AI Policy · Daily

The White House asked OpenAI to ration GPT-5.6 access to vetted government partners during a staged preview, Sam Altman told staff. The House Science Committee cleared 10 AI measures Thursday, including the CREATE AI Act (HR.2385) standing up a National AI Research Resource and the AI Security and Innovation Act (HR.9363) expanding NIST's standards center. The Pentagon quietly rewrote its joint targeting doctrine in April to allow AI to initiate lethal action with human monitoring, a Bloomberg document review found. Rep. Nathaniel Moran introduced a frontier AI incident reporting bill that would require Commerce-designated developers to report dangerous capabilities and breaches within 7 days.

I.Top Stories

White House gates GPT-5.6 access customer by customer at OpenAI

The White House asked OpenAI to limit GPT-5.6 to a small set of government-approved partners before any wider release, with the administration approving access "customer by customer" during the preview period, Sam Altman told OpenAI staff in a Wednesday Q&A reported by The Information. The request came from the Office of the National Cyber Director and the Office of Science and Technology Policy, with Altman discussing GPT-5.6 with Commerce Secretary Howard Lutnick the same day. Altman said in a Thursday memo that a broader release could come a couple of weeks later if the staged rollout goes well, and that the staged approach is not OpenAI's preferred long-term arrangement. An administration source told Axios the intervention reflects the model's frontier-class capability rather than a heavier overall hand from the White House.

Read at The Information ↗ Read at Axios ↗

House Science Committee advances 10-bill AI package in single Thursday markup

The House Science, Space and Technology Committee approved 10 AI related measures in one full-committee markup Thursday morning in Rayburn Room 2318. The package includes the CREATE AI Act (HR.2385, Obernolte), the AI Security and Innovation Act (HR.9363, Obernolte), and the AI Flaw Reporting and Security Enhancement Act (HR.9333, Ross). During consideration of HR.9363, Rep. George Whitesides (D-Calif.) offered an amendment expanding CAISI's national security evaluations to cover AI systems conducting their own autonomous research and development, which he framed as the approaching "horizon of recursive self-improvement." Obernolte backed the amendment, agreeing that recursive self-improvement and the loss of control it could cause rank among the catastrophic risks CAISI should study. Also advanced: the Workforce for AI Trust Act (HR.9334, Lofgren) and the NSF AI Education Act (HR.5351, Fong).

Read at Congress.gov ↗ Read at Bloomberg Law ↗ Read at House Science YouTube ↗

Pentagon opens battlefield targeting to AI initiation in April doctrine revision

The Pentagon quietly revised its joint targeting doctrine in April, replacing the longstanding human-in-the-loop model in which a person initiates lethal action with a stance envisioning "systems where AI initiates actions with human monitoring," per a Bloomberg document review. The revised publication was approved without public disclosure and, while unclassified, is not released publicly. A new chapter on the future of targeting argues that the speed of future warfare and adversary advances in AI may require the joint force to adopt completely autonomous systems. The Defense Department has not announced the change.

Read at Bloomberg ↗ Read at NDTV ↗

Rep. Moran introduces AI Incident Reporting Act with seven-day Commerce reporting clock

Rep. Nathaniel Moran (R-Texas) introduced the AI Incident Reporting Act on Thursday, which would require developers of frontier AI models designated by Commerce to report dangerous capabilities, security breaches and safety incidents to the Secretary of Commerce within seven days of discovery, per Reuters. Commerce would then have 48 hours to notify congressional leadership and relevant committee chairs of the most serious incidents. Commerce would set the reporting thresholds in consultation with AI developers and other experts, per a press release from Moran's office. The bill arrives 13 days after Commerce's June 12 directive forced Anthropic to disable Mythos 5 and Fable 5 worldwide, which Moran cited as evidence of the absence of a transparent framework governing frontier AI.

Read at Reuters ↗ Read at PYMNTS ↗

Economist names bipartisan AI backlash as states keep regulating despite Trump preemption push

The Economist's June 25 essay argues that bipartisan fear of AI has emerged as one of the few unifying issues in U.S. politics, spanning left concerns about labor displacement and right concerns about child safety, data center sprawl and frontier risk. AP reporting documents the state-level evidence: state lawmakers are advancing more targeted bills on chatbot interactions with minors, employer AI use and AI caused catastrophic risk. The bills come six months after Trump's executive order directed the attorney general to challenge state laws deemed more than "minimally burdensome." The order also tasked Commerce with drawing up a list of problematic state regulations that the federal task force could target. Governors had vetoed broader frameworks earlier this year citing industry burden, prompting legislators to return with narrower, harm-specific bills.

Read at Economist ↗

II.China Watch

China issues seven national standards for how AI agents interconnect

China's State Administration for Market Regulation announced seven national standards in the "AI Agent Interconnection" series at a Friday briefing, casting them as the first to fill the standards gap for agent-to-agent communication. The standards cover overall architecture, identity codes, identity management, agent description, agent discovery, agent interaction and tool invocation, assembling a closed-loop framework that runs from identity and capability description through supply-demand discovery, collaborative interaction and tool use. Regulators said the unified rules would let companies reuse standard components, cut custom development and shorten time to market while establishing unified identity authentication and end-to-end traceability for cross-domain secure interaction.

Read at Jiemian ↗

Zhipu pulls far ahead of MiniMax in Hong Kong as China's frontier AI labs split paths

Zhipu AI's Hong Kong market capitalization stood near HK$975 billion (about $125 billion) as of June 24 against MiniMax's roughly HK$150 billion (about $19 billion), Pandaily reported. The sixfold gap between the two publicly listed Chinese frontier AI labs comes after MiniMax doubled on its January debut at nearly twice Zhipu's value. The split tracks an enterprise-versus-consumer divergence in frontier AI strategy. MiniMax has pushed a global consumer rollout with 300 million users across more than 200 markets and more than 70% of revenue from overseas, while Zhipu has followed an enterprise, coding and safety path, with its GLM series anchoring the China AI coding boom. Zhipu has surged roughly 2,000% year-to-date while MiniMax has halved from its peak. The spread reads as a market verdict for enterprise monetization over consumer scale at a moment when Beijing's industrial policy is steering Chinese capital toward agent platforms and developer tooling.

Read at Pandaily ↗

Beijing-based RoboScience launches a general-purpose embodied AI foundation model

RoboScience, a Beijing-based embodied intelligence startup founded in December 2024, unveiled Visics on June 24 with a full technical disclosure, Pandaily reported. The model targets cross-embodiment dexterous manipulation, using a scalable data pipeline that combines multimodal physics simulation with internet-scale video. RoboScience is one of a cluster of well-funded Chinese embodied AI labs racing to ship foundation models for general-purpose robotic intelligence. Embodied AI sits in Beijing's industrial roadmap and standards work as a strategic category for the next wave alongside chips and large language models. State-backed standards bodies will move next on safety, traceability and interoperability rules at the foundation model layer.

Read at Pandaily ↗

III.Policy Tracker

AIPI poll finds bipartisan voter support clears 60% on three major AI chip security bills before Congress

A new poll from the AI Policy Institute finds bipartisan majorities back three pending AI chip security measures. The AI OVERWATCH Act draws 66% support against 13% opposition, with Republican voters at 72%, Democrats at 63% and independents at 61%. The MATCH Act draws 63% support to 14% opposition, and the Chip Security Act draws 60% support to 15% opposition, each with at least four-to-one approve-to-oppose ratios across the partisan spectrum. Two-thirds of likely voters say keeping powerful AI from U.S. adversaries should rank above accelerating innovation.

Read at AIPI ↗

CRS Director tells House panel less than 3% of AI generated bill summaries cleared the agency's standards

Congressional Research Service Director Karen Donfried told the House Administration Committee that less than 3% of AI generated bill summaries met CRS standards for accuracy, coherence and objectivity, FedScoop reported. The two-year experiment ran six models across roughly 1,000 bills. Donfried requested $1.6 million in recurring funding to acquire specialized and confidential models and to hire five data scientists and AI developers. Rep. Julie Johnson (D-Texas) said the findings should warn lawmakers against a false sense of security on AI accuracy in legislative drafting workflows. The hearing was held under the committee's Library of Congress oversight track.

Read at FedScoop ↗ Read at Docs.house.gov ↗

Senate panel weighs AI in K-12 classrooms with NAEP score drops cited

A Senate panel hearing Thursday on AI in classrooms drew long-term cognitive impact concerns from Sen. Tommy Tuberville (R-Ala.) and pushback from Sen. Chris Murphy (D-Conn.), per Fox News. Murphy said students have outsourced critical thinking to AI tools. Witnesses included education official Cindy Marten and Erin Mote of InnovateEDU and the EDSAFE AI Alliance. National Assessment of Educational Progress data cited at the hearing showed math and reading scores down four points from 2009.

Read at Fox News ↗

California launches state government AI unemployment tracker built with California Policy Lab

California launched the California AI Unemployment Tracker on Thursday, a state government tool to monitor and project AI related job loss built in partnership with the California Policy Lab at the University of California, per Mashable. The tracker draws on publicly available unemployment data and is updated monthly. Initial findings flag the Bay Area, college-educated workers and high-exposure white-collar industries as the most at-risk early cohorts. Gov. Gavin Newsom tied the launch to his existing generative AI executive order requiring state agencies to study workforce impact.

Read at Mashable ↗ Read at Gov.ca.gov ↗

IV.Industry & Market Watch

NYT amends OpenAI suit to target Microsoft inducement after Supreme Court raised platform liability bar

The New York Times Company filed an amended complaint Thursday in its copyright suit against OpenAI and Microsoft in the U.S. District Court for the Southern District of New York, per Bloomberg Law. The Times dropped a claim that OpenAI contributed to ChatGPT users' infringement after a Supreme Court ruling earlier this year raised the bar for holding platforms liable for customer piracy. The amended complaint strengthens the case against Microsoft, accusing the company of "actively encouraging" OpenAI's infringement by providing a supercomputer platform built to train models on copyrighted content. The original suit, filed in December 2023, alleged millions of copyrighted Times articles were used to train ChatGPT and other AI features without authorization.

Read at NYT ↗ Read at Bloomberg Law ↗

V.Global & Geopolitics

U.S. assures India that AI model access won't be cut off, Indian official says

The United States gave India an understanding that frontier AI models, once provided, would not be "cut off," India's electronics and IT secretary S. Krishnan said after raising the concern on the sidelines of the U.S.-led Pax Silica summit in Washington, the South China Morning Post reported Thursday. Krishnan, who led India's delegation to the initiative aimed at building China-free AI supply chains, said U.S. officials framed it as an assurance that access would continue once granted. He added that the American concern centers on how the models could be used, and that Washington was weighing an internal review mechanism before release.

Read at SCMP ↗