US Congress Gridlock Leaves AI Governance to 50-State Patchwork
Ai4
Election-year legislative paralysis guarantees federal preemption will stall in Washington, cementing a fragmented 50-state compliance burden that threatens domestic tech startups.
## Background
Published on YouTube by Ai4 on August 22, 2026, this panel discussion features former U.S. Congressman Dan Donovan, Portulans Institute President Soumitra Dutta, and AI investor and advisor Shreya Amin. According to the description, the session examines the constitutional and economic fault line of U.S. AI governance: the battle between federal preemption and sweeping state-level legislation in jurisdictions like Connecticut, Colorado, and Delaware.
## TL;DR
The panel explores the tension between a fragmented 50-state regulatory patchwork and centralized federal preemption over artificial intelligence. The speakers advocate for bifurcating the AI governance stack, assigning foundation models and national security to federal oversight while leaving downstream deployment, civil rights, and consumer protection to the states. While warning that state-level fragmentation could stifle startup innovation and mirror Europe's economic friction, the experts highlight state sandboxes as practical proving grounds. Former Congressman Dan Donovan concludes that election-year incentives and congressional gridlock make any imminent federal preemption bill virtually impossible, leaving governance momentum firmly in state hands.
## Atmosphere & style
The panel is candid, intellectually sharp, and grounded in unvarnished political realism. The dynamic between the speakers is collegial yet urgent, balancing complex regulatory frameworks with wry humor—exemplified by Donovan noting that renting a car from Hertz requires more qualifications than being a member of Congress. Watching the exchange feels like sitting in on a high-level, insider policy debrief that transitions from technical taxonomy to a sobering reality check on Washington's legislative paralysis.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) Former Congressman Dan Donovan questions federal lawmakers' readiness to regulate AI, noting there are fewer constitutional requirements to serve in Congress than qualifications needed to rent a car from Hertz.
- [02:34](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=154s) Soumitra Dutta advocates splitting AI governance by stack layer: centralized federal oversight for foundation models and national security, paired with state-level oversight for localized downstream applications.
- [05:23](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=323s) Divergent state policies are actively driving regulatory arbitrage, pushing capital and infrastructure investment away from restrictive states like New York toward pro-business jurisdictions.
- [05:23](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=323s) Total federal preemption of the downstream 'use' layer risks stripping state regulators of their authority to enforce civil rights, monitor hiring discrimination, and require algorithmic traceability.
- [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) A 50-state regulatory patchwork threatens to throttle early-stage startups and lone entrepreneurs who lack the legal resources to shoulder disparate compliance costs.
- [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) Historical precedent, such as the federal coercion of the 55 mph national speed limit via highway funding, suggests Washington will eventually assert sweeping interstate commerce authority over data crossing state lines.
- [12:07](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=727s) Congressional inertia and upcoming election incentives stall federal tech action, even as global competitors like China aggressively deploy multi-gigawatt data center capacity weekly.
- [12:07](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=727s) Hundreds of state-level bills introduce conflicting statutory definitions for identical features—such as chatbots—creating an inefficient compliance tax across jurisdictions.
- [15:06](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=906s) The United States risks repeating Europe's structural economic errors, where cross-border regulatory fragmentation historically hindered domestic startup expansion and favored well-resourced incumbents.
- [15:06](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=906s) State-level regulatory sandboxes in places like Utah and Connecticut provide a workable alternative to bans by granting temporary statutory waivers in exchange for algorithmic transparency and audit data.
- [18:39](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1119s) Widespread legislative hesitation and election-year gridlock make federal preemption or comprehensive AI legislation virtually impossible until the next Congress convenes.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) Introductions: The Hertz Rental Test for Lawmakers
The debate opens on the emerging constitutional clash between state legislatures passing bespoke artificial intelligence statutes and Washington’s drive to assert federal preemption. Moderator Sanjay frames the dilemma facing businesses and citizens alike: will your local state capital dictate how AI touches your life, or will federal standards from Washington, D.C. supersede them all? To unpack the issue, the panel brings together perspectives spanning federal politics, public sector governance, and global academic research.
Former U.S. Congressman Dan Donovan, now serving on the advisory board of the nonprofit RegulatingAI, kicks off introductions with an unvarnished assessment of congressional competence. Donovan points out that the very individuals drafting sweeping federal tech policy have remarkably few qualifications to hold office: "The only two requirements to become a member of Congress is you have to be over 25 and a citizen of the United States. There's more requirements to rent a car from Hertz than to become a member of Congress, and they are the people that are going to be regulating this."
Shreya Amin outlines her background bridging two decades across enterprise technology and public policy. Having served as the inaugural Chief AI Officer for New York State after previously working as Chief AI Scientist at a health tech enterprise deploying software across hundreds of hospital networks and school systems, she stresses that mitigating harm and fostering innovation are complementary pursuits. If systems are poorly architected, they cause real-world damage, but establishing rigorous safety and accountability frameworks creates the exact pathways needed to accelerate beneficial adoption. Soumitra Dutta, former dean at Cornell University and Oxford and the architect of the United Nations-adopted Global Innovation Index, rounds out the introductions by noting that the central governance challenge lies in how both governments and commercial enterprises manage technology adoption without suffocating economic dynamism.
[02:34](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=154s) The Governance Dilemma: Splitting the AI Stack Between Federal and State
When pressed by Sanjay for a direct, binary choice on whether Washington or the states should monitor and regulate AI—without offering evasive "lawyer answers"—Soumitra answers unequivocally that the solution requires both, differentiated by layers of the technology stack. Looking at traditional American federalism, Dutta observes that the federal government historically claims exclusive authority over domains requiring uniform central oversight, such as nuclear energy and civil aviation. In other sectors, such as environmental protection, Washington institutes baseline standards while permitting states the latitude to pass stricter statutes. Meanwhile, fields like the legal and medical professions are left almost entirely to state-level discretion.
AI, however, dynamically shifts jurisdictional boundaries depending on which layer of the stack is under scrutiny. Developing foundation models involves global inputs, vast compute resources, and heightened national security concerns over foreign open-source threats, such as models emanating from China. These foundation-level risks necessitate uniform, centralized federal mandates. Conversely, downstream application—how companies, municipalities, and local agencies deploy systems in specific workflows—is intensely contextual and localized, requiring state-level and regional rules tailored to local community needs and data privacy expectations.
```mermaid
flowchart TD
accTitle: Federal versus State AI Governance Stack
subgraph National["Federal Jurisdiction: Foundation Layer"]
A["Frontier Model Development"] --> B["National Security & Geopolitical Risks"]
B --> C["Export Controls & Open-Source Ingress"]
end
subgraph Shared["Cooperative Standards"]
C -.-> D["Common Definitions & Baselines"]
end
subgraph Local["State Jurisdiction: Application & Use Layer"]
D --> E["Local Context & Data Usage"]
E --> F["Workforce & Algorithmic Bias"]
E --> G["Consumer Protection & Sectoral Rules"]
end
```
[05:23](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=323s) State Innovation vs. Bans: Preemption and the "Build vs. Use" Divide
Addressing the fate of state laws already enacted in California, Colorado, and Connecticut if federal preemption occurs, Donovan argues that states can retain authority as long as their statutes do not directly conflict with the federal floor. Donovan notes that state divergence inevitably sparks regulatory arbitrage: capital and businesses will migrate to jurisdictions with welcoming regulatory environments and shun hostile ones. He points directly to his home state of New York, where an ongoing moratorium on data center construction is driving AI infrastructure investments straight into more business-friendly states that encourage growth rather than stifle it.
Shreya expands on how federal preemption would cripple state-level governance if Washington attempts to preempt the downstream "use" layer rather than confining itself to the "build" layer. Regulating the handful of frontier lab builders at the federal level leaves states with a clear implementation mandate. However, if federal preemption wipes out state oversight over how AI is used, fundamental consumer protections evaporate:
- **Civil Rights and Discrimination**: If an automated screening tool rejects a job applicant, state regulators already enforce longstanding employment discrimination and labor laws.
- **Traceability and Observability**: Preempting use-layer regulation would strip state regulators of the legal authority to inspect decision traces, examine which models touched applicant data, or verify human-in-the-loop compliance.
- **Scope of Preemption**: While early legislative drafts in Washington floated total preemption of downstream usage, more recent iterations have pulled back, focusing federal power primarily on foundation models and national security.
[08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) The 50-State Patchwork and the Threat to Basement Startups
Sanjay introduces the principal objection raised by frontier labs and federal policymakers: a fragmented patchwork of 50 disparate state frameworks threatens to throttle innovation, imposing compliance costs that a lone entrepreneur working out of a basement in Tennessee simply cannot shoulder. Soumitra confirms this danger with a "definite yes," pointing out that regulation must balance two crucial priorities: preserving the velocity of innovation and value creation on one hand, while safeguarding employment and citizen livelihoods on the other.
Donovan illustrates how the federal government routinely forces state compliance even in spheres constitutionally delegated to the states. Recalling the era when individual states determined their own roadway speed limits, he recounts how the federal government coerced nationwide adoption of a 55 mph limit by threatening to withhold federal highway funds. Donovan predicts federal authorities will adopt an equally aggressive interstate commerce posture over AI, asserting jurisdiction the moment code or data "goes over a wire." Nonetheless, market dynamics will ultimately dictate the landscape as emerging hubs like Nevada cultivate competitive, pro-business ecosystems, creating distinct "AI states."
[12:07](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=727s) Congressional Inertia, Election Incentives, and China's Infrastructure Push
When asked whether state governors are justified in passing their own legislation because Washington refuses to act, the panel underscores the deep structural paralysis within the federal legislative process. Donovan emphasizes that while the executive branch can issue directives with urgency, Capitol Hill operates at a glacial pace. With the entire House of Representatives and a third of the Senate facing reelection, lawmakers stall on controversial topics until the next Congress convenes in January: "It's easier in government to say no and be wrong than to say yes and be wrong. So, they caution on the side of no."
Shreya contrasts Washington's legislative stalling with the urgency of international competition, noting that while American stakeholders become entrenched in ideological pro- versus anti-regulation squabbles, China is aggressively scaling infrastructure, adding multiple gigawatts of data center capacity every week. She observes that among the hundreds of AI bills circulating across statehouses between 2024 and 2026, the overwhelming majority do not target frontier model providers—they regulate specific deployment formats, such as roughly 50 separate bills regulating chatbots. Subjecting developers to 50 distinct legal definitions for identical technical features creates an unsustainable compliance tax in the absence of national alignment.
[15:06](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=906s) The European Warning, Regulatory Capture, and State Sandboxes
Turning to the audience of founders and startups at the conference, Sanjay asks whether regulatory complexity serves as a vehicle for regulatory capture, wherein well-capitalized tech incumbents with armies of attorneys crowd out smaller players. Soumitra cautions that the United States risks repeating Europe's structural mistakes. Having lived and worked across Europe for years, he explains that despite the European Union single market, regional fragmentation makes it prohibitively expensive for an early-stage French startup to expand across continental borders. Failing to establish regulatory harmony across U.S. states could trigger the same dampening effect on domestic entrepreneurship.
In response to whether technical standards could substitute for absent federal legislation, Shreya notes that enterprise developers in regulated fields like healthcare and pharmaceuticals already build software to the single strictest state standard available. However, she argues standards must be paired with dynamic regulatory frameworks rather than outright bans:
- **The Sandbox Model**: States like Utah are pioneering regulatory sandboxes that bring regulators directly to the table with startups, temporarily waiving specific statutory hurdles in exchange for deep algorithmic observability—enabling innovations like autonomous AI-driven prescription platforms.
- **Policy Experimentation**: Connecticut is adopting similar flexible testing grounds to encourage controlled deployment.
- **Evidence-Based Policymaking**: Sandboxes generate empirical operational data, preventing sudden, politicized moratoriums like New York's ban on data centers.
[18:39](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1119s) The Odds of a Federal Preemption Bill Passing Congress
The discussion concludes with Sanjay asking the panel to give an unvarnished forecast: if the White House introduces a formal preemption bill in either chamber of Congress this year, can it actually pass? Donovan shuts down the prospect entirely, noting that legislative business in the House is functionally dead until the new Congress convenes in January, with reelection pressures preventing any substantive breakthrough.
[19:09](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1149s) Congressional Gridlock and Closing Remarks
The panel concludes on a sober note regarding the prospects of imminent federal action on AI legislation. The speakers candidly observe that lawmakers this year appear "afraid to vote on anything," making it exceedingly unlikely that comprehensive federal measures will pass in the near term. This legislative stalling extends across a wide spectrum of interconnected technological domains, bottlenecking bills related to data centers, core artificial intelligence, physical AI, and robotics. One panelist lightheartedly defers to their "two esteemed colleagues," acknowledging their deeper familiarity with the political realities holding these measures back.
A lingering sense of frustration surfaces over the lack of federal progress, with a speaker emphasizing the urgency of delivering results given public investment: "We got to get something done, man. I mean, our—these are our tax dollars at work." The session wraps up by noting that the political gridlock paints a challenging picture for Washington's ability to lead, leaving the momentum with individual states to navigate AI governance. The moderator thanks the panelists for illuminating the high-stakes friction between federal inaction and state-level regulation as the session closes to audience applause and cheering.
- **Stalled Federal Legislation:** Critical bills addressing data centers, artificial intelligence, physical AI, and robotics remain mired in political hesitation and inaction.
- **Federal vs. State Dynamic:** With Washington sidelined by gridlock, state-level initiatives and enforcement continue to drive the immediate future of AI oversight and policy.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) The only two requirements to become a member of Congress is you have to be over 25 and a citizen of the United States. There's more requirements to rent a car from Hertz than to become a member of Congress, and they are the people that are going to be regulating this. — Dan Donovan
> [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) goes over a wire — Dan Donovan
> [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) AI states — Dan Donovan
> [12:07](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=727s) It's easier in government to say no and be wrong than to say yes and be wrong. So, they caution on the side of no. — Dan Donovan
> [19:09](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1149s) afraid to vote on anything
> [19:09](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1149s) two esteemed colleagues
> [19:09](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1149s) We got to get something done, man. I mean, our—these are our tax dollars at work.
## Mentioned
Dan Donovan, Soumitra Dutta, Shreya Amin, Sanjay, RegulatingAI, Portulans Institute, Minsphere, U.S. Congress, OpenAI, Hertz, New York State, Cornell University, University of Oxford, United Nations, Global Innovation Index, White House, European Union, Colorado AI Act, California, Colorado, Connecticut, Delaware, Nevada, New York, Tennessee, Utah, China
## Video description
Dan Donovan - Congressman, U.S. Congress
Soumitra Dutta - President, Portulans Institute
Shreya Amin - AI Advisor / Investor, Minsphere
A multi-level debate bringing together sitting Attorney Generals, Governors, and Members of Congress to examine the most explosive fault line in U.S. AI governance: federal preemption of state AI laws. Connecticut just passed sweeping AI legislation (May 2026), Delaware is pioneering a first-in-nation OpenAI certification partnership, and Colorado's AI Act takes effect June 30, 2026 — while Washington pushes for a single national standard. Who wins, and what does it mean for every company operating across state lines?
As congressional gridlock stalls federal artificial intelligence legislation, policy experts argue that effective oversight requires bifurcating governance across the technology stack—assigning centralized foundation-model safeguards to Washington while leaving localized, downstream application rules to the states.
Bifurcating Oversight Between Build and Use Layers (2:34)
Soumitra Dutta asserts that AI governance cannot be treated as a binary choice between state and federal control, drawing parallels to established frameworks in nuclear energy, civil aviation, and environmental protection. Dutta argues that the foundation layer—involving frontier model training, vast compute resources, and national security threats from foreign open-source ingress—demands uniform, centralized federal mandates. Conversely, downstream application is inherently contextual and localized, requiring state oversight to address regional data privacy and community norms.
Shreya Amin expands on this divide by distinguishing the 'build' layer from the downstream 'use' layer. She warns that total federal preemption would strip state regulators of authority to enforce existing civil rights, labor, and anti-discrimination statutes when automated tools screen job applicants. Preempting the use layer would also eliminate states' legal authority to inspect decision traces, evaluate model data access, or mandate human-in-the-loop accountability.
Patchwork Compliance Risks and Regulatory Arbitrage (8:51)
Panelists caution that an uncoordinated 50-state regulatory patchwork threatens early-stage startups lacking enterprise legal resources. Dutta warns that the United States risks repeating Europe's structural mistake, where cross-border regional fragmentation made expansion cost-prohibitive for early-stage companies. Amin notes that among the hundreds of state bills introduced between 2024 and 2026, the vast majority target deployment mechanisms rather than frontier labs—including roughly 50 separate bills regulating chatbots—creating a burdensome compliance tax of conflicting technical definitions.
Dan Donovan highlights that divergent state laws inevitably spark regulatory arbitrage as capital migrates toward accommodating jurisdictions. Donovan points to New York, where an ongoing moratorium on data center construction is pushing computing investments into emerging pro-business hubs like Nevada. While Donovan predicts Washington will eventually assert broad interstate commerce jurisdiction the moment data crosses state lines, market competition currently dictates geographic deployment.
Regulatory Sandboxes Versus Blanket Statutory Bans (15:06)
Rather than relying on blunt statutory prohibitions or fragmented state mandates, Amin advocates for dynamic regulatory sandboxes to support safe deployment. In states like Utah and Connecticut, sandbox frameworks allow startups to temporarily waive specific statutory hurdles in exchange for granting regulators granular algorithmic observability. This model has already enabled empirical testing for sensitive applications, such as autonomous AI-driven prescription platforms.
Amin argues that technical standards alone cannot replace legislation unless paired with policy experimentation. Generating empirical operational data through sandboxes provides lawmakers with concrete evidence, preventing sudden, politicized measures like data center moratoria while maintaining rigorous oversight over safety and accountability.
Capitol Hill Gridlock Against Global Competition (12:07)
Despite mounting economic stakes, Donovan forecasts near-zero probability that Congress will pass a comprehensive federal AI preemption bill in the near term. Facing reelection pressures, lawmakers systematically avoid controversial votes, adhering to the political reality that it is easier in government to say no and stall than to vote yes and risk being wrong. This legislative paralysis bottlenecks policies across foundational AI, physical robotics, and data centers.
Amin contrasts Washington's legislative inertia with the speed of global competitors. While domestic stakeholders remain mired in ideological disputes over regulation, China is aggressively scaling physical infrastructure, deploying multiple gigawatts of data center capacity every week. Panelists stress that persistent federal inaction risks ceding technological momentum while leaving domestic governance unresolved.
With Washington paralyzed by election-cycle gridlock and institutional inertia, individual states remain the primary drivers of artificial intelligence policy and enforcement in the United States. Whether state-level experimentation will produce workable national standards or entrench an unmanageable regulatory patchwork remains the critical unresolved friction facing domestic developers.
## Background
该内容由行业峰会频道 Ai4 于 2026 年 8 月 22 日发布在 YouTube 平台上。视频简介指出，这是一场汇集了前国会议员、州级 AI 决策官员以及学术领袖的多方圆桌辩论，核心探讨美国各州竞相出台 AI 立法与华盛顿联邦政府推行“联邦优先权”之间的激烈权力博弈及其对跨州企业的深远影响。
## TL;DR
本场讨论围绕美国联邦与各州在 AI 治理上的管辖权冲突展开。专家指出应明确划分监管层级——联邦负责前沿底层大模型与国家安全标准，而各州则应保留在终端应用场景、反歧视与劳工保护方面的执法权。与会嘉宾警告，全美 50 州各自出台碎片化定义不仅会带来沉重的重复合规负担，甚至可能造成巨头垄断与“欧洲式”的创新停滞；但在大选年国会政治僵局下，短期内通过全美统一的联邦优先权法案几乎毫无可能，企业仍需依托最严标准或犹他州等地的监管沙盒谨慎前行。
## Atmosphere & style
对谈节奏紧凑、交锋直接，主持人摒弃了外交辞令式的含糊表态，频频以“直击痛点”的追问逼出真实底牌。前国会议员以熟稔的体制内黑色幽默撕开华盛顿“多做多错、不做不错”的政治怠惰，与前州级 AI 官员的技术务实、学术界导师的全球宏观视野相互碰撞交织。整场氛围既有对 50 州合规拼贴扼杀小创企的焦虑与现实批判，又在探讨监管沙盒与技术突破时透露出务实求解的张力，深刻展现出政策滞后于算力飞跃的时代阵痛。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) 联邦与州权博弈的核心在于“联邦优先权”（Federal Preemption）：随着各州相继立法，华盛顿试图统一标准以防碎片化，但这直接冲击了各州自主治理 AI 影响的宪政权力。
- [03:24](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=204s) AI 技术栈需实行分层协同治理：底层大模型与国家安全标准（构建层）应由联邦集中把控，而终端落地、具体场景与企业应用（使用层）则需各州根据本地语境监管。
- [05:25](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=325s) 各州政策环境差异驱动企业用脚投票：例如纽约州出台数据中心建设暂停令，直接促使算力资本与 AI 业务流向政策更友好、支持创新的州。
- [06:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=411s) 联邦优先权若强行覆盖“使用层”，将剥夺州政府利用反歧视法与消费者保护法审查 AI 决策链路追溯及人机协同流程的核心能力。
- [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) 联邦政府常通过财政杠杆干预州权：正如历史上以扣留公路补助金强制推行全美 55 英里限速一样，华盛顿未来可能借助资金与跨州商业条款介入 AI 治理。
- [11:24](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=684s) 大选政治周期与体制保守性加剧国会立法停摆：面临众议院全员及参议院三分之一改选，议员普遍规避风险，导致联邦立法在新国会就职前全面搁浅。
- [13:38](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=818s) 州级立法涌现导致法律定义严重碎片化：例如全美 50 种 AI 聊天机器人法案给出互不兼容的定义，给面临全球算力竞赛的科技企业带来极重的重复合规负担。
- [15:07](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=907s) 复杂的州级法律补丁存在“监管俘获”隐忧并面临欧洲化风险：大科技公司可凭借雄厚法务资源筑起合规护城河，而初创企业则面临被高昂跨州合规成本淘汰的困境。
- [17:03](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1023s) 犹他州与康涅狄格州的监管沙盒提供了循证决策新范式：企业以系统高透明度换取适度豁免，在医疗处方等前沿场景试点，有效对冲非理性的政治禁令。
- [18:39](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1119s) 选战之年联邦优先权法案落地基本无望：受制于连任表决压力与华盛顿立法停滞，年内几乎不可能出台废除各州自制 AI 法案的全国统一立法。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) 联邦与州权之争：谁来监管人工智能？
小组讨论由主持人 Sanjay 揭开序幕。他指出，随着全美多个州相继出台各具特色的 AI 法律，联邦政府正试图通过“联邦优先权”（Federal Preemption）来废止或统一这些州级立法。这不仅是一场关于美国联邦制（Federalism）的宪政理论探讨，更是一场正在发生的权力博弈：究竟是由各州政府还是由华盛顿特区来决定 AI 对普通人的影响？
随后，三位嘉宾依次进行自我介绍，并被要求回答一个核心问题——“究竟谁应该来监管 AI？”
```mermaid
flowchart TD
accTitle: AI Governance Level Debate
subgraph Federal["联邦政府层面 (Washington D.C.)"]
F1["前沿大模型监管"]
F2["国家安全与地缘竞争"]
F3["全国统一基准与防碎片化"]
end
subgraph State["各州与地方层面 (State & Local)"]
S1["终端使用与场景落地"]
S2["反歧视与劳工权益保护"]
S3["消费者保护与数据溯源"]
end
F1 -.->|"潜在联邦优先权覆盖"| S1
```
- **丹·多诺万（Dan Donovan）**：前美国国会议员，现任非营利组织 RegulatingAI 顾问委员会成员。他幽默地自嘲道，国会议员正是起草法律的人，但“当选国会议员只有两个基本要求：年满 25 岁且为美国公民——在赫兹（Hertz）租辆车的要求都比当国会议员要多”，而正是这群人在制定 AI 的监管规则。该组织的目标是在激发创新的同时遏制不良行为者的恶意活动。
- **施雷亚·阿明（Shreya Amin）**：曾任纽约州首位首席 AI 官（State AI Officer），在行业深耕近 20 年，曾担任医疗科技公司首席 AI 科学家（产品部署于数百家医院），并曾在学区部署机器学习系统。她强调不良系统会带来伤害，但安全、负责任的系统则能加速社会效益，监管与创新是“同一枚硬币的两面”。
- **苏米特拉·杜塔（Soumitra Dutta）**：牛津大学和康奈尔大学前院长及教授，UC 伯克利 AI 博士，也是联合国广泛采纳的“全球创新指数”（Global Innovation Index）的创立者。其研究重心长期聚焦于政府和企业如何落地创新技术。
---
### [03:24](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=204s) 监管层级划分：中央标准与地方治理的协同
面对主持人“不要律师式含糊其辞、必须明确给出由谁监管”的逼问，苏米特拉·杜塔明确指出答案必然是“两者兼需”（Both），关键在于厘清具体分工。他回顾了美国经典的监管架构：有些领域适合高度中央化管理（如核能、民航），有些领域采取混合模式（如环境保护，联邦设定底线标准，各州可自行制定更严格的法规），而某些行业则完全下放给各州（如法律、医疗执业许可）。
杜塔教授分析，AI 技术栈（AI Stack）本身跨越了多个层级，这使得界限划分尤为复杂。在“构建 AI”（Building AI）层面，底层模型汇集了全国乃至全球的数据与算法输入，同时涉及开源模型安全（如来自海外的竞争与安全风险），这一层级必须由联邦政府出台集中统一的红线与国家标准；而在“使用 AI”（Using AI）层面，AI 的部署高度依赖于城市、企业和区域的具体语境（Contextual），因此必须由各州和地方制定切合本地数据与实际场景的规则。
---
### [05:25](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=325s) 联邦优先权对各州立法的冲击与政策竞争
当主持人追问若联邦政府推行彻底的优先权政策，加州、康涅狄格州和科罗拉多州现已通过的 AI 法案将何去何从时，丹·多诺万指出，只要州法不违背联邦设定的基础底线，州级规则依然会发挥作用。未来不可避免地会出现联邦与州监管并存的局面，各州也将展开政策竞争。
多诺万以纽约州当前的政策为例说明了商业环境的敏感性：纽约州目前对数据中心建设实施了暂停令（Moratorium），这直接导致许多 AI 算力与业务流向了监管更友好、政策更支持创新的州。企业会用脚投票，选择前往那些能够促进增长而非抑制创新的地区。
---
### [06:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=411s) “构建”与“应用”的界限：使用层级监管被剥夺的风险
针对联邦优先权是否会瘫痪州级治理体系的问题，施雷亚·阿明深入剖析了“构建层”（Build）与“使用层”（Use）的区别。她指出，如果联邦仅仅是针对少数开发前沿模型（Frontier Models）的底层实验室进行集中监管，各州只需负责配合执行，冲击相对可控。但若联邦优先权强行延伸至“使用层”，将直接剥夺州政府保护民众的核心能力。
```mermaid
flowchart LR
accTitle: AI Application Discrimination & Regulatory Audit Flow
subgraph UserScenario["AI 应用场景 (使用层)"]
A["求职者投递简历"] --> B["企业 AI 筛选系统"]
B -->|"自动化拒绝"| C["求职者主张歧视 / 不公"]
end
subgraph StateEnforcement["州级现有监管与审查机制"]
C --> D["州反歧视法 / 消费者保护法"]
D --> E["调取 AI 审计追踪数据"]
E --> F["审查人机协同 (Human-in-the-loop)"]
F --> G["实施风险缓解措施"]
end
UserScenario --> StateEnforcement
```
阿明举例说明：当一名求职者申请工作被 AI 筛选系统淘汰，并怀疑自己遭遇了不公或歧视时，各州现有的劳工法、反歧视法和消费者保护机构本可以介入调查。然而，一旦联邦在“使用层”推行绝对优先权，州政府将失去要求企业提供 AI 决策链路追溯（Trace）、审查人机协同流程（Human in the loop）以及开展风险缓解审查的法定权力。她补充道，虽然去年底的某些联邦提案曾试图覆盖使用层，但近期较新的联邦立法草案已逐渐聚焦于国家级前沿模型层面，放开了对下游使用场景的过度干预。
---
### [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) 50州法律拼贴对创新的阻碍与联邦的“胡萝卜加大棒”
围绕科技大厂常提出的经典论调——“全美 50 个州若形成 50 套互不兼容的法律拼贴（Patchwork），将彻底摧毁创新，让田纳西州地下室里的初创开发者因雇不起合规律师而破产”，杜塔教授正面回应，明确承认这确实是一个非常现实的痛点。他强调在权衡监管体系时必须紧扣两大核心维度：
- **创新速率与财富创造**：监管体系是否会削弱国家孵化全新创意、保持技术领先和创造经济价值的能力；
- **就业保障与民生福祉**：监管体系能否有效保护劳动者权益，为普通民众提供优质的就业机会与生活保障。
针对各州作为“民主实验室”（Laboratories of Democracy）是否适用于 AI 的问题，多诺万通过一段生动的历史案例揭示了联邦政府干预州权的典型手段。他回忆道，历史上各州的最高车速限制原本属于州权范围，但当联邦政府希望推行全美统一的 55 英里/小时（55 mph）限速时，华盛顿采取了财政施压策略：“我们无权直接规定你们州的限速，但如果你们不改成 55 英里，我们就可以扣留联邦公路建设补助金。”最终所有州都被迫妥协。多诺万预言，即便在宪法权限存在争议的情况下，联邦政府未来依然会通过类似的杠杆手段强势介入 AI 治理。
### [11:24](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=684s) 州际竞争与国会僵局：“在政府里说‘不’永远比说‘行’更安全”
各州在人工智能监管上的态度差异，正在重塑美国的企业版图。正如“水往低处流”的自然法则，如果内华达州等地区选择打造对 AI 友好的营商环境，企业自然会蜂拥而至。然而，联邦政府绝不会轻易放弃管辖权，只要数据在电线或通信网络中跨州流动，华盛顿就会依据跨州贸易条款主张权力。这就导致不同行业不得不自行寻找监管阻力最小、政策红利最大的“AI 友好州”来落地业务。各州州长们频繁抱怨，若非华盛顿在国家层面行动迟缓、毫无作为，各州根本无需各自为政出台复杂的州级法案。
然而，寄希望于国会迅速立法无异于“痴人说梦”。前国会议员丹·多诺万（Dan Donovan）直言，联邦政府的运转极其迟缓。白宫的主人或许因任期限制而急于推进政治遗产，但国会两院的节奏完全由选举周期主导。当时众议院全部席位以及参议院三分之一的席位正面临改选，议员们在下一届国会于次年 1 月宣誓就职前几乎不会推进任何实质性法案。多诺万道出了官僚体制的生存哲学：“在政府运作中，说‘不’即使错了也容易脱身，而说‘行’一旦出错则代价巨大。因此，政客们在面对未知的 AI 议题时，第一反应永远是保守地说‘不’。”
- **联邦管辖权依据**：联邦政府通常以通信网络、跨州线缆数据传输为由，将 AI 活动纳入跨州商业条款（Interstate Commerce Clause）管辖。
- **政治周期停滞**：大选年面临众议院全员及参议院 1/3 席位改选，议员普遍规避争议性立法，导致联邦法案在次年 1 月新国会就职前全面停摆。
---
### [13:38](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=818s) 50 种定义的合规泥潭与全球算力竞赛
针对联邦与州立法的博弈，投资人兼顾问什雷亚·阿明（Shreya Amin）指出了两难困境背后的宏观矛盾。从全球竞争格局来看，当美国内部还在为“监管还是反监管”争论不休时，中国等竞争对手正在以每周数吉瓦（GW）的惊人速度扩张数据中心基础设施。面对这一态势，联邦政府最紧迫的任务并非全方位干预，而是确立统一的底层标准——在国家安全、网络安全以及通用技术术语层面给出清晰定义。
现实情况是，在 2024 至 2026 年间涌现的数百部州级 AI 提案中，绝大多数并非针对底层的模型基础提供商，而是聚焦于应用层场景。例如，全美出现了约 50 部针对 AI 聊天机器人的法案。尽管这些法案在技术本质上针对的是同一种模型风险与缓解机制，但 50 个州却给出了 50 种截然不同的法律定义。这迫使开发者必须为同一款产品做 50 套合规适配，给企业施加了极其沉重的运营负担。
```mermaid
flowchart TD
accTitle: AI法规碎片化与统一联邦标准对比
subgraph StatePatchwork["当前局面：州级立法碎片化"]
A["同一套 AI 聊天机器人底层技术"] --> B1["A 州法案：定义与合规要求 1"]
A --> B2["B 州法案：定义与合规要求 2"]
A --> B3["50 州各自出台独立法规..."]
B1 & B2 & B3 --> C["企业承担 50 次重复合规成本与风险"]
end
subgraph FederalStandard["业界诉求：联邦基线标准"]
D["联邦统一定义与国家安全基准"] --> E["通用合规框架 & 统一技术标准"]
E --> F["一次合规，全美通行"]
end
```
- **立法数量激增**：2024 年通过了 200 余部法案，2025 至 2026 年间又有上百部 AI 相关提案被引入各州议会。
- **定义碎片化**：仅 AI Chatbot 相关法规就有约 50 种不同表述，迫使技术团队面临重复开发的合规负担。
---
### [15:07](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=907s) 监管俘获与“欧洲化”隐忧：初创企业的生死劫
在大会展厅里汇聚了无数充满活力的年轻创业者，然而当前的政策乱局正在释放极其危险的信号。主持人尖锐地提出，复杂的州级法律补丁是否存在“监管俘获（Regulatory Capture）”的嫌疑——即资金雄厚的大型科技巨头能够轻松雇佣庞大的顶级律所团队搞定各州合规，而资源匮乏的初创团队则被直接挡在门外。
波图兰斯研究院（Portulans Institute）院长苏米特拉·杜塔（Soumitra Dutta）对此深表共鸣，并分享了他在欧洲长期生活与研究的惨痛教训。欧洲即便拥有欧盟单一市场的形式，内部市场实际上依然四分五裂；一个在法国起步的初创公司，光是满足全欧各成员国错综复杂的本地法规就已筋疲力尽，几乎无法成长壮大。杜塔警告，如果美国各州与联邦不能尽快理顺治理权属、拿出一套清晰协同的解决方案，美国可能会重蹈欧洲的覆辙。届时，高昂的跨州经商合规成本将彻底扼杀年轻一代创业者的创新活力。
- **监管俘获风险**：大型科技企业凭借雄厚法务资源建立合规护城河，中小创企因无力承担 50 州合规成本而被淘汰。
- **欧洲经验警示**：欧盟内部因碎片化监管导致法国等国初创企业跨国扩展极其艰难，美国若不解决各州法律冲突将面临相同困境。
---
### [17:03](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1023s) 严格标准与监管沙盒：犹他州与康涅狄格州的破局尝试
当联邦立法陷入僵局，行业标准与监管沙盒机制能否填补制度真空？什雷亚·阿明分析指出，在医疗健康和生物医药等高风险行业，企业为了生存，通常会主动对照全美“最严苛的标准”来构建底层系统，因为任何公司都无力承担针对各州定制 50 种不同技术架构的成本。但仅靠行业自律和最严标准仍无法解决全部问题，企业依然在很多具有不确定性的前沿领域畏首畏尾。
令人欣慰的是，部分州正在探索更具建设性的“监管沙盒（Regulatory Sandbox）”模式。犹他州在此领域走在最前列：政府建立沙盒机制，邀请监管机构与开发者同桌共议，企业提供极高的数据可见性与可观测性，以换取监管部门适度放宽合规约束，并在此框架下尝试了 AI 自主开处方药等前沿实验。康涅狄格州也在跟进类似的创新路径。这种将技术标准与实验室沙盒相结合的模式，能够推动“循证决策（Evidence-based policymaking）”，有效避免大选之年政客为迎合选民而仓促出台非理性禁令（例如纽约州在选举年直接暂停数据中心建设的极端做法）。
```mermaid
flowchart LR
accTitle: 监管沙盒运作机制
subgraph Sandbox["监管沙盒协作模型（如犹他州 / 康涅狄格州）"]
Ent["AI 创新企业"] -- "提供系统全透明度与可观测性" --> Reg["监管机构与政策制定者"]
Reg -- "放宽特定法规限制 / 允许前沿试点" --> Ent
Ent & Reg --> Evid["沉淀真实测试数据 & 循证政策"]
end
```
- **最严标准原则**：医疗医药等垂直领域企业普遍采取“就高不就低”策略，以全美最严标准研发单一产品线。
- **犹他州沙盒先例**：率先设立 AI 监管沙盒，在严格透明度要求下允许自主 AI 开具处方（Autonomous AI Prescriptions）等前沿场景测试。
- **政治情绪反例**：纽约州在选举年受政治情绪影响直接叫停部分数据中心建设，暴露了缺乏循证机制带来的政策风险。
---
### [18:39](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1119s) 选战年终极拷问：联邦优先权法案今年能否落地？
在小组讨论的尾声，主持人向三位嘉宾抛出了最核心的问题：“如果白宫今年向国会正式提交一项主张联邦优先权（Preemption Bill）、旨在废除各州自制 AI 法案的立法，它能在国会两院通过吗？”
嘉宾们的回答展现了残酷的政治现实：
- **前国会议员丹·多诺万（Dan Donovan）**斩钉截铁地表示绝无可能：“在新一届国会于次年 1 月就职之前，众议院不会通过任何实质性法案；而今年面临连任压力的三分之一参议员，甚至根本不敢就任何高风险议题进行表决。”
- **什雷亚·阿明（Shreya Amin）**同样持悲观态度，指出不仅宏观 AI 监管法案受阻，从底层数据中心审批到具身智能（Physical AI）与机器人技术的各项具体法案目前全都在国会陷入停滞。
- **苏米特拉·杜塔（Soumitra Dutta）**则幽默地表示完全赞同两位熟悉华盛顿内情的专家的判断。
主持人在掌声中总结道，纳税人的资金在运转，但华盛顿展现出的治理蓝图却令人沮丧。联邦与州在 AI 治理权上的角力远未结束，这将持续成为每一家跨州经营的技术企业必须直面的长期挑战。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) 当选国会议员只有两个基本要求：年满 25 岁且为美国公民——在赫兹（Hertz）租辆车的要求都比当国会议员要多 — Dan Donovan
> [00:04](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=4s) 不良系统会带来伤害，但安全、负责任的系统则能加速社会效益，监管与创新是“同一枚硬币的两面” — Shreya Amin
> [03:24](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=204s) 两者兼需 — Soumitra Dutta
> [08:51](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=531s) 我们无权直接规定你们州的限速，但如果你们不改成 55 英里，我们就可以扣留联邦公路建设补助金。 — Dan Donovan
> [11:24](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=684s) 在政府运作中，说‘不’即使错了也容易脱身，而说‘行’一旦出错则代价巨大。因此，政客们在面对未知的 AI 议题时，第一反应永远是保守地说‘不’。 — Dan Donovan
> [18:39](https://www.youtube.com/watch?v=cXaz0YSuKFg&t=1119s) 在新一届国会于次年 1 月就职之前，众议院不会通过任何实质性法案；而今年面临连任压力的三分之一参议员，甚至根本不敢就任何高风险议题进行表决。 — Dan Donovan
## Mentioned
Dan Donovan, Soumitra Dutta, Shreya Amin, Sanjay, RegulatingAI, Portulans Institute, Minsphere, Hertz, OpenAI, Global Innovation Index, 加利福尼亚州 (California), 康涅狄格州 (Connecticut), 科罗拉多州 (Colorado), 特拉华州 (Delaware), 纽约州 (New York), 犹他州 (Utah), 内华达州 (Nevada), 田纳西州 (Tennessee), 欧洲联盟 (European Union), 美国国会 (U.S. Congress)
## Video description
Dan Donovan - Congressman, U.S. Congress
Soumitra Dutta - President, Portulans Institute
Shreya Amin - AI Advisor / Investor, Minsphere
A multi-level debate bringing together sitting Attorney Generals, Governors, and Members of Congress to examine the most explosive fault line in U.S. AI governance: federal preemption of state AI laws. Connecticut just passed sweeping AI legislation (May 2026), Delaware is pioneering a first-in-nation OpenAI certification partnership, and Colorado's AI Act takes effect June 30, 2026 — while Washington pushes for a single national standard. Who wins, and what does it mean for every company operating across state lines?
[Source](https://www.youtube.com/watch?v=cXaz0YSuKFg)
Source quotes
[0:04] Dan Donovan: 当选国会议员只有两个基本要求：年满 25 岁且为美国公民——在赫兹（Hertz）租辆车的要求都比当国会议员要多
[0:04] Shreya Amin: 不良系统会带来伤害，但安全、负责任的系统则能加速社会效益，监管与创新是“同一枚硬币的两面”
[3:24] Soumitra Dutta: 两者兼需
[8:51] Dan Donovan: 我们无权直接规定你们州的限速，但如果你们不改成 55 英里，我们就可以扣留联邦公路建设补助金。
[11:24] Dan Donovan: 在政府运作中，说‘不’即使错了也容易脱身，而说‘行’一旦出错则代价巨大。因此，政客们在面对未知的 AI 议题时，第一反应永远是保守地说‘不’。
[18:39] Dan Donovan: 在新一届国会于次年 1 月就职之前，众议院不会通过任何实质性法案；而今年面临连任压力的三分之一参议员，甚至根本不敢就任何高风险议题进行表决。
Source quotes
[0:04] Dan Donovan: The only two requirements to become a member of Congress is you have to be over 25 and a citizen of the United States. There's more requirements to rent a car from Hertz than to become a member of Congress, and they are the people that are going to be regulating this.
[2:34] Sanjay: lawyer answers
[5:23] Shreya Amin: build
[5:23] Shreya Amin: use
[8:51] Soumitra Dutta: definite yes
[8:51] Dan Donovan: goes over a wire
[8:51] Dan Donovan: AI states
[12:07] Dan Donovan: It's easier in government to say no and be wrong than to say yes and be wrong. So, they caution on the side of no.
[19:09] afraid to vote on anything
[19:09] two esteemed colleagues
[19:09] We got to get something done, man. I mean, our—these are our tax dollars at work.
Original transcript
[0:04] Federal versus state, it's a pretty interesting
conversation. Uh, lot of states are kind of
[0:10] writing uh, different AI laws and the federal
government, I think you heard from Nate,
[0:14] is trying to have a preemption of those state
laws. And that's not just uh, conversation about
[0:21] federalism, that's actually happening. So, who's
going to decide how AI impacts you, whether it's
[0:28] your state or is it going to be Washington
D.C.? Well, we have three great panelists here
[0:35] to kind of enlighten us on that and they
have hopefully three different views.
[0:40] So, we will start uh, and we'll follow this format
today is ask you to introduce yourself and answer
[0:47] one common questions and the common question
is who should monitor and regulate AI? Should
[0:55] it be the state or should it be Washington
D.C.? We'll start with you, Congressman.
[0:59] My name is Dan Donovan. My name is Dan Donovan.
I'm a advisory board member of regulating AI,
[1:05] which is not for profit led by Sanjay, whose
purpose is to inspire innovation but at the
[1:11] same time curtail malicious activity uh, by bad
actors. Uh, prior to this I was a member of the
[1:18] United States Congress and those are the people
who are writing your laws and I just want you to
[1:23] keep in mind the only two requirements to become
a member of Congress is you have to be over 25
[1:29] and a citizen of the United States. There's
more requirements to rent a car from Hertz
[1:34] than to become a member of Congress and they are
the people that are going to be regulating this.
[1:41] Hi, I'm Shreya. Um, I I was the first
AI officer for New York State last year.
[1:47] Before that, I've been in this industry close
to 20 years. Uh, chief AI scientist at a health
[1:53] tech company where we kind of built and deployed
the platform in hundreds of hospitals before that
[1:59] uh machine learning systems in school districts.
Um so, all this to say kind of I've been on that
[2:05] the the industry side, and then last year I
kind of switched over to the public sector side.
[2:10] And what when you build these systems where if
you build it badly, it causes harm. So, you want
[2:15] to avoid that, what we talked about. But, there's
a flip side to that. When you actually build safe,
[2:20] accountable systems that are beneficial, you'd
build want to build pathways that'll help you
[2:24] accelerate that. Right? So, which is why this
topic is so close to my heart, and I think it's
[2:29] two sides of the same coin. So, really excited
to be here. Soumitra. Hello, everyone. So,
[2:34] my name is Soumitra, and I'm a former professor
and dean from Cornell University in New York, and
[2:41] from Oxford in the UK. A lot of my work over the
years has been around AI technology innovation.
[2:49] I did my PhD many years ago in AI from
Berkeley, but more of the last couple of
[2:55] decades I've done a lot of work in innovation,
having created the global innovation index,
[3:00] which is adopted by the United Nations, used
by more than 100 countries around the world.
[3:05] And a lot of my work really revolves around how
do businesses, how do governments actually adopt
[3:11] technology, adopt AI now, of course, as a
specific case of interest. Who should be
[3:16] regulating AI? States or Washington? They need a
clear answer. the answer always is case the only
[3:24] politician here. Everybody avoided the question.
I know. We don't want lawyer answers. Okay. So,
[3:29] if you ask me, the answer is clearly both. And
the question really is who does which part of it?
[3:36] So, look at typically how regulation
operates in, let's say, in the US.
[3:42] The federal government regulates those parts
which typically are best done centrally,
[3:49] like, for example, nuclear energy or aviation,
where there are rules that apply to everyone. Then
[3:54] there are some parts that the federal government
actually has a mixed approach. Where the federal
[4:00] government does regulate some parts, let's say
the environment has some standards, but the states
[4:04] are given some freedom and leeway to be able to
enforce more stricter standards if they wish to.
[4:09] And then there's some other parts, let's say for
example law or professions like medicine or other
[4:14] aspects which are completely given to the states
and given more leeway. Now the challenge in AI
[4:18] is of course AI changes and is the it's sort of a
changes the boundaries in a fairly flexible manner
[4:24] based on which part of the stack the previous
speaker, in fact, talked to the stack. And if
[4:29] you look at for example building AI, the building
AI actually takes input from all over the country,
[4:35] in fact, all over the world one might say, but
all over the country. And in fact is subject
[4:39] to a lot of concerns about you know, open source
models from China or other countries and so on.
[4:45] So there's a need for some kind of central federal
regulations about what is allowed, what is not
[4:51] allowed, what kind of, you know, constraints have
to be put. At the same time the people who use,
[4:56] whether in the city or whether in the companies,
they actually use AI at a much more local level,
[5:02] much more contextual. The context become much
more important out there. So you need much more,
[5:06] let's say, local rules, local regulations that
govern how and what kind of data should be used
[5:11] more regionally. So the answer, Sanjay, is
not easy, but the answer is actually both.
[5:16] And we are trying to figure out, I think, we're
in the process of figuring out which part and
[5:19] who does what. I think I'm going to stick to
the congressman now. Because he is giving us
[5:25] a yes and a no answer. Congressman, let's
say if the federal government were to preempt
[5:31] um the AI laws as they want to do with different
states, what happens to the laws that states like
[5:37] California, Connecticut, Colorado and others have
passed? What happens in that case? As long as they
[5:42] don't violate whatever the federal government
lays out as going to be the base for this.
[5:48] And you're going to find my my fellow panelists is
right on. There's going to be federal regulations,
[5:53] there's going to be state regulations. You're
going to find states that are more AI-friendly,
[5:58] and businesses will go there to do business. It's
just like any other industry. If you're allowing
[6:03] data centers to be built in your state, we have
a moratorium on them right now in New York. So,
[6:07] no one's going to be coming to do AI business in
New York. They'll go to friendly states where the
[6:13] regulations are I don't want to say more lenient,
but maybe more business-friendly, maybe more
[6:19] innovation um you know, guided um policies that
that help companies grow rather than stifle them.
[6:28] So, uh Sherrod, you were running AI for state of
New York, and there were governance infrastructure
[6:36] within the state, et cetera. So, if the uh
federal government had preempted and said no
[6:44] state AI laws, would it have impacted the
work that you were doing at that time?
[6:51] Yeah, I'm going to piggyback a little bit here and
Yeah, but no waffling. Okay, no waffling. Got it.
[6:56] Um I think uh the previous speaker like if it's
at the frontier model companies, right? So, if you
[7:02] take these handful of companies who are building
AI models, if you preempt that and regulate them,
[7:09] what falls on you at the state level is a little
bit of, you know, uh um uh implementing that,
[7:16] enforcing that with the right leadership that's
there. So, not so much. But, if you preempt usage
[7:22] of it, right? So, I think someone spoke about
build and use, it's a great starting architecture,
[7:26] not not complete, but so if you regulate kind of
how the state can use it. So, think of it this
[7:31] way, a worker uh applies to a job, uh AI system
is screening the application, she's rejected,
[7:39] right? So, So she can come in and say, "Wait
a minute, this was unfair." Unfair meaning,
[7:44] I think I was discriminated against. So the
state already has discrimination laws. There
[7:48] are regulators who do that anyway. So you have
consumer protection laws, workflow workforce laws.
[7:55] What happens when a federal government preempts
regulation at the use level? Now my ability
[8:01] to build you give you systems for you to probe
that is going away, right? So I can't put in
[8:07] systems that says, "Give me the trace of how
this AI which systems did it touch? Did we follow
[8:13] the right human in the loop?" There's there's
there's a lot of risk mitigation that you can do.
[8:18] So if you preempt at that level, then 100%
it gets affected. If you're Oh my you know
[8:24] what he was saying, which is like if you're
doing it at national security model level, um
[8:30] geopolitical race, all of those pieces, I
think they're not so much. So it depends on
[8:34] where you are at. And actually right now a lot
of federal regulation, the recent versions are
[8:40] not preempting the usage part. I think
there was an earlier version last December,
[8:44] which was. And right now a lot of it is at the
national model level. So uh one of the arguments
[8:51] that you hear from the federal government, from
all these big uh frontier labs is that if you have
[8:57] a patchwork of 50 laws with 50 different states,
it's going to hurt innovation. The entrepreneur
[9:04] in a basement in Tennessee is going to have to
have uh can't afford lawyers, etc. I'm going
[9:09] to start with you, Soumitra, because you did the
innovation index. And then I'll hopefully get an
[9:17] answer from you because this is the argument that
keeps coming at us. Can't have 50 patchwork laws.
[9:22] Just briefly because we have uh you know, Eric uh
hovering in the wings, but is that a good argument
[9:30] about innovation getting hurt if you have 50 so
to speak different laws? Yes, that's an issue.
[9:36] And let me say that's a definite yes. Definite
yes. So, no waffling this time. But, I think what
[9:41] is important for us is to also identify what I
think, in my view, the two important issues facing
[9:47] us. One is how is governance regulation affecting
the rate of innovation? And second is how it is
[9:54] affecting actually employment and employability.
I think those two important elements, how are we
[9:59] as a nation able to create new innovative ideas
and create value and wealth? And are we able to
[10:05] look after our people and guarantee or give them
in fact good jobs and good livelihoods? I think
[10:09] those are the two important dimensions that we
have to worry about regulation. So, Congressman,
[10:15] you know, we were you if you do a law class or
civics class, etc., states are the laboratory of,
[10:23] you know, some of the biggest laws that happen
in this country. But, is AI so different that
[10:30] we need the federal preemption that we kind of
say states kind of have to stay out of this?
[10:42] Well, I'm going to I'm going to take this. I think
the federal government is preempting your mic.
[10:50] Is it on? Oh, yeah. Uh so, so the federal
government's going to have their hand in this.
[10:54] I mean, even things that are delegated to
the states. I'm old enough to remember when
[10:59] all states had various because it was a state
authority that dictated what the speed limit
[11:06] So, the federal government wanted everybody
to drive at 55 mph. They said, "Well,
[11:11] we can't dictate your state speed limit, but we
don't have to give you federal highway money."
[11:17] And everybody went to 55 mph. So, they're going
to have their hand in this. But, I think the
[11:24] water is going to reach its level. If Nevada
ends up being an AI-friendly business state,
[11:34] Businesses will flock here to do their business.
And the other thing about the federal government
[11:38] is they're always going to grab jurisdiction cuz
they're going to say if anything goes over a wire,
[11:42] a telephone connection, anything, it's ours cuz
it's in the state. So, I I think there's going
[11:48] to be a balance, but I think you're going to find
that the industries are going to find their level
[11:53] of where they want where they're I don't want
to say least regulated, but where it benefits
[11:59] them the most, and you're going to find we're
going to have AI states. AI states. So, uh Shrab,
[12:07] uh when you talk to governors, you know, of the
states that have been enacted these laws, they
[12:13] say, "We wouldn't have to do this if Washington
would get its act together. If there was a federal
[12:20] so to speak regulation, states would not have
that." And this is a question for you, too.
[12:28] Getting something passed in Washington is
really kind of a fool's errand in many ways,
[12:33] to be very frank. Uh so, should states just keep
waiting, or should states uh should the uh just
[12:41] the federal government get in the mix? How How
is this going to play out uh for us? And Shrab,
[12:47] want your opinion, too. Federal government does
nothing fast. Other than the fact that you have
[12:53] a man who sits at 1600 Pennsylvania Avenue right
now, who only has 2 and 1/2 years to do everything
[12:57] he wants to do, regardless of whether he can run
for a third term. Uh so, so you're going to see
[13:03] some fast movement, but you're going to see slow
movement in the chambers where I used to work.
[13:11] Um members of Congress are up for re-election
this year, at least the entire House is,
[13:15] and 1/3 of Senate is. They're going to do nothing
until they have to the next Congress takes office
[13:20] in January 3rd or 4th. Then they're going to start
running for re-election again. And it's easier in
[13:27] government to say no and be wrong than to say yes
and be wrong. So, they caution on the side of no.
[13:38] Yeah, I agree with that. I think there's so much
of uh I think if you take data centers, right?
[13:45] Like and I think someone spoke about the entire AI
value chain. It's it's absurd at this point of uh
[13:51] how much people are either pro-regulation or
anti-regulation meanwhile China's building data
[13:56] centers that are like what? Some gigawatts per
week or something, right? So, I think the idea
[14:02] that if the federal government had something,
I think it's yes and no. Like yes, it's true.
[14:06] I think we do need the federal government to give
us something. Give us common definitions, common
[14:10] requirements, cybersecurity, national security.
Let's get all that written down somewhere.
[14:15] Uh beyond that, I think if you see the lot of the
like over 200 bills that have passed 2024 100 over
[14:21] 2025 to 2026 AI bills, a lot of it is has nothing
to do with any of the model provider companies.
[14:27] Almost all of them are how do you use it? So, you
have about 50 AI chatbot bills. Now, the question
[14:34] becomes if technically it's the same thing, I want
to avoid risk, I want to mitigate it this way,
[14:39] but I have 50 definitions of that technical thing.
Then I'm am I asking a company to do it 50 times?
[14:45] There is definitely a burden. Um so, I think yes,
federal regulation should be there. But short of
[14:52] that, I think a coalition or alignment is also
not there. So, we're defining things 50 different
[14:56] ways. Um you know, in some sense you look at it
as an experiment, but I don't know how much time
[15:01] we have for that. So, yeah. Uh Sumitra, we are at
AI 4. You see the entrepreneurs, the innovation.
[15:08] All you have to do is go to the exhibit hall, etc.
If you're an entrepreneur and you see the current
[15:17] you know, environment that we have here, states
added, federal government saying we're going to
[15:23] preempt it, but not happening. I mean, you've
done innovation, you say 50 state laws are not
[15:30] necessary. How does an a small
entrepreneur think? Do you think there is a
[15:38] element of regulatory capture by some of
these big companies who who can afford
[15:44] large law firms, lawyers, etc. that can deal with
many of these things, but you know, what is the
[15:52] message that we are sending? We want innovation,
but what is the message that is going out to all
[15:58] these folks in the exhibit booth here? So, I think
you hit the nail on the head out here, and one of
[16:03] my biggest worries is are we going to dampen
the innovative spirit of young entrepreneurs,
[16:09] especially if you have a patchwork of laws. And
I have seen this happen. I lived for many years
[16:14] in Europe. In Europe, as you know, despite the
European Union, the the the market is fragmented.
[16:21] And for a small entrepreneur who starts a company
in France, it's almost impossible for that person
[16:26] to satisfy all the conditions and the reg-
and the and the different regulations that
[16:30] are there in different parts of the of the
continent. I fear that if we do not get this
[16:36] messy situation sorted out in a cleaner manner,
we might end up with a similar outcome out here.
[16:42] Where people might find it very expensive to
do business out here, and in fact meet the
[16:46] constraints of all the different states actually
have to do business in. So, I really think it's
[16:51] very important for legislators and the policy
leaders across both federal and state to come
[16:56] up with some kind of a good solution to this so
that we don't dampen innovation out here. So, uh
[17:03] Again, this is a question for all three of you. If
we can't get regulation in place, is there a role
[17:11] for standards to take place instead of regulation
because it doesn't seem like as the congressman
[17:18] said, we are now in an election mode and we have
a panel on that coming up here on the ballot.
[17:24] Uh share, would standards play any kind of role
in this? I mean, I think we have standards that
[17:29] people are developing, right? And I think a lot
of kind of at least in healthcare and pharma,
[17:34] like what companies are doing are building to
the stricter standard. Just you can't afford to
[17:38] have 50 different products. So, you try to build
to the stricter standard as far as use goes. Um
[17:44] and so I think it helps, but I still think we need
regulations because if you see folks are kind of
[17:50] shying away from certain topics. And I think
we're seeing good innovation. Like Utah has a
[17:56] place where you could a regulatory sandbox is
what they're calling it, which means come in,
[18:02] we'll relax some of the constraints, we'll bring
the regulators at the table. You give us extra
[18:07] visibility, observability, and we will relax
them. So, they have autonomous AI prescriptions.
[18:13] There arguments on both sides, but at least
they're taking that shot in a good way. I think
[18:17] Connecticut is going after that kind of a model.
So, I think if you combine standards with these
[18:22] laboratory sandbox approach, you might start
seeing results. And then your policies can be
[18:28] more evidence-based versus kind of we just don't
agree and on an election year, New York state says
[18:34] no more data centers, right? So, we want to avoid
those problems, yeah. So, uh I would like to get
[18:39] an answer from all three of you. Do you think
if the uh White House puts in a preemption bill
[18:47] in the Senate or in the House uh in US
Congress this year, will it pass, congressman?
[18:57] Um nothing's going to pass the
House until the next Congress takes
[19:02] uh takes place in January. Um and a third
of the senators who are up for re-election
[19:09] this year would be afraid to vote on anything.
So, it's not going to pass. Right now. Sure.
[19:15] I agree. I don't think it passes. I
think there are a lot of bills that
[19:18] are getting stalled from data centers, AI,
physical AI, robotics, name it. So, yeah.
[19:24] I defer to my two esteemed colleagues.
They're more knowledgeable.
[19:30] We got to get something done, man. I mean, our
These are our tax dollars at work, but anyway,
[19:37] uh I've got uh Eric uh out there, but for folks
who live in Washington, it's not a good uh
[19:45] uh picture that is coming out, but thanks
so much to our esteemed panel for telling
[19:50] us about Washington versus state. So,
thank you so much. [cheering] [applause]
