AI Policy · Daily

The White House is controlling access to frontier AI models, with six cabinet agencies racing to finalize an evaluation process under Trump's June executive order. The European Commission published guidelines for the AI Act's Article 50 transparency rules, which require AI providers to label machine generated content and notify users of deepfakes starting Aug. 2. Labor, NSF, and Education are jointly building a K-12 AI curriculum modeled on China's, Acting Labor Secretary Keith Sonderling told the Senate HELP Committee. Moonshot's Kimi K3 rivals top U.S. models at far lower cost, the second Chinese release in about a month to do so, per the New York Times.

I.Top Stories

White House dictates access to frontier AI models

The Trump administration is controlling who gets access to the latest frontier AI models, CNBC reported Friday, citing people familiar with the matter. Trump's June executive order gave six cabinet agencies until early August to finalize an evaluation process, and former Trump AI adviser Sriram Krishnan told the Financial Times "there will not be an FDA for AI," per TechCrunch. National Cyber Director Sean Cairncross told reporters at the process's July 2 launch that the order also covers open source model scanning and deconfliction. A senior White House official signaled Tuesday that the administration has "plenty of ongoing work" beyond the executive order, per Semafor.

Read at CNBC ↗ • Read at Semafor ↗ • Read at TechCrunch ↗

European Commission publishes Article 50 transparency guidelines with 13-day compliance runway

The European Commission on Monday published guidelines to help providers and deployers of AI systems meet the AI Act's transparency obligations, which start to apply Aug. 2, 2026. Under Article 50, AI providers must design their systems to inform users when they are directly interacting with an AI and add machine readable marks to enable detection of AI generated or manipulated content. Deployers must also inform people when they are exposed to deepfakes, to AI generated content on matters of public interest without human review or editorial control, and to emotion recognition or biometric categorization systems. The Commission released a companion Code of Practice on Transparency of AI-Generated Content and a Q&A document alongside the guidelines.

Read at European Commission ↗

Acting Labor Secretary Sonderling tells Senate HELP that DOL, NSF and Education are building China-style AI curriculum

Acting Labor Secretary Keith Sonderling testified Thursday at his Senate confirmation hearing that Labor is partnering with the National Science Foundation and Department of Education to "develop that AI curriculum that China has," per FedScoop. Sonderling told Sen. Josh Hawley (R-Mo.) that Chinese students are learning AI in the second grade while U.S. students are not, framing the effort as national competitiveness and workforce development. He said the planned DOL AI workforce hub will let the agency sign memorandums of understanding with tech companies so Labor officials know what skills the private sector needs. The Senate Health, Education, Labor and Pensions Committee is scheduled to vote on Sonderling's nomination at a business meeting July 23.

Read at FedScoop ↗ • Read at Senate HELP Committee ↗

China narrows U.S. AI lead as cheaper open-source models spread

The U.S. still leads AI by most measures, but a run of Chinese releases is shifting perception, the New York Times reported. Moonshot AI's Kimi K3, unveiled Thursday and slated to be open-sourced, is nearly as capable as Anthropic's Fable 5 at far lower cost; its debut helped push the Nasdaq down 1.4% and the S&P 500 down 1% on Friday. It was the second Chinese model in about a month to rival top U.S. systems, after Z.ai's release drove six of the 10 most popular models on the OpenRouter leaderboard to Chinese technologies. In a Friday speech, Xi Jinping cast Beijing as champion of a global AI order built on open source, entering a long-running debate over how tightly the technology should be controlled. Anthropic recently shut down its two most powerful systems after the government demanded it bar foreign nationals, restored access under tight controls, and separately accused Alibaba of using 24,000 fraudulent accounts to copy its technology.

Read at NYT ↗

Sacks says U.S. is "tying itself in knots" over AI safety regulation

Former Trump AI czar David Sacks said Friday that the U.S. is "tying itself in knots" over AI regulation and risks losing its competitive edge to China, per The Hill. Sacks, now co-chair of the President's Council of Advisors on Science and Technology, criticized politicians and bureaucrats for banning new data centers and pushing federal agencies to pre-approve frontier models, per the Washington Post. He said Trump's "permissionless innovation" approach is how America won the internet and how it can win AI, per Axios. His remarks landed the day after Moonshot's Kimi K3 took first place on the Frontend Code Arena, a coding benchmark, over Anthropic's Fable 5 and OpenAI's GPT-5.6.

Read at Washington Post ↗ • Read at The Hill ↗ • Read at Axios ↗

Obama, Biden alumni launch group to shape Democratic AI agenda

The American Innovation Policy Forum, a new policy initiative founded by former Obama and Biden officials, is working to define a center-left vision for AI ahead of 2028, the group told Axios. Founder Alex Roark, a Biden-era Federal Communications Commission official, said the forum will hold listening sessions nationwide to surface community concerns, citing Utah, where a project he called the largest data center in U.S. history was announced and then rolled back, as a case of residents left out of the conversation. Those sessions will inform policy recommendations the group's fellows begin briefing lawmakers on this fall, with the fellows to be named in coming weeks.

Read at Axios ↗

II.China Watch

Kimi K3 demand pushes Moonshot to pause new consumer subscriptions days after launch

Moonshot AI suspended new consumer subscriptions to its Kimi service 48 hours after launching Kimi K3, saying user requests pushed its existing compute cluster close to maximum capacity, per TechNode. The Beijing lab said all available computing resources will be dedicated to current subscribers to preserve service quality, and future memberships will be split into two plans, including one dedicated to coding, to better match compute to demand. The 2.8 trillion parameter open weight model has full weights set for release on July 27.

Read at TechNode ↗ • Read at Reuters ↗

Alibaba unveils new open weight AI model as Chinese labs stack up open source releases

Alibaba previewed Qwen3.8-Max, a 2.4 trillion parameter model, and said it will release the full weights under an open source license, per Pandaily. The company said the model is the strongest available after Anthropic's Fable 5. Pandaily counted four Chinese labs launching trillion parameter open weight models within roughly one month: Moonshot's K3 at 2.8 trillion, DeepSeek V4 Pro at 1.6 trillion and MiniMax M3 Pro reportedly at 2.5 to 3 trillion parameters. The wave marks a shift from earlier Chinese open releases that competed mainly on price.

Read at Pandaily ↗

China's industry ministry says domestic firms have built more than 400 humanoid robot models, over half the global total

The Ministry of Industry and Information Technology said Monday that Chinese firms have developed more than 400 humanoid robot products, more than half of all humanoid designs worldwide, per TechNode citing Xinhua (state media). The ministry also said Chinese quadruped robots account for close to 70% of global sales. The disclosure gives Beijing an official industry ministry headline stat to anchor its embodied AI push, which industry regulators have positioned as a strategic industry alongside chips and models.

Read at TechNode ↗

China exports its AI weather warning system to 30 more countries as Xi frames it as an AI public good

Chinese President Xi Jinping used his opening address at the World AI Conference in Shanghai to announce that China's MAZU meteorological early warning system will be deployed in 30 additional countries over the next five years, per Xinhua's full-text release of the keynote. Pan Jinjun, chief engineer at the China Meteorological Administration, described the system as an "international public good" providing "tailored early warning services for developing nations," per SCMP. The platform combines Fengyun satellites, ground radar and machine learning models into a single early warning stack, with customized editions already delivered to Pakistan, Ethiopia, the Solomon Islands, Jordan, Sri Lanka, Mongolia and Djibouti.

Read at SCMP ↗ • Read at Xinhua ↗

III.Policy Tracker

DHS posts updated AI inventory months past OMB deadline; CDT flags "puzzling and uneven" revisions

The Department of Homeland Security this week posted an updated AI inventory reflecting its risk management practices for high-impact use cases, months past the April 3 deadline set by the Office of Management and Budget, per FedScoop. DHS has more than 50 high-impact use cases in various stages of deployment, and the revisions include walked-back deployment statuses and sunset use cases. Center for Democracy & Technology policy counsel Tom Bowman said DHS "deserves credit for revisiting" the inventory but called the revisions "puzzling and uneven." CDT senior policy analyst Quinn Anex-Ries said the record between April and July raised questions about use cases that had previously lacked risk management information.

Read at FedScoop ↗

Campaigns are learning to rewrite what AI chatbots tell voters

Political campaigns are using "answer engine optimization" tools to shape how chatbots describe candidates, the New York Times reported Sunday. The story centers on Dustin Lloyd, a Democratic primary candidate for Missouri's state legislature, who audited his AI presence with Run for Something's CampSight tool, covered in AIPD's July 16th edition. After Lloyd added a Q&A page to his campaign site, a chatbot that had recommended his opponent began describing his small business platform as the strongest in the race. Caucus AI, a political AEO research firm, estimates at least 16 million voters already get election information from AI and found that newly published Wikipedia content can surface in chatbot answers within about 12 minutes.

Read at NYT ↗

San Francisco city attorney demands Apple and Google delete 13 AI "nudify" apps

San Francisco City Attorney David Chiu on Thursday sent cease-and-desist letters to Apple and Google demanding they remove 13 face-swapping apps that generate nonconsensual nude images using AI and stop taking a cut of the in-app payments, per Wired. The letters say the platforms have likely "made millions of dollars in fees" from apps offering nudification and should improve moderation to keep such apps out of their stores. Chiu told Wired that "generating non-consensual intimate images is illegal, harmful, and completely unacceptable." California's laws prohibit supporting services that create deepfake pornography, the letters state; Chiu's office previously took legal action against 16 deepfake websites.

Read at Wired ↗

IV.Capability & Research Watch

arXiv preprint proposes harmonized AI safety thresholds across cyber, biological and R&D risks

A preprint submitted July 17 to arXiv proposes a methodology for deriving harmonized capability thresholds across three risk domains: cyber, biological and automated AI R&D. For misuse risks, the authors use expected harm as the key primitive with an explicit risk modeling approach that accounts for risk channels and model release conditions. For automated AI R&D, they base the proposed threshold on the observed rate of AI progress rather than expected harm. The paper said that without common minimum thresholds, risk mitigation may be inconsistent across labs, creating a potential race to the bottom in safety standards.

Read at arXiv ↗

AI Alignment Forum post from ex-DeepMind employee proposes red-line framework for government AI contracts

A framework published July 18 on AI Alignment Forum by a former Google DeepMind employee proposes two narrow standards for AI provided to government entities exercising coercive authority: human control over targeting and use of force, and a bar on untargeted AI profiling of persons not already under specific investigation. The first standard would rule out AI systems that select and engage targets for force without appropriate human control, while carving out anti-munition defensive systems, intelligence analysis subject to the second standard, logistics and R&D. The framework proposes an internal Review Body that assesses contracts against the standards and tracks leadership overrides in a yearly transparency report to all AI employees, while lacking authority to block contracts itself.

Read at AI Alignment Forum ↗

V.Industry & Market Watch

Pentagon uses surging budget to fund defense tech startups alongside legacy primes

Defense Secretary Pete Hegseth is using the Pentagon's surging budget to keep his promise to fund high-tech defense startups while also awarding ever-larger contracts to traditional prime contractors, the Wall Street Journal reported Monday. More than a year into his tenure, Hegseth has rewritten procurement rules and publicly criticized the legacy acquisition system as "slow and bloated." Venture capitalists and startup founders anticipate what they hope is a defense reformation that hands billions of dollars to defense tech startups previously excluded from the inner circle of weapons procurement.

Read at WSJ ↗

SpaceX in talks to sell Pentagon billions in AI compute capacity

SpaceX is in talks with the Defense Department about providing access to data center capacity worth billions of dollars for running AI models, deepening ties between Elon Musk's company and the Pentagon, the Wall Street Journal reported Friday. SpaceX employees have discussed plans to compete more directly with neocloud firms (upstart AI cloud providers) such as CoreWeave by selling computing capacity to AI customers at lower prices, per Reuters. The two sides are discussing an arrangement worth up to several billion dollars of compute capacity; the talks are ongoing and could still fall apart.

Read at WSJ ↗ • Read at Business Standard ↗

Meta and Anthropic in early talks for compute lease worth up to $10 billion over two years

Meta is in talks to lease AI computing capacity to Anthropic in a potential deal worth up to $10 billion over two years, the New York Times reported Friday, citing three people with knowledge of the discussions. Anthropic proposed the deal in June, and Meta is considering it despite not currently operating a compute-selling business, per Reuters. Anthropic already has multibillion-dollar compute deals with Google, SpaceX, Microsoft and Amazon, and SpaceX gave Anthropic access to its Colossus 1 data center in Memphis, Tennessee, under a May agreement, per CNN. Meta CEO Mark Zuckerberg told shareholders in May that entering cloud computing was "definitely on the table" and companies were approaching Meta "almost every week" to buy AI compute or model access.

Read at NYT ↗ • Read at Investing.com ↗ • Read at KTVZ ↗

Bezos and UK sovereign AI fund co-invest in Cambridge startup CuspAI at $2.6 billion valuation

Cambridge-based CuspAI raised $450 million (£330 million) in Series B funding from investors including Jeff Bezos and the UK government's sovereign AI fund, valuing the two-year-old startup at $2.6 billion, The Guardian reported Monday. Kleiner Perkins co-led the round. The company launched the AI Materials Foundry, a coalition of more than 48 tech, industrial and research firms including Nvidia, Meta and Hyundai, to build software for chipmakers to discover new materials. CuspAI aims to cut research times and reduce or eliminate use of rare metals such as iridium and ruthenium in chipmakers' supply chains.

Read at The Guardian ↗

VI.Global & Geopolitics

Incoming UK PM Burnham draws up plans to abolish DSIT, triggering tech industry backlash

Andy Burnham, the UK's incoming prime minister, has instructed officials to draw up plans to abolish the Department for Science, Innovation and Technology as part of a wider Whitehall reorganization, The Guardian reported Saturday. The plans have not been signed off but have triggered pushback from MPs, Whitehall officials and industry. Matt Clifford, an adviser on AI to outgoing PM Keir Starmer and his predecessor Rishi Sunak, called the move "a big mistake" that would tie up senior science and tech officials on reorganization when tech policy is under strain. Burnham's advisers have asked officials to hand much of science and technology policy to a more powerful business department led by the chief whip.

Read at The Guardian ↗

Rubio says Xi's U.S. visit remains on track after Trump's election fraud address

Secretary of State Marco Rubio said Monday that Chinese President Xi Jinping's upcoming U.S. trip is still on track, per Bloomberg. The confirmation follows Trump's Thursday night primetime address reviving claims of Chinese election interference, which had thrown the visit into doubt. The Xi visit is the diplomatic anchor of the current U.S.-China trade and technology truce that has kept Commerce from adding CXMT, a Chinese memory chipmaker, to the Entity List, the U.S. export blacklist.

Read at Bloomberg ↗

Australia sets rules for AI in agency decision-making ahead of Labor conference

Australia laid out consumer protection priorities for AI including a promise to better regulate the use of AI in federal agencies like Centrelink and Services Australia, the country's welfare and government service agencies, per ABC News. The priorities also include strengthening privacy laws and revisiting personal data protection statutes. AI is set as a major focus at the Labor Party's National Conference in Adelaide later this week.

Read at ABC News ↗

Singapore's IMDA publishes transparency infocard standard for AI chatbots

The Infocomm Media Development Authority, Singapore's telecom and media regulator, on Monday published voluntary guidelines encouraging generative AI chatbot providers to display standardized info cards to users, per Singapore Business Review. The info card should state what the chatbot can and cannot do, its reliability and safety, how data is handled, and how to raise concerns or report issues, and it should be written in plain language. Six organizations including Google, Meta and Singapore's DBS bank said they will use the guidelines to strengthen their transparency practices over the next six to 12 months.

Read at Singapore Business Review ↗ • Read at Xinhua ↗