AI Policy · Daily

Meta contractors posed as minors to prompt rival chatbots including Gemini and ChatGPT on suicide, sex and drugs for competitive benchmarking. California Gov. Gavin Newsom unveiled a partnership with Anthropic giving state agencies access to Claude at half the standard commercial rate, the first such state-level procurement deal. The House passed the Kids Internet and Digital Safety Act 267-117, its first floor vote on an online child safety package, requiring AI chatbots to tell users they aren't human and platforms to curb addictive features for minors. China's state security ministry warned that AR games may feed map data to foreign military AI, pointing to reporting on Pokemon Go maker Niantic, whose spinout used billions of player scans to train a 3D navigation model now tied to a U.S. defense contractor.

I.Top Stories

Hundreds of Meta contractors posed as teens to probe rival chatbots on suicide, sex and drugs

Hundreds of contractors working on a Meta benchmarking project pretended to be minors and prompted Gemini, ChatGPT and other rival chatbots on high-risk topics including suicide, sex and drug use, Wired reported Monday after reviewing internal documents and interviewing multiple contractors. The contractors were instructed to elicit the most damaging possible responses from competing systems for comparative scoring against Meta's own products. The project, code-named Cannes and managed by Meta contractor Covalen, was active as recently as April 21 and ran more than 45,000 prompts in a single August 2025 testing round, per Wired.

Read at Wired ↗

California signs Anthropic Claude deal at 50% of standard pricing, first U.S. state agreement of its kind

Gov. Gavin Newsom (D) and Anthropic on Monday announced a partnership giving all California state agencies and local governments access to Claude at 50% of standard commercial pricing, along with training and support from the company, the governor's office said. The state said Claude will help employees draft documents and analyze information, and Newsom said AI "should not replace the human work of government." The deal — the first formal state-level Claude procurement, per TechCrunch — follows Newsom's March executive order to speed government AI adoption while tightening safety standards. California's technology director, Chris Given, told POLITICO the risk designation "just didn't come up" during contract talks.

Read at POLITICO ↗ • Read at TechCrunch ↗ • Read at Governor of California ↗

2026 midterm campaigns route door-knock memos and voter data through AI synthesis pipelines

Both parties' 2026 midterm campaigns are running door-knock conversations and field-organizer notes through AI synthesis tools to direct strategy and write custom voter outreach, the New York Times reported in a Monday field piece. Democratic-aligned canvassing group Swing Left has canvassers in PA-10 record voter conversations into an app that feeds AI tools, which then synthesize hundreds of memos into targeted strategy guidance for campaign managers. AI generated images have become the public face of the cycle. The same back-end systems analyze voter data, craft materials and write personalized messages.

Read at NYT ↗

Palantir launches federal "intelligent engine" on NVIDIA Nemotron open models for classified, air-gapped agency use

Palantir on Monday introduced an AI product built on NVIDIA's Nemotron open-weightsOpen-weight modelAn AI model whose trained parameters are published so anyone can download, run and modify it, as opposed to one reachable only through the developer's API. The main policy tension is that open weights spread capability widely and cannot be recalled once released. models, designed to run in sovereign, air-gapped environments isolated from public networks, the companies said. The "intelligent engine" pairs NVIDIA's accelerated computing and model weights with Palantir's AIP, Ontology, Foundry and Apollo platforms, letting government agencies and critical infrastructure operators — including classified Intelligence Community and Defense Department customers — train AI on their own data, retain ownership of the resulting models and deploy them in air-gapped settings while controlling proprietary data and weights. Co-founder and CEO Alex Karp said the approach lets agencies use large language models without "proprietary insights migrating into the weights of closed models."

Read at Yahoo Finance ↗ • Read at NVIDIA ↗

House passes Kids Internet and Digital Safety Act 267-117

The House on Monday passed the Kids Internet and Digital Safety Act by a 267-117 bipartisan vote, the first time the chamber has cleared an online child safety package on its floor. The bill requires online platforms to offer tools that limit addictive features for minors and adopt policies addressing sexual exploitation, requires AI chatbots to disclose to users that they are not human, and mandates age verification for accounts viewing pornographic content, per Reuters. The Senate passed the older Kids Online Safety Act 91-3 in 2024, and Sen. Marsha Blackburn (R-Tenn.) has been negotiating a stricter Senate package with the White House. The bill now goes to the Senate.

Read at The Hill ↗ • Read at Reuters ↗

Supreme Court ends Humphrey's Executor protections for FTC, FCC and NLRB, spares Fed governor Cook

The Supreme Court on Monday ruled 6-3 in Trump v. Slaughter that for-cause removal protections for Federal Trade Commission members violate the separation of powers, overturning the 1935 Humphrey's Executor precedent that has shielded leaders of roughly two dozen independent agencies including the FCC, NLRB, Consumer Product Safety Commission and Merit Systems Protection Board. In a separate 5-4 ruling, Chief Justice John Roberts and Justice Brett Kavanaugh joined the three liberals to block Trump's at-will removal of Federal Reserve Governor Lisa Cook. The majority held the central bank's for-cause protections survive because its governors serve 14-year staggered terms tied to monetary independence. Justice Sonia Sotomayor read a rare dissent from the bench warning the majority grants presidents "unbridled authority" over agency commissioners. Trump on Truth Social pledged to "take appropriate action immediately" against Cook.

Read at NYT ↗ • Read at The Washington Post ↗ • Read at CBS News ↗

II.China Watch

China's state security ministry warns AR mobile games are feeding civilian map data into foreign military AI training

The Ministry of State Security on Monday cautioned that an unnamed AI company with overseas defense ties is collecting environmental scans through an augmented reality game and using them to train spatial AI models with potential battlefield applications, the South China Morning Post reported. The post cited concerns over "the militarisation of civilian data," pointing to roughly 30 billion player-uploaded scans being used to build navigation models for robots, autonomous systems and drones. The description tracks with recent reporting on Niantic, the U.S. maker of Pokemon Go, and its spinout Niantic Spatial, which is partnering with a U.S. defense firm on drone navigation in GPS-denied environments. The warning extends Beijing's reading of data security beyond financial and identity data into spatial AI training corpora, foreshadowing tighter enforcement against foreign app and game data collection under China's Data Security Law and counterespionage rules.

Read at SCMP ↗

CXMT lands $2.94 billion Tencent DRAM supply deal ahead of Shanghai STAR Market IPO

Chinese memory chipmaker ChangXin Memory Technologies has signed a multi-year agreement worth more than 20 billion yuan ($2.94 billion) with Tencent Holdings to supply DRAMDRAMThe working memory that phones, PCs, and AI servers use to hold data that a processor is actively handling. for the internet giant's server business, Reuters reported. Sources put the term at three to five years, making it one of the largest procurement commitments in recent years between a domestic Chinese chipmaker and a major internet platform. CXMT is preparing to list on Shanghai's STAR Market and has been approved for addition to the Commerce Department's Entity List, which restricts U.S. technology exports to the company, but has not yet been formally added. The contract signals a maturing domestic memory supply chain that pulls forward Beijing's chip self-sufficiency drive, just as AI driven demand squeezes global DRAM supply and U.S. export controls make domestic sourcing strategically necessary for Chinese cloud operators.

Read at Reuters ↗

Huawei open-sources Pangu flagship to lock developers into its domestic AI chip stack

Huawei on Tuesday released its openPangu 2.0 Flash model under an open license on GitCode, the first of seven planned components in a phased rollout through year end, Pandaily reported. The 92 billion parameter model is optimized for inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. on Huawei's Ascend AI accelerators, with a larger 505 billion parameter Pro variant slated for July. The package pairs model weights with training operators and pretraining code, scoped to let Chinese developers build without the dominant foreign GPU programming toolchain. The strategy gives Beijing a tightly coupled domestic model and silicon offering and hardens China's AI base against further U.S. Commerce Department restrictions on advanced AI accelerators and the design tooling needed to substitute for them.

Read at Pandaily ↗

Chinese AI chip designers turn to 3D stacking to bypass EUV restrictions and chase memory bandwidth

With advanced extreme ultraviolet lithography tools off limits to Chinese fabs under U.S.-led export controls, domestic AI chip startups are pivoting to 3D hybrid bonding and stacking as an alternative to transistor miniaturization, Pandaily reported. Suanmiao Tech unveiled an 8-layer DRAMDRAMThe working memory that phones, PCs, and AI servers use to hold data that a processor is actively handling.-on-logic processor called TokenPU designed for large language model inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training., while Kuaishou chip subsidiary Lingchuan said its SL200 stacked accelerator has shipped nearly 100,000 units across Alibaba Cloud, Baidu Cloud and Bilibili. Rockchip, Tsingway and others are pursuing similar architectures with backing from Tsinghua University, the Beijing Academy of Artificial Intelligence and Zhipu AI, framed by industry executives as a "curve overtaking" strategy.

Read at Pandaily ↗

III.Policy Tracker

Congressional and state tax writers eye AI compute and data centers as new tax bases

Tax writers in Congress and statehouses are looking at AI computeComputeThe processing power used to train and run AI models, usually measured in chips, GPU-hours or FLOPs., data centers and AI generated revenue as new tax bases, Roll Call reported Monday in a survey of pending proposals. Multiple bills are in scope at both the federal and state level, framed by sponsors as a fiscal response to AI's growing share of corporate value and electricity demand. Sen. Elizabeth Warren has proposed an excise tax on data center energy usage, Rep. Greg Casar has called for a tax on AI tokens and Sen. Ron Wyden has floated taxing tech companies to fund a wage-security program for workers whose jobs are cut due to AI, per Roll Call.

Read at Roll Call ↗

Warren and Scanlon to refile Health and Location Data Protection Act, expanded to cover AI chatbot disclosures

Sen. Elizabeth Warren (D-Mass.) and Rep. Mary Gay Scanlon (D-Pa.) are preparing to reintroduce the Health and Location Data Protection Act in the coming weeks, with a refreshed version that would ban data brokers from collecting or selling health and location information disclosed to AI chatbots including ChatGPT and Claude, The Verge reported. The original bill, first introduced in June 2022, prohibited brokers from trafficking such data; the 2026 version expands the ban to cover any company selling such data to brokers, and specifically names AI chatbot disclosures as covered information. The bill targets a gap that has grown as users routinely share medical and location detail with chatbots that are not bound by HIPAA, the federal health privacy law.

Read at The Verge ↗ • Read at Sen. Warren ↗

Democratic "Project 2029" rolls out "Kids Over Clicks" child safety plank as 2028-cycle opener

Project 2029, a liberal counterweight to the conservative Project 2025, released its first policy proposal Monday with a "Kids Over Clicks" framework that would ban social media accounts for users under 16, narrow Section 230 protections for AI generated content and paid advertising, ban AI chatbots "cosplaying as licensed professionals," and prohibit cellphones in schools, per Semafor. The group is led by former Biden adviser Chad Maisel and counts New Jersey Gov. Mikie Sherrill and American Federation of Teachers President Randi Weingarten among named supporters. Project 2029 has said it plans to release subsequent policy planks on other topics.

Read at Semafor ↗ • Read at Gizmodo ↗

Trump nominates acting Labor Secretary Sonderling, an AI focused former EEOC commissioner, to the permanent role

President Trump on Monday formally nominated acting Labor Secretary Keith Sonderling to the permanent post via Truth Social, two months after predecessor Lori Chavez-DeRemer resigned in April amid misconduct investigations, per AP. Sonderling previously served as deputy Labor Secretary and as a Republican commissioner on the Equal Employment Opportunity Commission, where he led federal guidance on AI enabled hiring discrimination. He has also held acting director posts at the Office of Government Ethics and the Institute of Museum and Library Services during the second Trump term. His confirmation will go to the Senate Health, Education, Labor and Pensions Committee.

Read at The Hill ↗ • Read at AP ↗

Hegseth names 15-member Defense Policy Board chaired by Lighthizer, with Andreessen among members

Defense Secretary Pete Hegseth on Monday announced 15 appointments to the Defense Policy Board, with former U.S. Trade Representative Robert Lighthizer as chair and former Sen. Norm Coleman (R-Minn.) as vice chair, the Pentagon said. Members include venture capitalist Marc Andreessen, former Arizona Senate candidate Blake Masters, retired Adm. Chas Richard who led U.S. Strategic Command, and China hawks Michael Pillsbury and Michael Anton. Hegseth had disbanded the prior panel last April after a 45-day review, with a Federal Register notice authorizing the board's return published last August. Per the notice, the board advises senior Defense Department leaders on national security and procurement strategy.

Read at The Hill ↗ • Read at Washington Examiner ↗

Trump's Axon stock purchase preceded ICE's pursuit of a $220 million Taser contract

President Trump bought between $1 million and $5 million of Axon Enterprise stock on Feb. 10, two weeks before Immigration and Customs Enforcement posted a notice seeking a five-year, $220 million deal for roughly 17,800 Tasers — with specifications that procurement reviewers and three policing experts told CNBC appeared to match only Axon products, CNBC reported, citing federal disclosures made public May 14. Axon's stock rose more than 34% in the week after the notice, and if Trump bought at the top of the disclosed range, his paper gain stood at roughly $350,000 as of June 26. The notice was a request for information rather than a binding solicitation, and no contract has been awarded. The White House said Trump's assets sit in a trust managed by his children and denied any conflict of interest, and presidents are exempt from the criminal conflict-of-interest statute that binds most executive-branch officials; CNBC found no evidence Trump knew of the procurement or that Axon knew he held shares.

Read at CNBC ↗

IV.Capability & Research Watch

More than 100 chatbots are marketed as mental health tools with thin clinical evidence

More than 100 AI chatbots are now being marketed to consumers as mental-health or therapy tools without the clinical evidence FDA reviewers typically require for medical devices, the Wall Street Journal reported. Apps such as Earkick offer cognitive-behavioral exercises and 24/7 conversational support while sitting outside FDA jurisdiction by avoiding explicit treatment claims, per AP. Earkick co-founder Karin Andrea Stephan told AP the company does not market itself as therapy. The market growth sharpens the FDA and state attorney general debate over whether AI mediated mental health support warrants pre-market review or post-market reporting.

Read at WSJ ↗ • Read at AP ↗

CAREBench releases child safety risk benchmark for frontier LLMs, finds failure rates of 2% to 58% across seven models

A new arXiv paper introduces CAREBench, a 500-prompt benchmark for evaluating upstream child safety risks in language models, with response annotations developed by parents and clinicians. The benchmark covers 12 risk categories including grooming and relationship engineering, deception and impersonation, surveillance and privacy, sextortion, AI anthropomorphization, emotional dependency and mental illness sensitivity, excluding explicit abuse material. The authors evaluated seven frontier models and found failure rates ranging from 2% to 58%, with failure patterns varying across risk categories. The benchmark scopes the evaluation to recognition, refusal, de-escalation and redirection before harm becomes overt.

Read at arXiv ↗

V.Industry & Market Watch

U.S. steelmakers warn AI data center electricity demand is squeezing reshoring margins

U.S. steelmakers have begun warning that surging AI data center electricity demand is competing for the grid power they need to operate domestic mills, the Wall Street Journal reported. Steel producers have been among the largest beneficiaries of the data center buildout through orders for structural components. Rising electricity prices and capacity constraints are eroding the margins the administration's reshoring agenda depends on, executives told the WSJ.

Read at WSJ ↗

Heatwaves and severe weather emerge as a material risk for AI data center siting and insurance

Severe weather, led by prolonged heatwaves, is straining the power grids that AI data centers depend on and driving up operators' insurance and repair costs, CNBC reported Monday. Severe weather has become the top cause of loss in Zurich's U.S. data center builders' risk portfolio over the past three years, now accounting for a third of the insurer's losses, and a First Street study found 79% of global data center capacity faces elevated risk from hazards such as flooding, extreme winds and wildfires. Hyperscalers are adapting: Microsoft cited site selection and redundant systems, and Nvidia said its newest AI servers can run cooling liquid at 45 degrees Celsius, noting that raising chiller temperatures by one degree cuts cooling energy costs about 4%.

Read at CNBC ↗

Study of 22,000 U.S. firms finds heaviest AI spenders hire faster than peers, including at entry level

White-collar headcount rose 10.2% at the companies using generative AI most intensively in the first two years after adoption, with entry-level roles up 12%, according to research covering almost 22,000 U.S. companies, the Financial Times reported. Firms in the bottom two-thirds of AI spending per worker showed no significant change against a control group, and co-author Ara Kharazian, Ramp's chief economist, cautioned that the lift appears only after a six- to 12-month learning curve and was concentrated almost entirely in the tech sector among white-collar roles. The findings cut against forecasts of broad AI-driven job losses, even as Oracle, Snap, Block and Cisco have tied layoffs to AI. But a labor economist warned the results are hard to read because heavy adopters skew toward small, fast-growing startups, and separate Stanford and Harvard studies found AI adoption reduced early-career and junior employment.

Read at FT ↗

VI.Global & Geopolitics

U.K. Defence Investment Plan allocates £5 billion to drones and AI alongside fighter jets

Prime Minister Keir Starmer will publish the U.K.'s long-delayed Defence Investment Plan on Tuesday, prioritizing £5 billion (about $6.4 billion) for drones, uncrewed ships and submarines, and autonomous systems alongside continued fighter jet and nuclear deterrent funding, Reuters and the AP reported. New Defence Secretary Dan Jarvis said uncrewed systems are defining modern warfare. Jarvis secured a £14.5 billion (about $18.6 billion) settlement, up from the £13.5 billion (about $17.3 billion) offered to predecessor John Healey, who resigned in protest June 11 saying the plan fell short of the £28 billion (about $36 billion) gap defense chiefs had identified. The plan formalizes AI and drones as top-tier U.K. military investment categories. Starmer is expected to leave office within weeks.

Read at Bloomberg ↗ • Read at AP ↗ • Read at FT ↗

EU and China set October deadline for "tangible" results on trade reset covering AI exports

EU trade commissioner Maros Sefcovic and Chinese Commerce Minister Wang Wentao agreed Monday in Brussels to an October deadline for progress on trade disagreements, including Chinese export controls that have limited European access to key inputs for AI and chip manufacturing, the EU and Chinese trade ministries said in their first joint statement since 2019. Sefcovic said the dialogue would also cover the EU's growing China trade deficit, which reached €360 billion (about $410 billion) last year, intellectual property rights and World Trade Organization reform. The two sides will set up a joint platform to monitor trade flows and identify import surges. Sefcovic plans to travel to Beijing in autumn to assess progress.

Read at Investing.com ↗ • Read at Euronews ↗

South Korea looks to cut nuclear plant construction time to meet AI power demand

The South Korean government will examine ways to shorten the nine-to-10-year timeline for building nuclear power plants to ramp up electricity supply for AI infrastructure, presidential chief of staff Kang Hoon-sik told reporters Monday. The acceleration plan will be detailed in the upcoming basic plan for long-term electricity supply and demand. South Korea already draws about a third of its electricity from nuclear and is orchestrating at least 1,350 trillion won (about $872 billion) in investment from Samsung Electronics and SK Hynix into chip and data center clusters, the industry ministry said this week. Small modular reactors remain unproven at commercial scale.

Read at Bloomberg ↗ • Read at The Edge Singapore ↗

Taiwan opposition KMT proposes NT$240 billion drone budget after blocking Lai government's special-budget version

Taiwan's main opposition Kuomintang on Monday submitted a draft six-year NT$240 billion (about $7.5 billion) drone procurement and industrial development budget to the Legislative Yuan, days after KMT lawmakers blocked President Lai Ching-te's NT$210 billion (about $6.6 billion) special budget for the same purpose, Taipei Times reported. The KMT plan would route NT$40 billion (about $1.3 billion) annually through the general budget for legislative review, require written notification for any procurement exceeding NT$100 million (about $3.1 million), and require 80% domestic production within four years. KMT caucus secretary-general Lin Pei-hsiang said the party objected to special budget mechanisms that bypass parliamentary scrutiny rather than to drone spending itself. The bill goes to the Legislative Yuan for review.

Read at Taipei Times ↗ • Read at SCMP ↗