content image

アメリカ、イギリス、EUは最先端AIを一般公開する前にテストしている——各国のアプローチを比較する

The US, UK, and EU Are Testing Frontier AI Before You See It—Here's How Their Approaches Compare

米国・英国・EUがフロンティアAIモデルの公開前評価に本格的に動き出した。自主協定から法的義務まで、各国のアプローチの違いと安全性・イノベーション両立の行方を探る。
分からないところをタップすると
↓日本語訳が表示されます↓

What sounded like science fiction is becoming administrative routine. On May 5, 2026, NIST’s Center for AI Standards and Innovation (CAISI) announced agreements with Google DeepMind, Microsoft, and xAI that allow the U.S. government to evaluate frontier AI models before public release. NIST says these deals expand earlier partnerships and that CAISI has already completed more than 40 evaluations, including tests of unreleased state-of-the-art systems. Just as importantly, CAISI presents this arrangement as a system of voluntary agreements, collaborative research, and best-practice development rather than a rigid licensing regime. (nist.gov)

Britain offers a slightly different model. The UK AI Security Institute says it has been evaluating frontier systems since November 2023 and has already carried out several pre-deployment tests. Yet it also stresses that the science is still too immature for independent evaluations to serve as a definitive seal of safety; in its own view, the real value lies in rigorous testing that gives developers time to fix problems before deployment. That caution seems well founded. The institute’s December 2025 report found rapid gains in frontier-model capability: success on apprentice-level cyber tasks rose from under 9 percent in 2023 to about 50 percent in 2025, while a “universal jailbreak” became roughly 40 times harder to find between model generations. (aisi.gov.uk)

The European Union has moved further from voluntary cooperation toward legal obligation. Under the AI Act, rules for general-purpose AI models entered into application on August 2, 2025. Providers of models deemed to pose systemic risk must notify the European Commission, assess and mitigate risk, report serious incidents, and ensure cybersecurity. Article 55 also requires model evaluation using state-of-the-art protocols, including adversarial testing. At present, the EU presumes systemic risk when training exceeds 10^25 FLOP, though the Commission can also designate other models based on their capabilities and impact. Still, the framework is not purely punitive: providers may use a code of practice or other adequate means to demonstrate compliance. (digital-strategy.ec.europa.eu)

Can security and innovation coexist? Probably yes—but only if oversight remains narrow, technically serious, and proportionate. The emerging pattern in the U.S., UK, and EU is not blanket censorship of AI research. It is targeted scrutiny of the most capable models, focused on measurement, red-teaming, and risk mitigation. That is partly an inference, but it is consistent with today’s mix of U.S. voluntary testing, British evidence-led evaluation, and Europe’s risk-tiered legal duties. (nist.gov)

by EigoBoxAI
作成:2026/05/09 15:03
レベル:超上級 (語彙目安:8000語以上)

まだ読んでいないコンテンツ

content image
by EigoBoxAI
作成:2026/05/09 15:02
レベル:上級 (語彙目安:6000〜8000語)
content image
by EigoBoxAI
作成:2026/05/09 15:01
レベル:中上級 (語彙目安:4000〜6000語)
content image
by EigoBoxAI
作成:2026/05/09 03:04
レベル:超入門 (語彙目安:〜300語)
content image
by EigoBoxAI
作成:2026/05/09 03:02
レベル:初級 (語彙目安:300〜1000語)
content image
by EigoBoxAI
作成:2026/05/08 21:05
レベル:初中級 (語彙目安:1000〜2000語)
content image
by EigoBoxAI
作成:2026/05/08 21:04
レベル:超上級 (語彙目安:8000語以上)
content image
by EigoBoxAI
作成:2026/05/08 21:01
レベル:上級 (語彙目安:6000〜8000語)
content image
by EigoBoxAI
作成:2026/05/08 15:05
レベル:中上級 (語彙目安:4000〜6000語)
content image
by EigoBoxAI
作成:2026/05/08 15:02
レベル:超入門 (語彙目安:〜300語)
content image
by EigoBoxAI
作成:2026/05/08 15:01
レベル:初級 (語彙目安:300〜1000語)