues to change and is likely to become even more difficult to identify in the future.

After writing this, I came across a thread from another editor who had also published (and unpublished) an AI-generated short story by one of the same people identified by the small press above (albeit using a different name).

And another post from a different small press that is nixing two now-deemed-AI-generated stories.

Clarify ‘hallucination’

I've never been opposed to the word "hallucinating" for describing how AI makes mistakes ... until now. The new problem? I just talked to someone who thought AI hallucinations would be obvious because it would be obvious if you talked to a person who was hallucinating. (Mental health experts say this isn’t true, but it’s a common misconception.)

In other words, the person I talked to equated "hallucination" with "sounds wacko" and accepted AI output as true because it sounded level-headed.

The word "hallucination" isn't going away — it's a widely used industry term — but I now realize we need to explain it better for beginners, with something like this:

"Hallucination" is just a fancy word for "produces things that aren’t true," and when AI does so, it does so confidently.

AI did a good job summarizing bills

The Congressional Research Service reports that the best available AI models did a perfect job summarizing bills from the U.S. 2025-2026 legislative session. Open-source models did almost as well for a fraction of the cost.

Three interesting things about the study:

  1. They included summaries written by humans, which you don’t often see in this kind of side-by-side study — and the human summaries had more errors than the LLM-generated summaries (although the number was still very small).

  2. All the results are available to view yourself on the website — another thing you don’t always see.

  3. The scoring was done by Claude Opus 4.8, and multiple studies have shown that models prefer their own writing. However, since all the commercial models got perfect scores, so it’s hard to say Claude was scoring itself more favorably.

Copilot uses stereotypes to categorize identical data

Adam Kucharski asked Microsoft Copilot to “look at differences in how people in the US and UK expressed emotions in an Excel dataset that contained thousands of survey responses,” and Copilot found what you might expect — that Americans were more straightforward, Britons more restrained, and so on. The problem is that the responses that supposedly came from the two groups were actually identical. Copilot had used stereotypes to rate the data.

I admit that my gut reaction was that Copilot just sucks and the best models would never do that, but Kucharski makes good arguments for why that’s a simplistic reaction and this is a real problem. (h/t Catharine Cellier-Smart of Smart Translate)

Quick Hits

My favorite recent pieces

Why is Meta destroying its engineering organization? [Jaw-dropping details about how Meta worked in the past and what’s going on there now.] — The Pragmatic Engineer

How agents are transforming work [Fascinating stats.] — OpenAI

The twilight of the chatbots [Why we’re in an exponential growth phase for AI and what it means.] — One Useful Thing (Ethan Mollick)

How tech workers are feeling in 2026: a workforce splitting in two [Lots of interesting findings.] — Lenny’s Newsletter

Using AI

11 Underrated Claude Desktop Features [Intermediate: Super useful tips for using Claude Code in the desktop app.] — Why Try AI

OpenAI's Codex can now watch you work once and repeat the task forever [Beginner: This looks like it will be very useful if you have well-defined repetitive tasks. It may be Mac-only to start. Claude has a similar tool that is limited to tasks in the Chrome browser.] — The Decoder

Tagging Motel Noir. An example of how LLM projects find depth, and how people find curiosity. [Beginner: A little more conceptual than a “how to.”] — Mike Caulfield

Setting up scheduled tasks and goals [Intermediate: Very clear.] — Why Try AI?

How to use Fable as an orchestrator model and save on token use [Advanced: Looks very useful if you are frequently running out of Fable tokens or use costly models outside of a monthly plan.] — Rundown AI

Audio

Bad stuff

The business of AI

ChatGPT's market share slips below 50% for first time [Users are moving to Gemini and Claude.] — TechCrunch

Education

Brown Professor Suspects Most of His Class Used AI to Cheat [He gave a take-home exam, and it’s quite clear nearly everyone cheated, which isn’t surprising. The bigger takeaway is that universities need to adapt their academic integrity processes to deal with cases of such widespread cheating.] — Inside Higher Ed

Government

Job market

Model & Product updates

Introducing Claude Tag [Use Claude in Slack. For Enterprise and Teams. Anthropic says they now use this extensively internally. Short YouTube video showing how it works.] — Anthropic

Introducing GPT-Live [Natural, overlapping conversations.] — OpenAI

ChatGPT Work with GPT-5.6 [At first glance, this essentially looks like Claude Cowork, but for the OpenAI ecosystem.] — OpenAI

Philosophy

Publishing

Robotics

Better Than a Robot Arm? Why I Built a Crane Robot to clean my house [I wouldn’t want this in my house, but it was a charming 6-minute video of a DIY project.] — Over Engineer on YouTube

Project Fetch: Phase two [Claude helped operate a robotic dog.] — Anthropic

Science & Medicine

Using AI to help physicians diagnose rare genetic diseases affecting children [ChatGPT o3 Deep Research provided leads that led to diagnosis of rare diseases in 4.8% of previously undiagnosable cases.] — OpenAI

Is AI ruining our skills? Early results are in — and they’re not good [Endoscopists who had access to AI tools on some days and not others performed worse than their previous baseline on days they did not have access to these tools.] — Nature

Climate & Energy

Audio

Other

77% of small businesses now use AI regularly. [In the U.S., the most common applications are marketing (45% of businesses using AI), customer service (37%), and bookkeeping (35%).] — Miami Herald

Germany's media rocked by AI scandal. Two German outlets were forced to delete articles that used undisclosed AI. Many fear that reliance on the technology will damage the credibility of German media. — DW

AI Self-preferencing in Algorithmic Hiring: Empirical Evidence and Insights [AI systems strongly prefer their own output — something to keep in mind when using AI systems to rank or select writing.] — arXiv

Inviting hard questions [Anthropic wants to know your concerns about AI.] — Anthropic

An OpenAI model crushed top human programmers at a world coding competition [Even in one category where they weren’t good enough to compete at all last year.] — Understanding AI

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