Fable is free. But the technology desperately needs better regulations
Illustration: Ben Hickey
New rules often come out of disasters. America’s Federal Reserve was founded following the Panic of 1907, when stock prices fell by half. Pressure from muckrakers such as Upton Sinclair brought about the Food and Drug Administration. The Securities and Exchange Commission was founded during the Great Depression. Artificial intelligence has not yet caused a calamity, but it might. Models like Anthropic’s Mythos and Open AI ‘s GPT -5.6 Sol are extraordinarily good hackers and could become capable advisers to bioterrorists.
Understandably, the Trump administration is trying to regulate the technology before catastrophe strikes. It is working with AI companies on voluntary standards that could soon be released. Unfortunately, its efforts so far have been a mess.
Anthropic was its first victim, slapped with export controls in mid-June after the release of Fable, a guardrailed version of Mythos. That came not long after a row between the company and the Pentagon. Fears of a grudge were allayed when the administration appeared to compel Open AI to limit access to its Sol model. Then, on June 30th, the Commerce Department abruptly lifted the ban on Fable after Anthropic fiddled with its safety protections.
Throughout, the government has seemed to make up rules on the fly. Its decisions have also had an unpleasant nationalistic tinge. At first it restricted Fable only for non-Americans; Anthropic decided that a wholesale block was the only way to comply. It looks as if the administration pulled the only lever available, knowing that it was a de facto ban. But throughout the past month’s brouhaha, the government has made clear that Americans’ AI access takes precedence over foreigners’.
Now that America has started licensing AI releases, it is unlikely to stop. But regulating frontier AI is tricky. China’s top models are only months behind America’s, and most are open-weight, meaning anyone can run or tinker with them. One recent release, GLM 5.2 from Z.ai, already matches the best of the last generation of American models. Chinese labs may take longer to catch up with Mythos, since they have fewer chips and American labs are cracking down on distillation, when competitors use the outputs of the best models to train their own. But that buys months, or a year at most. America could ban Chinese open-weight models and punish foreigners who use them. But even if it managed to enforce its ban, a vast home market would keep China’s model-makers going.
So a permanent block is unworkable. It is also undesirable. Many American AI firms and researchers rely on Chinese models, which are cheap and malleable. The intelligence that makes new models dangerous also makes them tremendously useful. Worries about China aside, a gulf between publicly available and restricted models is a problem. Societies adapt to AI best when improvements arrive gradually, not in a great lurch. Imagine the mess if regulators bottled up several generations of Mythos-style advances. The few with access would acquire great power. The sudden jump in capability whenever the models did get released would unleash chaos.
How then can models like Mythos and Sol be safely set free? The emerging norm provides for an evaluation period and a staggered release to trusted institutions. That is a good start, but it needs formalising. Some choices, such as how much risk to tolerate, belong to elected leaders. But politicians should not be micromanaging the process or horse-trading with AI companies, as they do today.
Once those goals are set, politicians should stand back. Evaluating a model is a technical problem. Governments have some expertise, for example in America’s Centre for AI Standards and Innovation or Britain’s AI Security Institute (AISI). But the private sector has more. The finance and electric-power industries offer structures where oversight is carried out by industry bodies, overseen by government. Something similar might help amalgamate knowledge from the AI labs, research groups and foreign bodies such as AISI.
Ideally, America would work with its allies on AI regulation. Alas, it is hard to imagine Donald Trump giving up the immense sovereign power that stems from control of frontier models. Locking others out of AI is not in America’s economic or strategic interest, but neither were indiscriminate tariffs or threats to Greenland.
So other countries must build leverage with their own AI sectors and regulations, and find fail-safes for American export controls. They could, for instance, ensure that businesses can easily switch to non-American models that run on non-American data centres. It is far wiser to depend on America than on China, but they would be mad to ignore the risks. ■
Subscribers to The Economist can sign up to our Opinion newsletter, which brings together the best of our leaders, columns, guest essays and reader correspondence.
논증 분석
유형: prescription
핵심 주장
Donald Trump 행정부의 최전선 AI 모델 규제—Anthropic의 Fable 수출 금지로 대표되는—는 즉흥적이고 민족주의적이며 실효성이 없다. China의 오픈웨이트 모델이 불과 수개월 뒤처진 상황에서, The Economist는 금융·에너지 산업을 모델로 한 공식화된 평가 프레임워크와 단계적 공개 체계를 처방한다.
논리구조
- 역사적 유추—재난이 규제를 낳는다: Federal Reserve는 1907년 공황 이후, FDA는 업튼 싱클레어의 폭로 이후, SEC는 대공황 이후 탄생했다. AI는 아직 재난을 일으키지 않았지만 잠재력이 있으므로 선제적 규제가 필요하다.
- 현행 규제의 실패 진단: Anthropic의 Fable은 6월 중순 수출 통제를 받았다가 6월 30일 갑자기 해제됐다. 정부는 규칙을 즉흥적으로 만들었고, 처음에는 비미국인에게만 제한을 가했는데 Anthropic은 전면 차단만이 유일한 준수 방법임을 알게 됐다. 전 과정이 불투명하고 민족주의적 편향을 드러냈다.
- 영구 차단의 실효성 부재: China 최고 모델들은 대부분 오픈웨이트 방식이어서 누구나 실행·수정 가능하다. Z.ai의 GLM 5.2는 이미 미국 이전 세대 모델과 동등하다. 증류 제한이 격차를 늦추더라도 몇 달, 최대 1년에 불과하다.
- 봉쇄의 부작용: 공개 모델과 제한 모델 간의 격차는 그 자체로 문제다. AI 개선이 점진적으로 도입될 때 사회가 적응하는데, 여러 세대 분의 능력 향상이 한꺼번에 풀릴 경우 혼란을 야기하고 접근권을 가진 소수에게 과도한 권력을 집중시킨다.
- 처방—공식화된 평가 체계: 신뢰할 수 있는 기관에 대한 평가 기간 및 단계적 공개라는 신흥 규범을 공식화해야 한다. 위험 허용 수준 같은 가치 판단은 선출직 지도자에게, 기술적 평가는 AI Security Institute 같은 전문 기관과 민간 산업 기구에 맡기는 금융·전력 산업 방식의 구조가 필요하다.
- 동맹 협력의 이상과 현실: 이상적으로는 America이 동맹국들과 협력해 AI를 규제해야 하지만, Donald Trump가 최전선 모델 통제에서 오는 막대한 주권적 권력을 포기할 가능성은 없다.
- 다른 국가들에 대한 처방: 미국의 수출 통제에 대비해 비미국 모델로 쉽게 전환할 수 있는 체계를 구축하고, 비미국 데이터 센터를 확보해야 한다. 미국에 의존하는 것이 China에 의존하는 것보다 현명하지만, 리스크를 무시해서는 안 된다.
결론
America은 최전선 AI를 ‘감금’하는 방식—불투명한 수출 통제와 즉흥적 제한—을 포기하고, 금융·에너지 산업의 선례를 따른 공식 평가·단계적 공개 체계를 구축해야 한다. 그렇지 않으면 실효도 없는 규제가 동맹국을 소외시키고 미국 기업에도 해를 끼치는 최악의 결과만 낳을 것이다.