U.S. bill proposes permanent ban on artificial superintelligence and creation of new federal AI agency

【Executive Summary & Core Event】

On September 23, Senator Bernie Sanders (I-VT) and Representative Greg Casar (D-TX) introduced a bill that would permanently prohibit the development, deployment, and use of artificial superintelligence within the United States. The legislation defines ASI as any AI system that substantially exceeds human cognitive performance across virtually all economically valuable domains of intellectual activity, drawing conceptual boundaries from Nick Bostrom's foundational superintelligence definitions. The bill also proposes establishing a new federal agency—tentatively referenced as the Federal AI Authority or equivalent—tasked with monitoring frontier AI development, certifying model compliance, and enforcing prohibitions against superintelligence creation.

The legislation arrives amid intensifying global debate over frontier AI governance, following the EU AI Act's phased implementation, the UK's AI Safety Institute establishment, and executive orders from the Biden administration requiring safety testing disclosures for models exceeding specific training compute thresholds. Unlike these approaches, which emphasize risk mitigation and transparency, the Sanders-Casar bill adopts a hard prohibitionist stance—criminalizing not just deployment but the fundamental research and development of systems that could plausibly reach superintelligence. This positions the bill as the most restrictive AI governance proposal introduced in any major Western legislature to date, though its legislative prospects remain uncertain given industry opposition and the technical challenges of defining and detecting ASI thresholds.

【Technical Architecture & Key Innovations】

The bill's central technical challenge lies in operationally defining 'artificial superintelligence' in a manner that is both legally enforceable and scientifically coherent. The legislation reportedly draws on capability-based definitions rather than architectural specifications, meaning the prohibition would apply to any system—regardless of underlying architecture, whether transformer-based, mixture-of-experts, neurosymbolic, or novel paradigms—that demonstrates performance exceeding human expert baselines across comprehensive cognitive benchmarks. This approach mirrors methodologies proposed by AI safety researchers including Bostrom, Yudkowsky, and more recently Christiano and Cotra, who have advocated for capability evaluation frameworks as governance instruments. However, the bill faces significant measurement challenges: current benchmark suites like MMLU, GPQA, ARC-AGI, and SWE-bench remain imperfect proxies for general intelligence, and the absence of consensus on ASI detection thresholds creates enforcement ambiguities.

From a technical enforcement perspective, the proposed federal agency would need to develop sophisticated model auditing capabilities, including access to training compute logs, weight inspection protocols, behavioral evaluation harnesses, and potentially runtime monitoring systems for deployed models. The bill's prohibition on development—not merely deployment—implies requirements for pre-training registration, compute threshold monitoring (likely building on the Biden executive order's 10^26 FLOP reporting threshold), and red-team verification of capability emergence during training. This creates substantial technical infrastructure demands: the agency would require expertise in mechanistic interpretability, scaling law analysis, and emergent capability detection. Critics note that defining 'substantially exceeds human cognitive performance' remains philosophically and technically contested, with some researchers arguing that superintelligence may emerge gradually through distributed systems rather than discrete model deployments, complicating any single-model enforcement framework.

【Industry Context & Competitive Landscape】

The Sanders-Casar bill positions the United States at the extreme regulatory end of the global AI governance spectrum, contrasting sharply with the EU AI Act's risk-tiered approach, the UK's voluntary safety-testing framework, and China's algorithm-specific regulations. Where the EU prohibits certain AI applications (social scoring, real-time biometric identification) while permitting frontier model development under safety obligations, this bill would fundamentally constrain the trajectory of American frontier AI labs including OpenAI, Anthropic, Google DeepMind, Meta AI, and emerging players like xAI and DeepSeek's international operations. The legislation effectively challenges the stated missions of OpenAI (whose founding charter references AGI) and Anthropic (whose safety research explicitly contemplates steering systems beyond human-level capability), creating existential strategic questions for these organizations should similar provisions gain legislative traction.

Competitive dynamics introduce additional complexity. A unilateral U.S. ban on ASI development would not constrain parallel efforts in China, where entities including DeepSeek, Qwen (Alibaba), and Zhipu AI continue advancing frontier capabilities, nor in other jurisdictions with permissive regulatory environments. This creates what governance scholars term a 'regulatory asymmetry problem'—where prohibition in one jurisdiction potentially cedes strategic AI advantage to others. Industry advocates argue this could drive talent and capital to less restrictive jurisdictions, mirroring concerns raised during crypto and biotechnology regulatory debates. However, proponents counter that the bill's enforcement mechanisms could include extraterritorial provisions similar to export controls, and that the establishment of a federal AI agency would centralize oversight currently fragmented across NIST, FTC, SEC, and agency-specific frameworks. The bill also intersects with ongoing discussions about compute governance, as effective ASI prohibition likely requires monitoring of advanced semiconductor clusters—a domain where U.S. export controls on TSMC, NVIDIA, and ASML technologies already provide enforcement infrastructure.

【Developer & Enterprise Implications】

For developers and enterprises, the bill's immediate implications depend critically on how 'superintelligence' thresholds are operationalized. If the prohibition applies only to systems demonstrably exceeding human performance across all cognitive domains—a high bar that current frontier models including GPT-4, Claude 3.5, and Gemini 1.5 do not meet—then near-term commercial AI development would continue largely unaffected. However, the bill's prohibition on development toward ASI could be interpreted to restrict research trajectories explicitly aimed at surpassing human capabilities, potentially impacting OpenAI's stated AGI objectives, Anthropic's capability scaling research, and academic work on recursive self-improvement. This ambiguity creates significant compliance uncertainty for organizations whose research roadmaps contemplate human-level or superhuman systems.

The proposed federal AI agency would introduce new regulatory overhead for frontier model developers, including registration requirements, pre-deployment safety audits, ongoing capability monitoring, and potential mandatory red-team access. For enterprise consumers of AI services, the agency may establish certification standards affecting procurement decisions, particularly in regulated sectors like healthcare, finance, and defense. Implementation costs could be substantial: building a technically competent federal agency requires recruiting top-tier AI researchers, safety engineers, and policy experts in an intensely competitive talent market where senior AI researchers command compensation packages exceeding $1 million annually. The bill's enforcement mechanisms—reportedly including civil penalties and potential criminal liability for willful ASI development violations—would necessitate robust compliance programs at frontier labs, potentially redirecting R&D investment toward capability verification and safety research rather than raw scaling. Smaller developers and open-source communities face particular challenges, as compliance costs could disproportionately burden organizations lacking dedicated legal and safety teams, potentially consolidating frontier AI development among well-capitalized incumbents capable of absorbing regulatory overhead.

【Key Takeaways & Strategic Outlook】

The Sanders-Casar bill represents a paradigm shift in AI governance discourse—moving from risk management and transparency frameworks toward outright prohibition of a specific capability frontier. Regardless of its legislative prospects, the bill signals growing political appetite for aggressive AI regulation and establishes a policy anchor that may influence future legislative negotiations. The proposal's most significant contribution may be elevating the 'ASI prohibition' concept from academic safety literature into mainstream political discourse, potentially reshaping how frontier AI labs communicate about their long-term objectives and how investors evaluate AI company trajectories. The creation of a dedicated federal AI agency, even if decoupled from the ASI ban, addresses a genuine governance gap and may gain bipartisan support as a standalone initiative.

Strategically, the bill highlights fundamental tensions in AI governance: between unilateral action and international coordination, between prohibition and managed progress, between safety precautionism and competitive necessity. The technical challenge of defining and detecting ASI thresholds remains unsolved, and any enforcement framework will require unprecedented collaboration between government, academia, and industry on capability evaluation methodologies. For AI organizations, the bill underscores the importance of proactive safety research, transparent capability reporting, and engagement with governance frameworks—not merely as compliance activities but as strategic imperatives that shape the regulatory environment in which they operate. The coming legislative cycle will likely see modified versions of these proposals, with the federal AI agency concept having greater viability than the absolute ASI prohibition, but the Sanders-Casar bill has established a new ceiling for regulatory ambition that will define AI policy debates for years to come.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.