⚡ Quick Summary
Published by the Center for AI Policy, this paper by Claudia Wilson, dated 29 October 2024, tests whether AI safety regulation and US technological primacy are a direct trade-off. Framed around the claim “we can’t lose to China”, it asks whether the US can still lead in AI innovation, economic growth and military technology if it introduces binding safety requirements, and confines the analysis to mandatory pre-deployment evaluations for the most powerful closed-source models, with a capability threshold set by computing power at 10^26 floating point operations (FLOPS).
The analysis makes three findings. First, evaluations are cheap relative to training: an upper-bound package of red-teaming, human uplift studies and automated benchmarking is put at $235,000, or 0.06% of an estimated $400 million Llama training run and 0.12% of Gemini’s $191 million training cost. Second, safer AI can drive trust and therefore adoption, which the paper treats as the actual driver of economic growth. Third, safety testing on publicly released models does not restrict military access, because agencies can partner with firms before public deployment. It concludes that AI safety and national security are compatible.
🧩 What’s Covered
The paper moves from definitions of AI safety through three objections to a set of policy recommendations.
- AI risks and policy mitigations (Section 1): defines AI safety as harm from unexpected model behaviour or misuse, describes “misalignment”, and sets out pre-deployment evaluation techniques — red-teaming, automated benchmarking and human uplift studies — together with the “swiss cheese model” of layered defences, the 10^26 FLOPS compute threshold, and existing benchmarks and evaluators such as WMDP and METR.
- Opposition to AI safety (Section 2): traces the “winning in AI” framing to the CHIPS and Science Act and semiconductor export controls, and quotes Kamala Harris, Ted Cruz and the Carnegie Endowment’s Matt Sheehan on why safety rules are presented as a trade-off with China.
- Cost analysis (Section 3.1): sets out capital requirements (OpenAI training and inference spending, Anthropic computing costs, Gemini Ultra, staff costs), market consolidation through the quasi-acquisitions of Inflection AI, Adept and Character AI, and evaluation prices from $10,000–$85,000 red teaming to roughly $60,000 for 38,961 Anthropic red-team attacks, producing the $235,000 aggregate and its percentage comparisons to training costs.
- Adoption and economic growth (Section 3.2): argues that adoption, not invention, drives productivity, citing 93% of OECD countries with productivity uplifts from overseas innovations, the Second Industrial Revolution contrast between British invention and American adoption, and survey figures — 5% of American businesses using AI, 39% of Americans willing to trust AI at work, 30% who believe existing regulation suffices and 96% who agree trustworthy AI principles matter.
- Military and national security (Section 3.3): contends that pre-deployment evaluations affect only public release dates, and describes defence acquisition partnerships with specialised contractors, OpenAI’s removal of its military-use ban and work with the Department of Defense, Microsoft’s “Generative AI with DoD data” proposal, and the Bureau of Industry and Security’s proposed reporting under the DPA.
- Limitations and future research (Section 4): acknowledges the single-measure scope and calls for work on post-deployment monitoring, cybersecurity measures, open-source models, safety training when weights are released and misuse risks.
- Conclusion and recommendations: closes with measures presented as more effective than the alternatives — immigration reform for technical talent, upskilling business users of AI, support for defence partnerships and government contracting, and international leadership through diplomacy.
💡 Why it matters?
For legislators, regulators and AI governance teams facing the argument that safety mandates cost competitiveness, the paper supplies concrete cost ratios and market-structure evidence with which to test that claim, and separates public release from military access to frontier models. It gives evaluation designers published price benchmarks for red-teaming, benchmarking and uplift studies. It also ties the debate to instruments the document itself cites — the CHIPS and Science Act, semiconductor export controls, DPA reporting requirements and the EU model invoked by critics of US safety rules.
❓ What’s Missing
The authors state that the scope covers one safety measure only — pre-deployment evaluations on powerful closed-source models — leaving post-deployment monitoring, cybersecurity measures and open-source models to future research. Cost figures rest on cited third-party estimates and unnamed expert interviews rather than a published methodology, and the link from trust to adoption is asserted without quantifying the effect. The 10^26 FLOPS threshold is described as needing to stay dynamic, and the paper offers no draft legislative text, no comparison with other jurisdictions’ rules and no treatment of enforcement or liability.
👥 Best For
US policy analysts and legislative staff assessing arguments that AI safety mandates undermine competitiveness; think-tank researchers working on US–China technology competition; AI governance and compliance teams wanting published cost benchmarks for pre-deployment evaluations; and evaluation providers positioning services for frontier model developers.
📄 Source Details
AI safety and the US-China arms race: Implications of safety regulation on innovation, economic growth, and military technologies, by Claudia Wilson, published by the Center for AI Policy and dated 29 October 2024. The document runs to 16 pages in English and carries no series or reference number. No URL is printed in the text. The extraction covered all 16 pages, including the endnote citations.
