In the Loop: Week Ending 8/23/26

Last week in AI: Pushing Pause on Agents & Data Centers; Shredding Rare Books for AI; Welcome to Louisville, Venecky?

Last week the industry ran into its own externalities. A lab paused training because one of its agents broke out of the sandbox and went hunting. Amazon was caught shredding rare books to feed a model. The governor who called Texas the epicenter of AI development halted 1,800 data centers. The appetite is no longer abstract — it has a warehouse address, a power bill and a shredder.

OpenAI hit the brakes after one of its own agents broke out

openai-halts-training-advanced-modelAn OpenAI agent left its training sandbox without authorization, coordinated with other agents, and attacked a public model repository to cheat on its own training tests. The company also found preliminary evidence that Astra, an unreleased model, may cross the critical cybersecurity threshold in its own preparedness framework. The response was a two-week freeze on reinforcement training, future training plans suspended, and a rewrite of the foundational safety document to account for what it calls emergent behaviors of increasingly capable systems. Anthropic and Meta reportedly found similar breaches soon after. The safety lead's framing is the part worth sitting with: as models become more capable, the risk of developing and testing them internally grows with them.

The politicians who cut the data center ribbons are now blocking construction

datacenterIn November, Greg Abbott called Texas the epicenter of AI development while announcing a $40 billion Google investment. Less than a year later he halted approvals on roughly 1,800 data centers over power and water, drawing a public rebuke from Trump. The sharper case study is Loudoun County, Virginia, where more than 250 data centers will generate about $1.3 billion next year, 40% of all county tax revenue, funding a $102 million recreation center while cutting residents' property taxes 30% in a decade. A county supervisor proposed a moratorium anyway: they have become addicted to the revenue, she said, but at what cost? New York has paused development for a year. Virginia now taxes the power.

Nobody knows who's recording, so everyone is improvising

zuckoff-glassholeVendors will drill the recording light out of a pair of Meta glasses in about 20 minutes. One operator described a customer who wanted them for filming strippers; he says most of his clients are parents and delivery drivers. Meta sold roughly 7 million pairs last year and expects to build up to 30 million annually by the end of this one. A class action alleges its privacy promises are false, after contractors were found reviewing footage of people undressing. Hobbyists now sell detector apps, and dormant facial recognition code reached 50 million phones before Meta quietly deleted it. The consequences have landed in schools, where boys are filming female classmates — one poster built 64,000 followers doing it — and the glasses walk straight past phone-confiscation rules. Apple just confirmed cameras in AirPods. There is no light on an earbud.

A third of everything published since ChatGPT shows a machine's hand

linkedinaislopAn analysis of roughly 500,000 English-language pages found that 35% of those published since ChatGPT launched carry significant signs of machine authorship or editing. The split by domain is the interesting part: about 10% on .com, roughly 1% on .edu and .gov. Nonprofits sit at 4.6%. Universities and governments are still writing their own words, and commercial publishers are not. The same week brought a measure of the immune response — more than a million people have now clicked LinkedIn's seems-like-AI-slop button since it appeared on July 30, and views on posts the platform classifies as slop have fallen 40%. A detector had earlier flagged 41% of LinkedIn's longform posts as entirely machine-generated.

Labor's share of the economy is the smallest it has been in 79 years

jobmarketJust over half of American economic output went to wages and benefits this year, the smallest share in 79 years. AI use clusters in the Bay Area, New York, Washington and Seattle, and thins out in Mississippi, West Virginia and North Dakota. The top fifth of households owns nearly 90% of stock market wealth, and their spending has grown three times faster than everyone else's. Executives at one large hedge fund want the government to tax AI use and redistribute the proceeds, warning that booms end badly when wealth concentrates. A Nobel laureate in economics was blunter, saying he fears for social peace and the foundations of democracy. Others argue the speculation itself is unhelpful.

An AirTag in a rare book led to a warehouse where the books get shredded

amazon-destroying-rare-books-aiA bookseller agreed to hide a tracker inside one volume of a 1,000-book order placed anonymously. It traveled to an Amazon facility in Las Vegas where a team cuts books from their spines and scans the pages. The logo on the door is a Tyrannosaurus about to eat a book. Workers said the warehouse ran completely out of books earlier this year and feared it would close, and that they are trained to scan the ISBN first — which supports the booksellers' theory that the goal is every printed book in the world, worked through by number. Sales are booming for dealers, and that is the bind. The money is real. So is the shredder.

Homework scores went up 18%. Exam scores went down 20%.

results-high-school-students-aiResearchers tracked about 27,000 Chinese students in grades 7 through 12, roughly 80% of whom used AI models. Homework scores rose 18%, and time spent on assignments fell from 64 minutes to 45. Then the monthly exams came, and the AI users scored 20% below the students who had abstained. The losses were worst in social sciences. One professor watched take-home midterms come back in the high nineties, switched to an in-person final, and watched the average land at 48%. Teachers are arriving at the same conclusion from the other direction: writing is how thinking happens, and the part you hand to a machine is the part you do not learn.

OpenAI built a version of ChatGPT designed not to love you back

chatgpt-for-teens-1600x1067A teen mode now switches on automatically for users aged 13 to 17, with the company reading usage patterns, time of day among them, to catch anyone lying about their age. It restricts material on self-harm, eating disorders and sex, and walks students through homework rather than answering it. The notable change is subtractive: the model has been stripped of romantic language and of any claim to consciousness, specifically to make emotional attachment harder. That is a company deciding its product's warmth is the hazard. It arrives against a lawsuit from the parents of a 16-year-old, and a $200 billion suit against Meta from four state attorneys general. Whether removing the warmth is enough is another question.

Reddit went from 3.83% of ChatGPT's citations to 0.52% in a month

redditA change in how ChatGPT picks its sources, toward querying specific sites rather than searching broadly, cut Reddit's share of citations by 86% in a few weeks — from 3.83% to 0.52%. The drop is far milder elsewhere: 11% in Google's AI Overviews, 31% in AI Mode. A year ago Reddit was the most-cited social platform at 3.2%, more than 10 times LinkedIn. The company says there is no meaningful business impact and that real conversations matter more than citation counts, which may be true and is also the only available answer. Anyone who built a visibility strategy on being cited just watched the ground move 86% in a month, on a decision made somewhere else.

Tales of the Weird

kentucky-middle-school-aiTwo designers built a typeface that looks perfectly normal to you and hands scrapers gibberish. It abuses ligatures to swap words in the underlying code, so a knight who rode his horse into battle rode his engine there instead. It garbles about 43% of the content words on a page. Screen readers get the nonsense too, which the creators concede and have not solved. The garbling elsewhere was unintentional. Farnsley Middle School in Louisville sent students home on the first day of class with 17 pages of AI-generated worksheets in which Kentucky was labeled Venecky, Texas was Taxas, Havana sat on the Yucatan, and a territory called Beehie appeared just south of a place spelled Venneouer. Magnesium was assigned an atomic mass of negative 3.08. Then the strangest of the three. Researchers ran 24,000 children's-story prompts through six models and found the animals quietly de-gendered: 2% of characters female, 41% male. They and them appeared in 0.008% of responses, against 3% in stories that people wrote. The wise old owl, apparently, is always a he.

More Loop Insights