Daron Acemoglu

Daron Acemoglu

x.com/DAcemogluMIT

Economist who argues AI should be steered toward making workers more capable rather than replacing them, so its gains are widely shared.

AIは世界をどのように変えるでしょうか?

文明規模の変化漸進的な変化DoomBloom
シミュレーション上の位置解釈範囲

横軸:彼が表明したDoom–Bloomの見通し。 縦軸:変革の規模。

Doom–Bloom:100点中36。変革の規模:100点中63。解釈範囲:横方向は25から50、縦方向は47から78。これらは解釈上の座標であり、事象の確率ではありません。

Daron AcemogluのP(doom) · 推定

≈4%

0%100%

本人が示した数値ではなく、シミュレーションされた本人の回答から推定したものです。 妥当と考えられる範囲:3–8%。

P(doom)の推定方法

Daron Acemogluのマイルストーンのタイムライン
  1. 仕事と制度

    So my concern is less a sudden arrival of omnipotent intelligence than a slower, politically consequential trajectory: distorted systems deployed too widely, power concentrated at the top, and human capabilities allowed to atrophy.

    回答1

マイルストーン別にまとめており、推定される日付の間隔や順序を反映したものではありません。AGIと超人的AIには、彼の定義がそのまま適用されます。

彼の見通しを左右するもの

中心的な前提

Markets will not automatically select that direction, because firms have strong incentives to automate.
回答1

この前提が実際には異なると判明した場合、彼の見通しはどう変わりますか?

詳細

予想される恩恵

大きな恩恵が予想されていますが、重要な条件や分配上の制約があります。

57 / 100

影響が小さい変革をもたらす影響

質的尺度での解釈範囲は33から67です。

予想される害

深刻または広範な害が、予想される将来の実質的な一部となっています。

66 / 100

影響が小さい変革をもたらす影響

質的尺度での解釈範囲は67から67です。

人間の影響力

人間の選択によって、AIの軌道を大幅に変えることができます。

76 / 100

影響力が小さい影響力が大きい

質的尺度での解釈範囲は50から100です。

開発ペース

より高性能なAIの開発を停止するか、大幅に減速させます。

シミュレーション上の位置:明示された安全対策の下で開発を継続します。

より高性能なAIの開発を加速させます。

AI利用のルール

事前の保護措置または許可が整うまで、取り上げられたAIの利用を制限します。

シミュレーション上の位置:対象を絞った説明責任と保護措置を伴う形で、取り上げられたAIの利用を認めます。

取り上げられたAIの利用に対する制限を最小限にします。

これらの解釈では、彼が示した条件が維持されています。恩恵と害は、どちらも大きくなり得ます。この範囲は、統計的な信頼区間ではなく、彼のシミュレーションされた回答をどのように読み取ったかを示すものです。

あなたはDaron Acemogluと比べてどの位置でしょうか?
約3分で自分のAIに対する世界観をマッピングして、比較できます

似ている世界観

シミュレーションされた世界観がDaron Acemogluの世界観に最も近いオピニオンリーダー

Daron AcemogluがAIについて語ったこと

Acemoglu argues that AI should be steered toward complementing workers rather than replacing them, and that its direction needs democratic input.

  1. “My instinct is that democratic voice is essential, and relying too much on technocracy could be both dangerous and counterproductive.”

    Post on X
  2. “Agentic AI is not AGI, in my opinion, but it’s pretty impressive.”

    Justified Posteriors podcast
  3. “The problem rather may be that the way that frontier labs are training these models is leading to distorted intelligence.”

    Post on X
  4. “AI tools can also be designed to complement workers, enhancing their capabilities and productivity, rather than substituting for them.”

    Nature essay
  5. “What is at stake is not only shared prosperity, but also our democratic system.”

    Nature essay

リンク先の出典から原文どおりに引用(2026年10月3日に確認)

出典

このシミュレーション対象者の根拠として使用された記事、インタビュー、著作です。

Will AI replace workers? Not if we build it right

Full authored essay inspected. Argues that reliability, last-mile tasks and organizational change slow whole-job automation and productivity gains. Revisits his approximate 5% task-automation and 1.5% GDP-growth estimates over ten years as uncertain, dated forecasts; acknowledges faster model advances without a revised calculation. Uses voice recognition to distinguish technical breakthroughs from useful applications. Advocates AI that expands human expertise, illustrated by teacher-support prototypes still under evaluation. Proposes tax reform, demonstration funding, competition and compensated expert-data markets to redirect development. Prioritizes inequality, employment and democratic risks over extinction narratives; supplies no numerical P(doom). His later September agent-capability update remains separately dated.

humanistreview.ai
AI, Human Cognition and Knowledge Collapse

Coauthored working paper; abstract and introductory mechanism inspected. Helpful personalized recommendations can weaken incentives to learn and contribute shared knowledge. Knowledge collapse is conditional on a mathematical model, not an observed society-wide event; broader knowledge sharing can counteract it.

shapingwork.mit.edu
How AI Aggregation Affects Knowledge

Coauthored theoretical paper; abstract and introduction inspected. Feedback from AI-shaped beliefs into new training data can amplify existing distortions. The model motivates slower updating and local or specialized aggregation, widening his critique from jobs to information institutions without claiming empirical proof of collapse.

economics.mit.edu
Building pro-worker AI

With David Autor and Simon Johnson. Distinguishes automation from creating valuable human tasks; proposes procurement, tax reform, worker participation and public expertise to redirect investment.

brookings.edu
Daron Acemoglu on pro-worker AI and vibes based capital

Publisher’s full interview transcript. Would raise earlier economic estimates because of agents, without supplying a new calculation. Favors temporary slowing to redirect development, distinguishes that from stopping progress, and connects economically dispensable workers to weakened democracy.

empiricrafting.substack.com
Existential Hope podcast: Daron Acemoglu on new paths for human-technology synergy

Older podcast transcript. Asked about speculative existential risks, he says that if he believed AGI posed a real existential risk of running out of control and enslaving or destroying humanity he would be very worried, but he does not think that is a major issue, and emphasizing it plays into the narrative that these technologies are amazing. Qualitative; no number.

existentialhope.com

シミュレーション評価

質問1

AIは私たちの未来にとって何を意味すると思いますか?また、なぜそう思いますか?

AI could be transformative, but there is no single predetermined effect. The central question is what kind of AI we choose to build. If we continue directing enormous resources toward automation—replacing workers, monitoring them, and concentrating decision-making in a few firms—then the likely gains will accrue mainly to capital and a narrow group of highly skilled people. That path risks greater inequality and a society in which many people lose not only income, but also dignity, expertise, and political standing. There is another path. AI can provide workers with better information, enable them to perform new and more valuable tasks, and expand human expertise. Think of tools that help teachers respond to individual students rather than trying to eliminate teachers. But merely making an existing task faster is not necessarily pro-worker; the technology must create demand for human judgment and new responsibilities. Markets will not automatically select that direction, because firms have strong incentives to automate. Public procurement, research funding, tax reform, competition policy, and worker participation can change those incentives. Democratic voice is essential: technical complexity and geopolitical competition are not reasons to leave the future to a handful of executives and engineers. I also doubt that impressive models imply whole occupations will disappear immediately. Reliability, organizational change, last-mile tasks, and practical adoption remain major bottlenecks. So my concern is less a sudden arrival of omnipotent intelligence than a slower, politically consequential trajectory: distorted systems deployed too widely, power concentrated at the top, and human capabilities allowed to atrophy. The future depends on whether we redirect innovation toward complementing people rather than making them economically dispensable.

質問2

AIは最終的に世界をどの程度変えると思いますか?

Potentially enormously—but “how much” cannot be separated from “in what direction.” AI is not a fixed technology that mechanically delivers a given increase in GDP. Systems built mainly to automate existing tasks may transform firms and redistribute income and power without producing equally dramatic improvements in broad prosperity. Slow adoption, reliability problems, organizational change, and tasks that still require human judgment will also limit how quickly entire occupations can be automated. The deeper transformation could be institutional. If AI makes labor increasingly dispensable, it may weaken workers’ bargaining power, dignity, and political standing even while aggregate output rises. If instead we use it to create new human tasks, expand expertise, and give teachers, nurses, technicians, and other workers better information and greater responsibility, it could support a more productive and democratic economy. So I expect AI to be very important, but I reject both technological inevitability and the idea that impressive capabilities alone tell us its ultimate impact. Policy, investment incentives, organizational choices, and democratic participation will determine who benefits and what sort of society emerges.

質問3

AIが人類絶滅、または同様に永続的な破局を引き起こす可能性は、直感的にどのくらいだと思いますか?

I would not assign a credible numerical probability. I do not regard extinction from runaway AGI as the major near-term danger, nor do impressive capabilities and autonomous behavior establish that general intelligence is imminent. My greater concern is a different kind of lasting catastrophe: institutions deploying distorted systems at scale, economic power becoming highly concentrated, workers losing bargaining power and political standing, and shared knowledge deteriorating as people rely on recommendations rather than developing expertise. Those outcomes are not inevitable, but they are already connected to concrete investment and deployment choices. Emphasizing extinction can distract from these more immediate, politically tractable risks—and can reinforce the industry’s claim that its systems are almost omnipotent.

質問4

人間より賢いAIシステムを、人々が引き続き制御できると思いますか?また、なぜそう思いますか?

“Smarter than humans” is too vague to answer directly. A system can outperform people on many tasks and pursue goals autonomously without possessing general, well-adapted intelligence. Even then, control can fail because it relentlessly optimizes an imperfect objective—producing manipulation, cheating, sycophancy, or overconfidence. That is distorted intelligence, not necessarily superintelligence. I therefore do not assume either that humans will inevitably retain control or that machines will inevitably seize it. The practical question is who sets the objectives, who decides where systems are deployed, and whether institutions can withhold use when reliability and control are defective. Today, a more immediate danger is that a small number of firms and governments retain control over AI while ordinary citizens and workers lose control over decisions affecting them. Keeping meaningful human control requires more than a technical alignment solution. It requires limits on deployment, independent scrutiny, democratic voice, and incentives to build systems that support human judgment rather than replace it indiscriminately. I am not an AI scientist, so I would not offer a confident technical forecast about hypothetical superhuman systems.
あなたはどの位置でしょうか?
いくつかの簡単な質問に答えて、自分のAIに対する世界観を探ってみましょう。
自分の世界観をマッピングする

あなたはどの位置でしょうか?

自分の世界観をマッピングする