Elizabeth Kelly: Laying the Foundation for the U.S. Artificial Intelligence Safety Institute Custom Case Solution & Analysis

Evidence Brief: U.S. Artificial Intelligence Safety Institute

1. Financial Metrics

• Initial Funding: 10 million dollars reallocated from the CHIPS and Science Act for fiscal year 2024. (Paragraph 12)
• Budget Request: 50 million dollars requested by the Biden-Harris administration for fiscal year 2025 to sustain operations. (Exhibit 4)
• Comparative Scale: The 10 million dollar budget represents less than 0.1 percent of the annual R and D spend of major frontier AI labs like Microsoft or Google. (Paragraph 15)
• Personnel Costs: Federal pay scales for technical talent capped at approximately 183300 dollars, significantly below private sector total compensation packages exceeding 500000 dollars. (Exhibit 2)

2. Operational Facts

• Organizational Structure: Housed within the National Institute of Standards and Technology (NIST) under the Department of Commerce. (Paragraph 4)
• Mandate: Established via Executive Order 14110 to develop standards, tools, and tests for AI safety and security. (Paragraph 2)
• Core Functions: Red-teaming protocols, evaluation of dual-use foundation models, and development of risk management frameworks. (Paragraph 8)
• International Context: Modeled partly after the United Kingdom AI Safety Institute, necessitating cross-border technical coordination. (Paragraph 22)

3. Stakeholder Positions

• Elizabeth Kelly (Director): Prioritizes building technical credibility and institutional permanence within the federal bureaucracy. (Paragraph 6)
• Gina Raimondo (Secretary of Commerce): Views the institute as a critical bridge between national security requirements and economic competitiveness. (Paragraph 10)
• Frontier Model Labs (OpenAI, Anthropic, Google): Publicly support safety standards but remain protective of proprietary weights and training data. (Paragraph 28)
• Congressional Critics: Divided between those demanding more oversight of tech giants and those fearing over-regulation will stifle American innovation. (Paragraph 31)

4. Information Gaps

• Compute Access: The case does not specify the amount of compute power available to the institute for independent model verification.
• Enforcement Mechanisms: Absence of clear legal authority to compel private labs to submit models for testing prior to public release.
• Staffing Levels: Exact number of PhD-level researchers successfully recruited versus vacancies as of the case date.

Strategic Analysis

1. Core Strategic Question

• How can the U.S. AI Safety Institute establish a definitive technical authority and secure industry compliance without formal regulatory powers or market-competitive financial resources?

2. Structural Analysis

The institute operates in a high-stakes environment where the pace of private innovation exceeds the speed of public policy. Applying a Value Chain lens to AI development, the institute seeks to insert itself at the Evaluation and Validation stage. However, it faces a structural disadvantage in the Resource-Based View (RBV). It lacks the specialized compute assets and the financial capital to attract top-tier human capital away from firms like Anthropic or OpenAI. Its primary asset is its institutional neutrality and its proximity to the Department of Commerce, which provides a platform for setting industry-wide norms.

3. Strategic Options

4. Preliminary Recommendation

The institute must pursue the Technical Specialist path. Credibility in the AI community is built on technical proficiency, not policy white papers. By developing superior evaluation tools that the labs themselves find useful for internal safety checks, the institute creates a voluntary pull-factor. This technical authority is the only viable precursor to future regulatory influence. Attempting to be a coordinator or a policy bridge before proving technical mastery will result in institutional irrelevance.

Implementation Roadmap

1. Critical Path

• Month 1-2: Finalize the technical leadership team using Intergovernmental Personnel Act (IPA) agreements to bypass federal salary caps.
• Month 3-4: Secure Memoranda of Understanding (MOUs) with at least three major frontier labs for pre-release access to specific model versions.
• Month 5-6: Publish the first set of standardized benchmarks for dual-use capabilities in chemical and biological risk.
• Dependency: Successful recruitment of a Chief Scientist is the prerequisite for all technical workstreams.

2. Key Constraints

• Talent Scarcity: The institute is competing for a global pool of fewer than 500 individuals capable of auditing frontier models.
• Political Volatility: Funding is tied to annual appropriations, making long-term research hiring difficult.
• Compute Access: Without a dedicated national AI research resource, the institute remains dependent on the very companies it is meant to oversee for the hardware required to run evaluations.

3. Risk-Adjusted Implementation Strategy

To mitigate recruitment risks, the institute should pivot toward a fellowship model, bringing in top academic talent for 12-to-24-month stints. This avoids the long-term salary cap issue while ensuring a steady influx of fresh expertise. To address compute constraints, the implementation plan must prioritize partnerships with the National Science Foundation to utilize existing academic supercomputing clusters. Execution success will be measured by the adoption rate of NIST-developed safety benchmarks by non-U.S. labs, establishing a global de facto standard.

Executive Review and BLUF

1. BLUF

The U.S. AI Safety Institute must prioritize technical benchmarking over policy advocacy to establish legitimacy. With a budget of only 10 million dollars, it cannot outspend the industry. It must instead out-think it by becoming the primary source of safety measurement tools. Success requires bypassing federal hiring constraints through academic fellowships and securing compute access via public-private partnerships. The institute has an 18-month window to become indispensable before its voluntary cooperation model is challenged by political shifts or industry fatigue. Failure to deliver high-quality technical standards by fiscal year 2025 will result in the agency being relegated to a minor advisory role.

2. Dangerous Assumption

The most consequential unchallenged premise is that frontier AI labs will continue to provide meaningful, pre-deployment access to their models voluntarily. As competitive pressures increase, the incentive for labs to withhold data to protect trade secrets will likely outweigh their desire for a safety seal of approval from a non-regulatory body.

3. Unaddressed Risks

• Regulatory Capture (High Probability, High Consequence): The institute relies on industry experts for staffing and industry models for testing. This creates a feedback loop where standards are tailored to what the current leaders can already achieve, preventing new entrants and softening safety requirements.
• International Divergence (Medium Probability, Medium Consequence): If the UK or EU institutes develop more rigorous or different testing protocols, the U.S. institute risks irrelevance as global labs gravitate toward the most stringent or most widely accepted standard.

4. Unconsidered Alternative

The analysis overlooks the option of a purely Defensive Focus. Instead of general safety standards, the institute could focus exclusively on the safety of AI applications within the federal government itself. This would provide a controlled environment for testing, a guaranteed user base, and a clear legal mandate under existing procurement rules, avoiding the need for voluntary industry cooperation.

5. MECE Verdict

APPROVED FOR LEADERSHIP REVIEW. The analysis correctly identifies the tension between technical goals and resource limits. The recommendation to focus on technical standards is the only path that builds long-term institutional value.