Frontier Model Gatekeeping | AI Field Notes #57

AnalysisA frontier model now reaches paying users only after a federal office signs off on each one. OpenAI put GPT-5.6 into limited preview on June 26 in three tiers, Sol for hard problems, Terra for everyday use, and Luna for cheap high-volume calls, with Sol priced at $5 per million input tokens. Then the Office of the National Cyber Director and the Office of Science and Technology Policy capped early access at roughly 20 vetted partners. Executive Order 14409, signed June 2, insists it creates no licensing requirement. Sam Altman told staff the government approves access one customer at a time, which is a license wearing a plainer coat.

AnalysisApple stopped chasing the assistant crown and started threading AI through the dull corners of the phone. iOS 27 spreads machine learning across small daily chores: splitting a restaurant bill, replacing a breached password on its own, drafting message replies, turning a sentence into a calendar entry, grouping smart-home alerts, sorting Safari tabs. None of it makes a keynote highlight reel. That is deliberate. After two years of Siri stumbles, Apple is wagering that a dozen features people actually use beat one that dazzles on stage. The plan concedes the chatbot race and fights on the ground Apple still owns, the device already in your hand.

AnalysisTelling a chatbot your problems is now legally separate from getting therapy, at least in Rhode Island. Governor Dan McKee signed a measure this week barring AI from providing therapy unless a licensed professional delivers it, part of a package that also forces chatbots to say they are not human and to run self-harm prevention protocols. A companion law makes healthcare providers tell patients when AI wrote their clinical notes. It follows a year of reports about people leaning on general-purpose chatbots for mental health support the tools were never built to give. The state drew its line at the word 'licensed.'

AnalysisA bill on Gavin Newsom's desk would put into law that a teacher has to be a person. California's AB 2148 says any elementary or secondary public school employee, or a contractor doing that job, must be 'a natural person,' shutting the door on districts quietly swapping instructors for chatbots to cover shortages. The Assembly passed it 76 to 0 on May 4 and the Senate 38 to 0 on June 18 before it reached the governor on June 24. Unanimous votes on an AI restriction are rare. The unsettling part is that the question needed a statute at all.

AnalysisStealing a rival's model now looks like 25,000 sham accounts quietly milking it for answers. Anthropic says operators linked to Alibaba ran that many fake accounts to pull roughly 28.8 million exchanges out of an early Claude model between April 22 and June 5, a campaign aimed at distillation, training a cheaper model on a pricier one's outputs. It is the same charge Anthropic has aimed at DeepSeek, Moonshot, and MiniMax, all Chinese labs racing to close the gap for far less money. The fight reframes a frontier model's real moat: the answers it gives are now the training data someone else wants most.

AnalysisA one-person newsletter can now decide, bot by bot, whether an AI company gets to read its work. Cloudflare and beehiiv, the newsletter platform that competes with Substack, rolled out joint controls on June 26 that show publishers which AI crawlers are hitting their pages and let them block or wave through each one. It is the same gatekeeping Cloudflare already sells big sites, pushed down to a solo writer. The timing matters: model makers are scraping harder for fresh text while paying less for it. Putting a newsroom-grade crawler dashboard in one writer's hands nudges the leverage over training data, slightly, back toward the people who make it.

AnalysisMost people meeting a customer-service bot are already trying to get past it. In a Parloa survey of 1,001 US adults, more than half said they actively try to bypass the chatbot, 43.9% reported repeating 'human' or 'person' to force a handoff, and just 13.6% trust AI with a complex request. Many give up after three minutes. The numbers land while companies race to replace call centers with exactly these systems to cut headcount. The gap between what firms are deploying and what customers will tolerate is the whole story, and right now the customers are losing.

AnalysisThe promise was that AI would write the code; the bill is arriving as senior engineers spending their days fixing what it wrote. Deedy Das, a partner at the venture firm Menlo Ventures, describes 'workslop,' low-quality machine-generated code that floods review queues and lands on the most experienced people to untangle. One company is reportedly spending around $500 million a month on Claude alone. Meta has started tying AI tool use to performance reviews, making the tools mandatory rather than optional. At Meta's @Scale conference, Claude Code creator Boris Cherny demoed 'Loops,' agents prompting other agents to write code with no human in between. The cleanup did not vanish. It moved up the seniority ladder.