Publishers Bill AI $1.5 Billion as Other Newsrooms Invite It In

01A Filipino restaurant fed its menu to AI, and the adobo came back uncanny

An Austin food writer looked up a Filipino and Hawaiian bar's menu before heading over, curious the way she always is about Filipino restaurants, where the dishes run to family-style comfort food. A friend at the table had eaten there many times and liked the old menu. Then they saw the new one. The restaurant had rebuilt its menu with AI-generated plate photos, and the food no longer looked like food.

She described the images as uncanny, the plates rendered into something she didn't want to eat. She had come craving comfort food and planned to order the adobo. She pivoted to spamsilog, a $10 plate she says she could make at home if she had the patience to crisp the Spam.

What stuck with her was her own vantage point. She works with small businesses in Austin, and she reads these redesigns as ignorance, not any claim over design work. "People see button. They press," she wrote. Her friends had come straight from a doodle-date meetup, artists now watching machine images stand in for a menu a person once drew.

Her account reached the front page of Hacker News with 367 points and nearly 300 comments. Days later, a different writer made similar rounds with the opposite instinct. Instead of documenting the slop, he built something to route around it.

He had spent his first year of college shelving library books in the A-through-F section: religion, philosophy, sociology, history. To fight the tedium, he flipped each book to a random page and read one sentence. Most left him cold. A few, like a study of Hellenistic mystery cults, pulled him through several pages, and then he would read the books shelved on either side. He calls it one of the most important intellectual experiences he has had.

So he vibe-coded a tool to surface quality non-fiction, which he frames as the opposite of AI slop. The software recreates the shelf-browsing accident that taught him, pointing readers back toward books written by people.

Neither is debating whether AI will replace writers. One stopped ordering the adobo. The other wrote code to find better books. Both are sorting by hand.

02Enterprises stopped grading AI on benchmarks. Now they count the hours it removes

NTT DATA Group says it cut incident analysis to 30 minutes using Codex, and put ChatGPT Enterprise in front of 9,000 employees. Its stated goal was automating work and scaling secure adoption, not chasing model quality. What justified the rollout internally was the compressed timeline, according to OpenAI's account.

That framing shows where enterprise AI procurement has landed. Buyers have stopped asking whether a model can perform a task. They ask how much labor it subtracts, then settle the bill against that number. Incident triage that once ran for hours collapsing into a half-hour job is the kind of line item an operations budget owner can defend to finance.

OpenAI is building a product business around that same accounting. It launched Presence, a platform for deploying voice and chat agents across customer service and internal workflows. OpenAI calls the product "proven" and positions it as packaged capability, moving the company from selling model access to selling agents that deploying organizations run themselves.

The repackaging changes who OpenAI sells to and how the cost is measured. A model subscription is priced per seat or per token. An agent platform is priced against the workflow it replaces, which pulls the vendor directly into a customer's staffing math. For the deploying organization, that reframes the AI budget from an experimentation cost into a headcount-equivalent one, evaluated by the same managers who own the process being automated.

It also shifts where the burden of proof sits. NTT DATA's 30-minute figure is a claim the deploying company can measure against its own prior baseline, unlike a benchmark score generated in a lab. Support teams, incident responders, and internal operations are the first roles where that measurement gets applied, because their work already produces timestamps and ticket counts. Whether Presence spreads will depend less on model capability than on whether early deployers publish numbers like NTT DATA's, and whether those numbers hold once the automated agents run unsupervised.

03One corner of publishing billed AI $1.5 billion. Another invited it into the newsroom.

A federal judge on Monday approved a $1.5 billion settlement requiring Anthropic to pay thousands of authors about $3,000 per book. The company had trained its Claude chatbot on pirated copies of their work. District Judge Araceli Martínez-Olguín called the class-action deal "meaningful relief" for writers and publishers.

The settlement covers more than 482,000 books, and roughly 91% have already been claimed by authors or publishers now due payment. Plaintiff attorney Justin Nelson called it "the largest known copyright recovery in history." That claim rate shows how many writers lined up to collect.

The award rested on a narrow point. Judge William Alsup, who granted preliminary approval last September before retiring, had ruled last summer that training AI on copyrighted books was not illegal. Anthropic's exposure came from how it obtained the books: millions pulled from pirate sites. Its deputy general counsel, Aparna Sridhar, highlighted the fair-use finding as a landmark for the industry.

While authors extracted a record check, another part of the content business was writing AI into its own operations. OpenAI published an account of newsrooms using its tools to strengthen reporting, grow audiences, and improve business operations, the company says. The framing is a customer showcase, not an independent audit.

The split runs along incentives. Book authors hold back catalogs that were already scraped, so litigation converts past infringement into cash. News organizations run on daily deadlines and thin margins, where the same technology promises faster production and lower cost. One group is billing for what AI already took. The other is betting on what it can still deliver.

The two outcomes set different precedents. The payout affirms that pirated training data carries a price, but Alsup's ruling leaves lawful training as fair use. Future author suits will turn on how models acquire text, not whether they train on it.