Who Will Control Frontier AI: Washington, States, or Independent Boards?

WASHINGTON — Major AI developers, state regulators, and federal policymakers are building competing systems for governing frontier artificial intelligence models. The proposals include centralized federal licensing agencies, independent auditing bodies modeled on international accounting systems, and state compliance requirements.

Before Congress settles on a framework, OpenAI, Anthropic, Google DeepMind, and Microsoft have established internal risk management programs and governance bodies ahead of formal statutory mandates. Major laboratories also publish “system cards” with new model releases, detailing operational capabilities, safety evaluation results, architectural constraints, and known failure modes.

Microsoft’s internal oversight relies on a corporate Responsible AI Standard drafted by senior executives and enforced by an Office of Responsible AI. Dedicated staff are embedded across engineering teams, while the company’s Sensitive Uses review committee has evaluated more than 1,900 deployment cases since 2019, including 450 in the past year. The committee blocked the integration of real-time facial recognition technology onto police body cameras, and the entire compliance structure reports to a dedicated committee on Microsoft’s board of directors.

The federal debate intensified after the President wrote on social media that the primary safeguard required for artificial intelligence is “a STRONG AND SMART (High IQ!) PRESIDENT.” The post came amid administrative changes following the Supreme Court’s decision in *Trump v. Slaughter*.

Under that ruling, federal administrative agencies—with the exception of the Federal Reserve—answer directly to executive control. Any newly established federal AI regulatory body would therefore operate under White House direction and be subject to direct presidential removal authority. The precedent has complicated Capitol Hill efforts to create long-term oversight.

Sen. Bernie Sanders (I-Vt.) has called for a temporary moratorium on advanced model releases until Congress creates a dedicated Cabinet-level agency to oversee the industry. Sen. John Thune (R-S.D.) has proposed a federal office with statutory authority to block the public release of frontier models that fail risk assessments.

The Department of the Treasury has drafted a different proposal: a self-regulatory organization similar to the Financial Industry Regulatory Authority (FINRA) in the financial sector that would report directly to Treasury officials. White House AI Adviser David Sacks, however, has warned laboratory executives against using safety discussions to bypass competition rules, telling companies to “stop pretending antitrust law has to be suspended so you can form a cartel.”

Google DeepMind used multi-disciplinary review procedures before deploying AlphaFold, its protein-structure prediction system. The company consulted more than 30 external experts in biosecurity, biology, bioethics, and human rights to assess potential dual-use biological risks.

Anthropic has also imposed voluntary release controls. It withheld its most advanced AI model from broad public distribution because of identified safety concerns, initially limiting access to a select group of defensive cybersecurity partners before expanding public deployment after federal sign-off.

Pre-release “red-teaming” is another part of current industry practice. These adversarial evaluations are designed to identify vulnerabilities such as automated prompt injections and dangerous knowledge generation involving chemical, biological, radiological, or nuclear domains. Industry coordination has expanded through the Frontier Model Forum, an industry coalition formed by leading laboratories to establish shared safety benchmarks and evaluation practices.

External evaluation organizations have increasingly received pre-release access to frontier AI models for systemic-risk assessments. METR (Model Evaluation and Threat Research), an independent evaluation group, conducted a month-long assessment of unreleased frontier models from four major AI laboratories simultaneously.

Separately, the U.K. AI Safety Institute evaluated OpenAI’s GPT-5.5 before its launch and identified a universal jailbreak vulnerability that allowed internal safety controls to be bypassed. California rules now require developers of the largest AI models to publish safety frameworks identifying capability thresholds that would halt a release, report serious security incidents to state authorities within 15 days, and submit to evaluations by state-registered independent inspectors.

The European Union’s Artificial Intelligence Act structured much of its compliance framework around internal review procedures, risk monitoring protocols, and executive oversight mechanisms that private technology firms had originally established. State regulators’ moves have added another layer to the divide among Washington policymakers, state authorities, and technology developers.

Legal experts and industry analysts examining long-term governance have pointed to financial accounting standards as a potential operational model. In 1973, professional accounting organizations established the International Accounting Standards Committee (IASC) in London as a private industry body.

Following pressure from international regulators, including the U.S. Securities and Exchange Commission (SEC), the organization was restructured in 2001 as the International Accounting Standards Board (IASB). Its board members became full-time professionals severed from industry financial ties and were funded through an independent non-profit foundation.

In 2009, an international Monitoring Board comprising securities regulators, including the SEC and the European Commission, was established. It received veto power over board appointments without directly dictating technical standard-setting. More than 140 jurisdictions now mandate IASB standards for corporate financial reporting.

The United States uses a similar independent structure through the Financial Accounting Standards Board (FASB), maintained by the non-profit Financial Accounting Foundation and formally recognized under federal securities laws. Proponents of applying this template to artificial intelligence argue that an independent, non-governmental standards board—funded by independent foundations, monitored by international regulatory bodies, and verified by certified third-party auditors—could establish globally accepted safety benchmarks without requiring centralized federal enforcement agencies or triggering cross-border regulatory conflicts.