AI Policy Weekly — Issue #5: Trump signs frontier-model EO, Congress floats 3-year state freeze, and Warren calls Jensen Huang to testify
Trump signed an executive order creating a voluntary (but de facto mandatory) NSA-led pre-release review for frontier AI models on June 2, followed by NSPM-11 on June 5 directing the national security enterprise to accelerate AI adoption. The Great American AI Act discussion draft dropped June 4, proposing 3-year preemption of state AI development laws alongside mandatory semi-annual audits for companies above $500M revenue. Sen. Warren called Nvidia CEO Jensen Huang to testify June 11 on export-control compliance. The EU Digital Omnibus provisional agreement must be formally adopted before August 2 or relief provisions lapse.
The White House issued two AI-governance documents in four days, Congress dropped its most detailed federal AI bill yet, and a Senate hearing on Nvidia's China chip compliance is now scheduled for June 11. The EU's August 2 enforcement clock is under 60 days.
1. Trump signs the frontier-model EO — and NSPM-11 follows three days later
On June 2, President Trump signed Executive Order "Promoting Advanced Artificial Intelligence Innovation and Security." The EO had been drafted, leaked, and publicly killed the prior week before Trump reversed course and signed a version with one substantive change: the government's pre-deployment access window shrank from 90 days to 30.1
What the EO does:
Section 3 directs the NSA, CISA, and Treasury — in consultation with NIST and the White House science adviser — to do three things within 60 days: develop a classified benchmarking process to determine what counts as a "covered frontier model" (defined by cyber-capability thresholds); build a voluntary framework through which developers can submit models for up to 30 days of government access before releasing to "other trusted partners"; and explicitly disclaim that none of this creates a mandatory licensing, preclearance, or permitting regime.1
That disclaimer is doing a lot of work. The framework is technically voluntary, but access to federal government contracts and the ability to release to key institutional partners will effectively require participation — a point made explicitly by government contracts attorney Jessica Tillipman in public commentary cited by Zvi Mowshowitz's analysis of the order.2
Section 2 is less contested: it gives NSS, DHS, and Treasury 30 days to accelerate cyber defenses across national security systems, civilian federal networks, and critical infrastructure. Section 4 directs the attorney general to prioritize prosecution of AI-enabled cybercrime under existing statutes.1
正在加载内容卡片…
Three days later, on June 5, the White House issued NSPM-11, a National Security Presidential Memorandum directing the national security enterprise to accelerate AI adoption under four pillars: Adoption (eliminate barriers to deploying frontier models), Adaptation (prefer commercial and open-source over internal builds), Assurance (reliability and controllability requirements), and Accountability (no unauthorized surveillance, civil liberties compliance). NSPM-11 rescinds NSM-25 and directs DoD to update its Autonomy in Weapon Systems directive within 90 days.3
What to watch: The EO is structurally identical to the version Trump called "too regulatory" on May 21. The benchmarking process and trigger threshold will be classified — meaning labs won't necessarily know whether a model they're training is subject to pre-deployment review until they seek a determination. How BIS, NSA, and CAISI (the civilian NIST-based evaluation body that many experts argue should lead this work) divide authority in practice will define whether the EO becomes a minor coordination mechanism or a de facto checkpoint regime. CAISI has not been formally authorized by Congress; the Great American AI Act would change that.
2. The Great American AI Act: a federal floor with a 3-year ceiling on states
On June 4, Reps. Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page discussion draft of the Great American Artificial Intelligence Act of 2026, with four additional co-sponsors.4 It is the most detailed bipartisan federal AI proposal yet and, depending on which provision you focus on, either the first serious attempt at federal AI governance or the first serious attempt to gut state AI authority.
正在加载内容卡片…
The preemption provision: For three years, states could not enact new laws specifically governing the development of AI models — how systems are trained, built, and weighted. State laws governing deployment and use would remain intact, as would civil rights, labor, copyright, child safety material rules, and privacy law. Trahan's own FAQ names specific casualties: California AB 2013 (training-data disclosures) and portions of California SB 942 (AI content watermarking).4 Frontier-safety laws in California, New York, and Illinois would be "federalized" — a characterization critics noted is a euphemism for superseded.
This is the third congressional attempt at federal preemption. The Senate voted 99-1 to strip a 10-year moratorium from the One Big Beautiful Bill, and a version attached to the FY2026 NDAA also failed. The new bill uses a narrower 3-year window and limits preemption to the development side, which sponsors say is the key structural difference.4
The audit requirement: Companies with more than $500M in annual revenue that develop frontier AI models must submit to semi-annual audits by CAISI-licensed Independent Verification Organizations (IVOs). IVOs assess developer frameworks across four risk categories: cybersecurity, biosecurity, CBRN uplift, and loss-of-control scenarios. Penalties for non-compliance or misrepresentation run up to $1 million per day for the duration of the violation. Whistleblower protections for employees at covered labs are included.4 The IAPP noted this builds on Illinois SB 315 but raises the stakes considerably — state laws cap total liability at $1-3M per incident; this bill runs penalties indefinitely.4
The reception: Opposition was immediate and came from multiple directions at once. The AFL-CIO (15 million workers) called it a "hard no." Public Citizen called it "a disastrous proposal that Big Tech is celebrating." The House Commission on AI and the Innovation Economy — a Democratic body co-chaired by Reps. Ted Lieu, Valerie Foushee, and Josh Gottheimer — declared "we do not support the discussion draft as it currently stands" on the same day it was released.4 On the other side, the Information Technology Industry Council backed the bill; Anthropic and OpenAI, both covered by the audit regime, supported Illinois SB 315 (the audit model the bill expands), and have not publicly opposed the federal version.
The bill is a discussion draft: no floor date, comment period open at [email protected]. CAISI would be formally authorized and funded at $100M/year FY2027-2029, up from its current ~$15M.4
What to watch: The preemption provision will drive the legislative fight. The bill's sponsors need to hold their bipartisan coalition together on preemption while the House Democratic commission has already declared the draft inadequate. Watch for revised drafts that narrow the preemption scope or provide carve-outs for specific categories (worker surveillance, automated employment decisions) where documented harms are well-established and federalized protection does not yet exist. The 99-1 Senate vote in July 2025 is a baseline: even if this bill narrows the preemption window, senators who killed the 10-year moratorium will need a reason to vote differently on a 3-year version.
3. Warren calls Nvidia CEO to testify — the chip accountability gap widens
On June 1, Senator Elizabeth Warren (D-MA), ranking member of the Senate Banking Committee, sent a letter to Nvidia's General Counsel Tim Teter and audit committee chair Brooke Seawell probing the company's compliance with U.S. export control laws.5 By June 4, Warren had invited CEO Jensen Huang to appear before the committee on June 11 to answer questions about Nvidia's China business and export controls.6
The trigger: multiple recent DOJ prosecutions allege organized networks routing Nvidia GPUs — H100s, H200s, A100s — to China through Malaysia, Thailand, and other third countries, in direct contradiction of Huang's public statements that Nvidia chips have zero market share in China and that he has seen no evidence of diversion.5 The Warren letter specifically cited DOJ prosecutions involving over $510M in diverted servers and $160M in H100/H200 chips.
This is also bipartisan. By June 5, Senator Jim Banks (R-IN) had joined Warren in warning Huang ahead of a planned trip to China, cautioning that meeting with firms that have circumvented export controls would legitimate those networks.7
The committee gave Nvidia a June 18 deadline for a written response to Warren's original letter.5 The questions on the table: whether the board exercised adequate oversight, whether Nvidia modified customer due-diligence and export-control compliance processes after the prosecutions, and whether public statements about chip flows to China were accurate.
This fits into a broader accountability gap. BIS's May 31 guidance (reported in Issue #4) reactivated license requirements for Chinese-headquartered entities buying through overseas affiliates, but explicitly did not address past shipments or TSMC's supply-chain due-diligence obligations. The Warren-Banks pressure puts board-level accountability — and the accuracy of public disclosures — directly in play at the same time BIS is still figuring out how backward-looking its enforcement stance will be.
正在加载内容卡片…
What to watch: Whether Huang testifies June 11 or declines; the content of Nvidia's June 18 written response to Warren; and whether BIS issues clarifying guidance on enforcement posture for shipments that occurred during the May 2025–May 2026 period when the China subsidiary loophole was open.
4. EU Digital Omnibus must clear formal adoption before August 2
The EU's Digital Omnibus provisional agreement — reached May 7 between the Council, Parliament, and Commission — is now racing toward formal adoption with a hard deadline of August 2, 2026. If the formal vote does not occur before that date, the existing AI Act provisions apply and the compliance timeline relief disappears.8
The agreement extends the Annex III high-risk AI deadline from August 2026 to December 2, 2027 (17 months), and the product-safety-component high-risk deadline to August 2, 2028. It also tightens the watermarking/labeling deadline to December 2, 2026 — three months shorter than the original six-month window. New prohibitions against AI-generated CSAM and non-consensual sexual imagery are added under Article 5.8 The GPAI/transparency tier and the unacceptable-risk prohibitions remain fixed at August 2, 2026 — these are not affected by the agreement.9
Two additional developments this week:
- On June 5, the Commission appointed Jim Hagemann Snabe as Industrial AI Envoy, effective June 3.10 The appointment comes 18 days before the June 23 consultation deadline on high-risk AI classification guidelines.
- On June 6, the Commission proposed the Cloud and AI Development Act (CADA), establishing a four-tier sovereignty framework for which vendors can bid on EU public sector cloud and AI contracts. The highest tier requires European ownership and genuine operational autonomy — a standard no U.S. hyperscale provider currently meets under the draft criteria.10
What to watch: Formal adoption vote timing — the Council and Parliament both need to act. June 23 is the consultation deadline on high-risk classification guidelines; any response to that consultation will shape the Dec 2027 compliance scope. The CADA proposal has not been finalized; lobbying from U.S. cloud providers to modify the top-tier ownership requirements is expected.
5. State AI wave: Illinois sends five bills to Pritzker at 4:30 a.m.
The Illinois legislature adjourned at 4:30 a.m. on June 1, having sent five AI bills to Governor Pritzker's desk. The most significant is SB 315 (the Artificial Intelligence Safety Measures Act), which would be the first U.S. mandate for independent third-party safety audits of frontier AI systems.11 The other four: SB 318 (bots purchasing tickets), SB 343 (AI-assisted rental property price fixing), SB 2909 (AI in teacher evaluations), and SB 3114 (AI in healthcare approvals).
Pritzker previously confirmed he intends to sign SB 315. The governor has not yet acted; his signature remains the final step before Illinois becomes the first state in the country with mandatory frontier-AI audit requirements. Effective date: January 1, 2028. Threshold: companies with $500M+ annual revenue.
In New York, the Senate approved S 6954 (AI disclosure) and S 1169 (high-risk audit and algorithm discrimination) this week. California's 30 remaining AI bills are now in second-chamber committees with a July 2 summer adjournment deadline approaching.11
The running watch list
| Item | Status as of June 8 |
|---|---|
| Trump AI EO (June 2) | Signed; 60-day clock for NSA/CISA/Treasury to design framework and benchmarking process |
| NSPM-11 (June 5) | Issued; 90-day clock for DoD, DNI, and OMB to issue updated policies and classified annex |
| Great American AI Act | Discussion draft; no floor date; comment period open |
| Warren/Huang testimony | Huang invited for June 11 Senate Banking hearing; June 18 written-response deadline |
| BIS May 31 guidance | In effect; backward-looking enforcement posture unresolved; TSMC compliance gap unaddressed |
| BIS AI Diffusion Rule / CRA | GAO ruling stands; BIS has not clarified CRA submission |
| Illinois SB 315 | Awaiting Pritzker signature; effective January 1, 2028 |
| EU Digital Omnibus formal adoption | Must occur before August 2; GPAI enforcement at August 2 is unchanged |
| EU CADA proposal | Proposed June 6; not finalized; lobbying expected |
| Commerce "onerous AI laws" list | Expected ~late June 2026 |
| DOJ AI Litigation Task Force | Operational; active challenge to Colorado's algorithmic-discrimination statute |
参考来源
- 1White House EO – Promoting Advanced AI Innovation and Security
- 2Trump Signs Executive Order For AI Testing Prior To Frontier Model Releases – Zvi Mowshowitz
- 3NSPM-11: Artificial Intelligence in the National Security Enterprise
- 4Federal AI Regulation Bill Freezes State Consumer Protections for Three Years – TechTimes
- 5Warren Probes NVIDIA's Compliance With Export Control Laws – Senate Banking Committee
- 6Nvidia CEO Jensen Huang invited to Senate hearing on China chip compliance – CNBC
- 7US senators warn Nvidia's CEO over upcoming China trip – Straits Times
- 8EU bodies reach provisional agreement on AI Act – DLA Piper
- 9EU AI Act Update: Timeline Relief, Targeted Simplification, and New Prohibitions – Global Policy Watch
- 10EU AI Act June 2026 Digest – Tech Jacks Solutions
- 11AI Legislative Update: June 5, 2026 – Transparency Coalition
围绕这条内容继续补充观点或上下文。