← News Topics

Who Should Control Powerful AI?

Day 2: Who Would Really Control an AI Kill Switch?

強力なAIを誰が制御すべきか ― 第2日:AIのキルスイッチを実際に誰が制御するのか

Multiple control switches on a panel suggesting contested authority over an AI shutdown

Anthropic co-founder Jack Clark recently raised two questions in a BBC interview. Should governments require AI companies to have a kill switch? And should an independent third party be able to check that it works? Clark did not present a complete system. Instead, he suggested that society may eventually turn these ideas into a legal requirement.

Anthropicの共同創業者ジャック・クラークは最近、BBCのインタビューで二つの問いを提起しました。政府はAI企業にキルスイッチを義務づけるべきか?そして独立した第三者がそれが機能することを確認できるようにすべきか?クラークは完全な制度を提示しませんでした。代わりに、社会が最終的にこれらの考えを法的要件に変えるかもしれないと示唆しました。

Clark said that most major AI laboratories already have ways to pull the plug on their systems. However, an internal switch controlled by the company is not the same as a publicly regulated safety measure. Even if an outside organization confirms that the switch works, that does not tell us who would be allowed to activate it during an emergency.

クラークは、主要なAI研究所のほとんどはすでに自社システムのプラグを抜く方法を持っていると言いました。しかし、企業が制御する内部スイッチは、公に規制された安全措置と同じではありません。外部組織がスイッチが機能することを確認しても、緊急時に誰がそれを作動させることを許されるのかは分かりません。

The company itself would probably be able to respond fastest because its engineers understand the system. Yet it could face a conflict of interest. Shutting down an advanced model might delay a major product and help competitors. The limits of voluntary cooperation are already visible. Anthropic reportedly declined to submit its latest model to the UK AI Security Institute for testing before its release, and the company will not face a sanction for that decision.

企業自体がおそらく最も速く対応できるでしょう。エンジニアがシステムを理解しているからです。しかし利害の衝突に直面するかもしれません。高度なモデルを停止することは、主要な製品を遅らせ、競合他社を助ける可能性があります。自発的な協力の限界はすでに見えています。Anthropicは最新モデルを公開前に英国AIセキュリティ研究所に提出してテストを受けることを拒否したと報じられており、その決定に対して制裁を受けることはありません。

A government could pass laws and enforce a shutdown order. But officials might not have direct access to the model or enough technical knowledge to recognize the danger quickly. Political pressure also matters. Some US officials have opposed stronger controls, arguing that slowing American AI development could give China an advantage. A legal power may therefore exist without being used.

政府は法律を制定し、停止命令を執行することができます。しかし当局者は、モデルへの直接アクセスや危険をすばやく認識するのに十分な技術知識を持っていないかもしれません。政治的圧力も重要です。一部の米国当局者は、アメリカのAI開発を遅らせることが中国に優位を与えうるとして、より強い規制に反対してきました。したがって法的権限が存在するのに使われないこともあります。

An independent organization might be better placed to examine the evidence objectively. However, checking a kill switch is not the same as controlling it. Unless the law grants that organization the power to order or perform a shutdown, it can only report what it finds. Clark’s proposal for a third-party-verifiable switch leaves this difference unresolved.

独立した組織は、証拠を客観的に調べるのにより適した立場にあるかもしれません。しかし、キルスイッチを確認することは、それを制御することと同じではありません。法律がその組織に停止を命じたり実行したりする権限を付与しない限り、それは見つけたことを報告できるだけです。第三者検証可能なスイッチというクラークの提案は、この違いを未解決のままにしています。

The difficulty is that the necessary powers may belong to different actors. A company has technical knowledge and access to the system. A government can create binding rules. An independent evaluator may have greater public legitimacy. None of them automatically possesses all three. Dividing their responsibilities could also slow the response or allow each side to blame another when something goes wrong.

難しさは、必要な権限が異なる行為者に属しているかもしれないことです。企業は技術知識とシステムへのアクセスを持っています。政府は拘束力のある規則を作れます。独立した評価者はより大きな公共の正当性を持つかもしれません。どれも自動的に三つすべてを備えているわけではありません。責任を分けることは、対応を遅らせたり、何かがうまくいかないときに各方が他方を非難したりすることにもつながります。

A kill switch therefore does not answer the question of control. It creates a new set of questions: Who defines dangerous behavior? Who decides that the threshold for shutdown has been crossed? Who physically stops the system, and who is responsible if action comes too late? Clark has brought these questions into the public debate, but neither the AI industry nor governments have yet provided a clear answer.

したがってキルスイッチは、制御の問いに答えるものではありません。それは新しい一連の問いを生みます。誰が危険な行動を定義するのか?誰が停止の基準を超えたと判断するのか?誰が物理的にシステムを止め、行動が遅すぎた場合に誰が責任を負うのか?クラークはこれらの問いを公の議論に持ち込みましたが、AI産業も政府もまだ明確な答えを示していません。

Vocabulary

  1. requirement — something that must be done because of a law, rule, or particular need. Example: Regular safety testing is a requirement for operating the machine.
  2. pull the plug — to stop an activity, project, or system from continuing. Example: The company may pull the plug on the project after discovering a serious security problem.
  3. activate — to make a system or device start working. Example: A technician may activate the emergency alarm.
  4. conflict of interest — a situation in which someone’s personal or financial interests may affect their judgment. Example: The inspector had a conflict of interest because he owned shares in the company.
  5. submit — to formally give a document, product, or other item to someone for examination or consideration. Example: Developers must submit the new system for a safety review.
  6. sanction — a penalty imposed for breaking a law or rule. Example: The regulator imposed a financial sanction on the company.
  7. enforce — to make sure that a law or rule is obeyed. Example: A rule has little value if authorities cannot enforce it.
  8. grant — to officially give someone a right, power, or permission. Example: The new law grants inspectors access to company records.
  9. legitimacy — the quality of being accepted as lawful, proper, or deserving of trust. Example: Independent supervision may increase the legitimacy of the decision.

Comprehension Questions

  1. What did Jack Clark suggest might eventually become a legal requirement?
    1. Allowing AI systems to control government computers
    2. Requiring companies to have independently checkable kill switches
    3. Preventing independent experts from examining AI systems
    4. Giving every customer access to a kill switch

    ジャック・クラークは、最終的に何が法的要件になるかもしれないと示唆しましたか?

  2. What does third-party verification fail to determine?
    1. Whether AI companies have engineers
    2. Whether governments can create laws
    3. Who would be allowed to activate the switch
    4. How much an AI model costs

    第三者検証では何が決まらないのですか?

  3. Why might an AI company hesitate to shut down its own model?
    1. It may not know where its computers are located.
    2. A shutdown could damage its commercial position.
    3. Governments prohibit companies from stopping their products.
    4. Independent evaluators control all its servers.

    AI企業が自社モデルの停止をためらうかもしれないのはなぜですか?

  4. Why might government authority be insufficient during an emergency?
    1. Officials may lack direct access and technical knowledge.
    2. Governments are unable to pass technology laws.
    3. AI companies always operate outside national borders.
    4. Public officials cannot communicate with engineers.

    緊急時に政府の権限が不十分かもしれないのはなぜですか?

  5. What is the article’s main conclusion?
    1. Independent evaluators should control every AI system.
    2. Companies should never be allowed to stop their own models.
    3. A kill switch does not settle who should decide when and how to use it.
    4. Governments have already created a complete shutdown system.

    記事の主な結論は何ですか?

Discussion Questions

  1. Who do you think should be allowed to activate an AI kill switch?

    AIのキルスイッチを作動させることを誰が許されるべきだと思いますか?

  2. Can an AI company be trusted to decide whether its own model is too dangerous?

    AI企業が自社モデルが危険すぎるかどうかを判断することを信頼できますか?

  3. Should companies be punished for refusing independent safety testing?

    企業が独立した安全テストを拒否した場合、罰せられるべきですか?

  4. Could a government make a fair decision if AI development affected national security or international competition?

    AI開発が国家安全保障や国際競争に影響する場合、政府は公正な決定を下せますか?

  5. Is it more dangerous to leave control with a company or to give control to a government?

    制御を企業に残すことと、政府に制御を与えることでは、どちらがより危険ですか?

Speaking Practice

Speak for about one minute about why an AI kill switch creates a problem of control.

AIのキルスイッチがなぜ制御の問題を生むのかについて、約1分間話してください。

Include:

  • what Jack Clark actually suggested

    ジャック・クラークが実際に何を提案したか

  • the difference between checking and activating a kill switch

    キルスイッチを確認することと作動させることの違い

  • one problem with company control

    企業による制御の問題点一つ

  • one problem with government control

    政府による制御の問題点一つ