Good morning. It's Friday 21, and we're covering the White House’s murky AI testing plans, the political backlash over data centers, and Meta’s push into AI-powered game creation.

Plus: yesterday’s poll results on AI ads and a Friday Fact on @.

Ramp is turning its expense-management expertise toward AI inference with Router, a new service for switching among major language models through one API. The U.S.-only service offers models from OpenAI, Anthropic, DeepSeek, xAI and others, with routing based on benchmarks, cost or difficulty. It is free through 2026 and includes a $26 launch credit.

AI is now the most cited reason for U.S. job cuts, but CEOs are becoming more cautious about saying so. Etsy, Patreon and Microsoft all recently stressed that layoffs were not directly driven by AI while acknowledging it is reshaping required skills. Challenger says employers attributed nearly 113,000 cuts to AI this year, or 24% of announced reductions.

More than one-third of web pages published since ChatGPT’s launch show signs of AI authorship, according to Pew Research. An analysis of nearly 500,000 English-language pages found 35% of post-November 2022 pages showed significant AI writing or editing. AI signals were about 10 times more common on .com sites than .edu or .gov domains.

OpenAI launched an Apple Messages plugin for ChatGPT that lets users connect their inbox to the chatbot for sorting, editing, deleting, and drafting or sending messages on their behalf, with a promo clip showing ChatGPT suggesting follow-ups based on yesterday's inbound texts. Privacy questions came immediately: OpenAI told Bloomberg the plugin runs locally and "doesn't create an index of all someone's messages."

The @ symbol had no name in most of the world until email gave it one, and different languages settled on wildly different versions. Answer at the bottom.

The White House briefed OpenAI, Anthropic and Google on August 4 about a voluntary framework for testing frontier AI models before public release, but has not publicly released the framework or distributed it to attendees. Companies were shown paper copies during the private meeting and allowed only to take notes, leaving unresolved questions about which models qualify and how testing will work.

Officials reportedly said the framework currently applies only to closed-source models, though the written language appears less explicit about open models. The uncertainty comes as the Trump administration reshapes AI policy and some companies, including OpenAI, are independently slowing model development over safety concerns. → Read the full article here. (Paywall)

Opposition to AI data centers is emerging as a bipartisan political issue less than three months before the 2026 midterm elections, with campaigns and companies increasingly focusing on their energy and water demands. The issue featured prominently in Florida’s Republican gubernatorial primary, where Rep. Byron Donalds won, and in Pennsylvania, where Gov.

Josh Shapiro signed an executive order imposing tougher standards on data center development. Beverage brands Liquid Death and Garage Beer also turned the backlash into a satirical ad about water consumption, featuring Jason Kelce. In Ohio, data centers have become a central issue in the Senate race, prompting the National Republican Senatorial Committee to warn that voter opposition could spread nationally. → Read the full article here.

As AI systems become more human-like, debate is intensifying over whether they could eventually develop consciousness—or simply convince people that they have it. Researchers are exploring models with features associated with human cognition, while companies are designing systems that can display introspection and emotional understanding.

The distinction matters because people may increasingly treat AI companions as beings with rights and welfare, even without evidence of subjective experience. The article argues that granting legal or moral rights to potentially superintelligent AI could create serious risks, while acknowledging that advanced AI could challenge human control regardless of whether machines become conscious. → Read the full article here. (Paywall)

11 use cases for Grok Bot, from email and calendars to coding, shopping, meetings, school tasks, and everyday life.

Mark Cuban Wants Doctors Teaching AI: Mark Cuban says doctors should help patients use AI alongside treatment.

Berkeley Professor Admits AI Editing: Professor used AI to edit an op-ed criticizing students’ math skills.

Binance Lets AI Agents Trade: Binance’s Agent OS lets AI tools analyze markets and execute trades.

SpaceX Approached Cognition (Paywall): SpaceX sought to buy AI coding startup Cognition.

Meta AI Launches Mac App: Meta’s Mac app adds system-wide dictation and screen-aware answers.

Meta Becomes Major Microsoft AI Customer (Paywall): Meta spends hundreds of millions yearly on AI models through Microsoft Azure.

Before 1971, @ was a mostly forgotten commerce symbol meaning "at the rate of." Then Ray Tomlinson picked it to separate usernames from hosts in the first email address, and suddenly every language on Earth needed a word for a squiggle they'd rarely had to say out loud.

Most countries reached for animals or food:

Same character in every email address. Depending on where you are, it's a snail, a duckling, a spider monkey, or a pickled fish.

Here’s how you voted: Keeping ads out of AI won, with 60% rejecting ads entirely, while 28% accepted clearly labeled ads outside the answer.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team