8:57Daron Acemoglu： 都有。你举的每个例子里，推动力也来自内部，但最后都得建立新的监管机构。而且有时候人们围绕一个从大局看其实是次要的议题凝聚起来，后来又汇入了别的东西。比如美国历史上的人民党，最初的民粹运动，诉求主要是农村、农业和经济上的，几十年后并入了进步运动。在我看来，数据中心不是主战场。但如果这是能让人动起来的事，也许它提供了组织起来的机会。这就是领导力该起作用的地方：眼睛要盯着大目标。大目标不是阻止 AI，那是浪费，它是非常有前途的技术；也不是在社会保险上做点小修小补，让人们因 AI 失业后不至于挨饿。大目标是给 AI 找到一个对社会更有益的方向。
10:25 · 为什么不该拿工业革命作比
10:23Sinead Bovell： 你在 X 上发过一条帖子，回应几位经济学家签的一份请愿，说你不喜欢把 AI 对经济的影响比作工业革命，觉得那是拿苹果比橘子。考虑到工业革命对亲历者有多糟糕，为什么这个比较不值得做？你对劳动力市场的预测又是什么？
11:09Daron Acemoglu： 具体原因是，科技圈的人拿英国工业革命来比时，有意无意地忘掉了它的头 80 年，指的其实是工业革命后来带来的种种好处。我和 Simon Johnson 写《权力与进步》，就是为了强调这一点，但它还是被忽略。所以我担心，拿工业革命作比往往是一句暗语，意思是「这一切都会很美好」。
12:02Daron Acemoglu： 第二个原因是，尽管经济史学家记录了种种苦难，工业革命本身非常慢，也非常局部。定义它头 80 年的只有两个部门：纺织和煤。经济的其他部分变化不大，那些苦难就是在这个背景下发生的。现在想想 AI 演进的速度。我第一个承认，我们还没看到大规模的岗位替代；会不会出现、以什么方式出现，还非常不确定。但想象几个大部门同时开始裁员，那会比工业革命惨烈得多。所以这个比较在这一点上也很牵强。当然，我们依然能从历史里学到很多。
16:52Daron Acemoglu： 顺便说，我是互联网的忠实拥趸，虽然我们误用了其中一部分，社交媒体和网上交流一直很糟。但总体上，我是早期用户，当时每个人多少都明白互联网会做什么。互联网泡沫破裂时倒闭的那些公司，商业模式也都是对的。大家都笑 pets.com，可它的模式是对的：用互联网这个新平台，更便宜地提供更多服务和更多样的商品，给顾客更多选择、更好的信息，去和实体店竞争。这个模式延续到了今天。再看 AI：我们仍然完全不知道 AI 经济会长什么样。这种不确定让过渡变得非常复杂。
18:03Daron Acemoglu： 我很想告诉你 20 年后的 AI 经济是什么样，但我完全不知道。我猜里面会有工人，但不知道他们在做什么。做社会性的工作？全都是照看 AI 模型的工程师和技术员？模型还会不会参差不齐，以至于需要人类工人大量手把手地扶着？工人能不能找到新的任务，为创新和生产率作贡献？太多不确定了，包括 20 年后我们的社会体制会是什么样。如果我们多造出几个亿万富翁，一路上替代掉大量工人，那就祝我们好运吧。
19:22 · AI 的方向由我们决定
19:29Sinead Bovell： AI 不是宇宙送来的礼物，是我们在造它，所以没有理由不能引导它。但要理解工作会怎样，是不是得先理解未来的市场结构和经济形态？
19:57Daron Acemoglu： 我经常说：预测 AI 的影响不像预测天气，因为 AI 是我们控制的，取决于我们拿它做什么。你还问 AI 时代的民主治理是什么样。我不知道。在某种意义上，我的想法很传统：我认为托克维尔说对了。民主和地方的、全国性的社团紧紧缠在一起，人们参与其中，形成交叉的成员身份，在这个过程中提供信息、公共服务、履行公民义务。这些社团被我们一点点毁掉了，罪魁不止一个，社交媒体也有份。也许有了 AI 我们能建起新的社团，那会很棒，但我没看到任何这样的计划。你强调的原则是核心：AI 没有唯一的方向，我们必须塑造它的方向，靠我们的制度，靠民主进程，还得让走在 AI 最前沿的公司参与进来。坐在白宫或者渥太华的某位官员决定不了 AI 的方向，方向会是领先公司决定的样子。但我们对这些公司有巨大的影响力，只是到现在都没用过。
22:23Daron Acemoglu： 我分三个层次回答。第一，现在大家强调通用人工智能和超级智能，好像 AI 模型和人非常像，所以理所当然应该做人能做的一切。但人工智能和人类智能非常不同，两者有很多互补的可能。特别是，AI 非常擅长给人提供依情境而定的、可靠有用的信息，让人去解决问题、做新的任务、发展新的专长和能力。我们得看到 AI 能做的事情的整个光谱。
24:17Daron Acemoglu： 第三点要小心对待。OpenAI、Anthropic 和其他公司告诉我们 AI 对工人有用，但他们说的意思和对工人有利的 AI 很不一样，经济后果也很不一样。他们举的例子是：如果你是最早用 AI 的那几个记者或作者，你的生产率会提高，做得更快更多，比同行占优势。这是真的，但这不是对工人有利的 AI。对记者或作者这个群体来说，等到大多数人都用上 AI，他们的技能就被商品化了。这和创造新的专长、新的能力完全不同。所以当科技公司说「看，我们的模型已经在帮工人了」，不，并没有。那只是一个过渡阶段，帮了少数人，代价是其他人，所有人迟早会醒过来，闻到硝烟味。
25:43Sinead Bovell： 那这个设计选择的代价谁来承担？我们的市场结构看重利润，公司对股东有受托责任。而且在美国、加拿大、欧洲，大多数新公司都依赖开源模型。如果我今天创业，我会从一开始就以 AI 为核心，商业模式可能怪得像 2005 年的 YouTube，组织架构很流动，也许根本不需要全职员工。这已经不是「工作」了，而是不断变化、流动的公司和人。
26:59Daron Acemoglu： 这些都对，也都很令人困惑。我很乐意把很多事怪到市场头上，但 AI 现在的方向，我不确定该怪市场。毕竟这些公司亏的钱多到你无法想象，市场并没有在奖励它们。这也是不确定的一部分。还有一个不确定：就算 Anthropic、Google、OpenAI、Meta 的承诺大体兑现，这些模型变成非常值钱的生产工具，也不保证它们能赚钱。
27:49Daron Acemoglu： 如果有两三个模型，再加几个能力差不多的开源模型，它们就收不了多高的价。利润会跑到经济的别处，跑到 AI 技术栈的别的环节。所以特别刺眼的是，风投、私人财富、主权基金砸了几千亿美元，却没有一条清楚的回本路径。这加剧了不确定，也带来一个我不愿看到的风险：到某个时候钱枯竭了，我们就会迎来衰退。
28:39Daron Acemoglu： 但关键正是你的问题：这些设计选择从哪里来，是谁做的？我的感觉是，领着这场冲锋的是一个非常紧密的小圈子，读同样的科幻，有同样的感性，在同样的环境里受训练、被塑造。Google 的创始人、Elon Musk、Sam Altman、Dario Amodei、他们的一些核心工程师，都出自同一个圈子，读同样的书，在成为敌人之前都曾是朋友。在人类历史上，我从没见过这么紧密的一小群人掌管这么多东西。你必须认清 AI 背后的意识形态。
29:30 · 几种可能的未来
29:51Sinead Bovell： 假设未来接近你的预期：AI 推动了生产率和一些 GDP 增长，但没有冲上天，也许像计算机时代，要等很久才在数据里看到效果。同时我们的起点是财富不平等本来就不乐观。你会拉哪些杠杆？
30:35Daron Acemoglu： 如果 AI 走对工人有利的方向，我预测，当然我不能确定，它会同时推动生产率增长和工资增长，成为共同繁荣的引擎，那时我们希望它尽快扩散。但如果发展中国家在基础设施和教育上掉队，对工人有利的 AI 就只会推动加拿大和美国的增长，而不是墨西哥或巴拉圭，这我们也得担心。而且未来不止一种，而是一个多维的连续谱。有中国式的未来，某些方面比美国好：他们坚持把 AI 融入生产过程，所以可能比美国更快拿到生产率收益；不过他们大量抄袭美国技术，这长期不可行。坏的一面是，AI 是强大的监控工具，已经让民众变得驯服。
32:00Daron Acemoglu： 如果走 Elon Musk 那条路，就是私营公司把持 AI，让它成为一种集中化的技术，可能导致两级社会，大量工人被边缘化；也可能导致一个非常不民主的社会：要么两级结构动摇了民主的根基，要么不平等高到当权者决定需要中国式的镇压。还有一种我前面提过：投资大幅放缓，一些公司破产，就像过去 AI 的一次次春天和冬天，这会是史上最大的一个冬天。其中有些在我们的掌控之内，这就是民主进程该起作用的地方；有些则完全不在。最后一点，我们的民主本来就在生病，这就是我写那本书的原因。我们得现实地看待能做成什么，并且趁还来得及去做。
05
UBI 不是答案：转向，而不是停车
UBI 融不起、等于认输、也挡不住两级社会；别再往通用人工智能竞赛砸一万亿，轻点刹车，把人才和资金转向对工人有用的应用。
33:41 · 就算快速起飞，UBI 也不是答案
33:40Sinead Bovell： 如果没有快速起飞，而我们又把 AI 塑造成对工人有利的，工人就可能像历史上那样受益，工资上涨、生产率成果共享，甚至不需要 UBI 这类新的再分配机制。但还有情景 B：出现突破，也许 Transformer 不是终点，起飞更快，过渡期动荡十年甚至更久。那我们需要不同的经济结构吗？所得税、整个体系是不是要改？
35:38Daron Acemoglu： 而且即使有了对工人有利的 AI，我们仍然需要对工人的大量公共支持。劳动力市场几乎肯定会变。今天仍有很多人在做重复性的工作，这些工作会交给 AI。我说对工人有利的 AI，不是说 AI 不该用于自动化。它会被用于自动化，也应该；只要自动化同时伴随着创造就业的东西，我们就欢迎。所以人们需要不同的技能，一部分来自职业培训，但主要来自学校，学校需要调整。很多工作会变得更动态，工人需要更灵活，顺便说，这是一种很难教的能力，在低收入的学校和社区尤其难。光有灵活还不够，我们还需要更好的社会保险。制度具体怎么调整，取决于我说的那些不确定性最后怎么落地。
36:50Daron Acemoglu： 没错，UBI 是非常线性的思维，我很喜欢这个说法。它是一个非常简单化的方案。第一，它行不通：你能想象以我们现在的政治经济状况，去资助一个像样的 UBI 吗？我想象不出来。第二，它等于认输，等于说 AI 已经失控，我们唯一能做的就是搞一套现代版的「面包加马戏」。第三，就算它行得通、而且很慷慨，也阻止不了我所说的两级社会。所有人都会明白，80%、70%，不管多少，的人是靠科技亿万富翁的面包屑过活的，这会造成一个非常陡峭的地位等级，和自由民主格格不入。
37:47 · 不是停车，是转向
37:53Sinead Bovell： 有些人提 UBI 是出于好意，但它把我们带到一种无能为力的状态，然后问：接下来呢？未来还没发生。而且「阻止这项技术」其实给了公司一个很方便的替罪羊。
38:17Daron Acemoglu： 我研究政治经济学及其历史花的时间，和研究技术一样多。我第一个告诉你：过去几乎每一次试图阻止或封杀技术，结果都是灾难。所以每当要人们抵制技术时，一定要非常小心，你需要一个积极主动的计划。我想传达的画面不是让 AI 放慢或停下来，而是：我们坐在一辆时速 200 英里的车里，冲向一道陡峭的悬崖，我们得把方向转开。时速 200 英里的车没法猛打方向，也许得轻点一下刹车，然后慢慢转。我们该考虑的是这种放慢：别再往通用人工智能竞赛里砸下一个一万亿美元了，想想怎么用我们非常稀缺的工程人才、资金和创业能量，去开发对社会、对工人更有用的应用。这就是转向。
45:56Daron Acemoglu： 我们已经在衡量的一些劳动力市场指标依然核心：人们拿到的工资、就业的选择和流动性。我们大概还需要更好地了解工作满意度，人们从哪里获得尊严。但关于 AI 模型，要衡量的东西多得多。我们需要一套审计 AI 模型的框架，知道它们能被拿去做哪些危害社会的事。有时这些被夸大了，但我不怀疑，比如 Mythos，确实有一些非常强大、可能被用来作恶的能力。监管机构和政府怎样才能理解这些，提前建好合适的护栏？这也是衡量问题的一部分。
48:13Daron Acemoglu： 我分别给企业 CEO、公民社会和决策者一些建议。先说 CEO，我的经验是，真懂自己生意的 CEO 会认同这条：别把劳动力当成要削减的成本。尤其在我们这个时代，世界在变，需要更多创新、更多新商品和新服务，劳动力是你最重要的资源。从这个心态出发，你对技术的需求会更多地是提高生产率、更好地发挥员工的作用，而不只是一味自动化；自动化既没那么容易，往往也没那么赚钱。
49:01Daron Acemoglu： 对公民社会，第一，我们都要先了解情况。第二，要认识到公民社会整体的力量，比我自己十到十五年前能想象的大得多。如果今天没有监管、没有护栏、没有人努力引导 AI，那是因为主流叙事支持的是这样一种 AI 图景：人人都会受益，AI 不可避免，天才们在领路。如果想从 AI 得到更好的未来，我们需要一个不同的叙事，而这要从公民社会开始。
49:47Daron Acemoglu： 对决策者，难处在这里。说 AI 有两颗跳动的心脏，一颗在中国，一颗在美国，有点夸张，但不算太夸张。我对美国的决策者有很明确的建议，但加拿大呢？你们可以在本地做事：鼓励本国公司用对 AI，鼓励人才进入 AI 的正确领域。但那是小事。决策者能做的更大的事，是推动建立制定政策的国际联盟。归根结底我们需要国际政策。加拿大、英国这样的国家，影响力比表面上看起来大，也有专业能力；如果能把很多国家团结在明确的目标周围，一起把 AI 引向正确方向，它们的力量还会更大。
54:32Daron Acemoglu： 不能。未来很难预知，也许这些 AI 投资会亏得一塌糊涂，让一些万亿富翁变成百万富翁。但问题在这里：今年大概有一万亿美元投进去，这会产生回报。如果回报接近人们的预期，资本拿到的就更多，劳动拿到的就更少，我们已经走在不平等扩大的路上了。对工人有利的方向能缓解这一点，对工人有利的模型会更专注于特定领域，也没那么集中。但即使把 AI 往对工人更有利的方向推一些，它的集中化倾向依然会在，所以我们还需要其他政策，反垄断在这里真的非常重要。财富不平等的问题应该和「我们想要多集中的经济」放在一起看。有些人以为集中化经济就是苏联。但如果一家公司、一个 AI 模型扮演非常重要的角色，那也是集中化，它会伴随着不平等和权力失衡。
58:59Daron Acemoglu： 我会跟学生强调三个原则。第一，学 AI。如果你从没用过 AI、不理解 AI，未来不可能过得好。第二，建立适应 AI 时代的规范，包括我们在 AI 时代的公民责任：负责任地使用 AI，参与辩论和民主政治。民主不能脱离人们的参与而存在，而这一点被我们忽视了。我们一直没有做好公民教育，要么自上而下地在学校灌输价值观，要么完全不管。到了 AI 时代，这更加重要。第三，灵活。没有哪个领域完全不受 AI 影响，但一个职业需要的技能里，会有很多部分仍然需要人来做。是哪些部分，常常并不清楚：不会是最重复的那些，但也许是需要更多经验、专长、社交互动或技术理解的部分。所以你需要足够全面的理解和扎实的基础，才能在任务之间灵活切换，找到自己贡献最大的地方。
A Nobel economist says AI is a real innovation, but the industry's “it will all be wonderful” is mostly untrue, and the Industrial Revolution analogy is a smokescreen.
Part 1 of 7 · 1:11
Trust the technology, doubt the claims
AI is a real innovation, but the industry's “it will all be wonderful” is mostly untrue; social media and past automation are the track record. Institutions must be as innovative as the moment.
AI is a real innovation, but the industry's “it will all be wonderful” is mostly untrue; social media and past automation are the track record. Institutions must be as innovative as the moment. Read this part →
The Industrial Revolution analogy is a smokescreen
People forget its brutal first 80 years. It was slow and local; AI is fast and broad, and several big sectors laying off at once would be far worse. Read this part →
Nobody knows what an AI economy looks like; we choose the direction
Everyone understood the internet's business model; not AI's. Pro-worker AI expands what people can do, rather than rewarding early users and then commodifying a whole profession's skills. Read this part →
Two or three models plus comparable open-source ones can't charge much, so profits move elsewhere. Hundreds of billions invested with no path to payback; if the money dries up, recession. Read this part →
UBI can't be funded, amounts to giving up, and can't prevent a two-tier society. Stop pouring another trillion into the AGI race; tap the brake and redirect talent and money to applications that help workers. Read this part →
Carte blanche first and reactive rules later is a recipe for disaster. We need a framework for auditing AI models and an international alliance; the AI race has wrecked cooperation with China. Read this part →
Audience questions: inequality, careers, population, a crash
AI will raise capital's share, so tax all income the same; students should learn AI, civic responsibility and flexibility; a crash is a risk, but the clock hasn't started. Read this part →
Indigo's conclusion
The weightiest opponent of the Industrial Revolution analogy, because he wrote that history. And from labor economics he independently reaches two conclusions: the path to returns is unclear, and profits will leave the model layer.
How to read this An institutional economist who trusts the technology and doubts the industry's claims, consistently making the case for “pro-worker AI” from his book Power and Progress. Not a doomer, not a bull: a moderate who says steer, don't stop. Epistemically honest: he keeps saying there's huge uncertainty, that he doesn't know what an AI economy looks like, and that pro-worker AI may not be feasible.
What to remember
Trust the technology, doubt the claims: AI is real, but “it will all be wonderful” mostly isn't; social media and past automation are the track record.
The Industrial Revolution analogy is a smokescreen: AI is fast and broad, so simultaneous layoffs would be worse. The nightmare is the 10-to-15-year transition, not the end state.
Returns are the biggest uncertainty: two or three models plus open source can't charge much, so profits leave the model layer; hundreds of billions invested with no clear payback.
Steer, don't stop: no more trillions into the AGI race; carte blanche first and reactive regulation later is a recipe for disaster.
What would change my mind
comprehensive AGI arrives quickly and models beat people at everything; then the pro-worker path no longer holds.
How to read this
An institutional economist who trusts the technology and doubts the industry's claims, consistently making the case for “pro-worker AI” from his book Power and Progress. Not a doomer, not a bull: a moderate who says steer, don't stop. Epistemically honest: he keeps saying there's huge uncertainty, that he doesn't know what an AI economy looks like, and that pro-worker AI may not be feasible.
AI is a real innovation, but the industry's “it will all be wonderful” is mostly untrue; social media and past automation are the track record. Institutions must be as innovative as the moment.
00:58 · Technology, companies, or institutions?
1:11Sinead Bovell: Today I'm in conversation with Daron Acemoglu, an MIT professor and one of the world's leading voices on how technology, power and institutions shape prosperity. In 2024 he won the Nobel Prize in Economics for his research on how institutions form and why they determine whether a society prospers. I want to start by taking stock of where we are with AI. It feels as though AI isn't working for society in many ways: people are pushing back on data centers, there's a lot of fear about a job apocalypse and concentration of power, and it all sits on top of an already wobbly tech stack, with social media for instance. Is the struggle we see a result of the technologies themselves, of the companies building them, or a failure of the institutions we have right now?
2:25Daron Acemoglu: Thank you for that question, it gets to the heart of it, and it has a very simple answer: yes. Can we move on? No, I'm kidding. It is the result of technology. It is the result of how the technology has been steered. And it's our responsibility, because we built the institutions and empowered the politicians and bureaucrats who allowed that to happen. That's why the issue of tech is intimately intertwined with democracy: what do we want our democracy to look like, and what can we do within a democratic framework to interact with a technology that's going to shape every aspect of our lives?
3:20Daron Acemoglu: So let me make my view clear. AI is a very powerful technology. It's not a gimmick; it is truly a big innovation with far-reaching consequences. That does not mean that everything the industry tells us is true. I'm still suspicious that we're going to have as much productivity growth as the industry tells us, and I'm certainly suspicious about some of the other beneficial claims. I would have been a little less suspicious if this had happened in 2005, before we saw what social media did, and less suspicious if it had happened before we saw what previous automation technologies did to the working and middle classes in the United States and other countries. We have a track record of technology having negative effects on at least some segments of society, and of our democratic system remaining passive. Those are the issues we have to grapple with.
4:47Sinead Bovell: Most of the institutions we lean on today hardened in the second industrial revolution or in the post-war environment. Institutions aren't eternal. It's one thing to make tweaks within them, but another to ask what we have to invent, the way we invented the Bureau of Labor Statistics or the FDA. Maybe institutions have to be as innovative as the moment. What could that look like?
5:22Daron Acemoglu: 100%. First, labor voice was very important during the industrial age and remains very important today, but trade unions as constituted in the 19th and early 20th century are no longer going to work. Labor organizations have to represent workers from much more diverse backgrounds, not centered on blue-collar workers, and they have to be much more conversant, even expert, in AI. Labor voice is still critical, but we need a different set of labor organizations. Same with self-government, which I think is the most important part of liberalism's ideas and the foundation of democratic institutions; what we mean by self-government may need to change. Many of the institutions we rely on require regulatory muscle, and you build that by exercising it. Right now in the United States we have no antitrust left, plain and simple. We have not exercised any antitrust oversight in the tech sector; we've allowed the biggest companies humanity has ever seen to take over all their rivals. We also don't have regulatory muscle or expertise when it comes to AI.
6:59Daron Acemoglu: But we also have to be realistic. Not every institution we wish for can exist, and some may not be feasible in the current polarized environment. Not every adaptation is feasible either. When I worry about people losing their jobs, being sidelined, losing the source of their dignity, some people in tech respond that they can find dignity in other ways. Perhaps that's a great idea, but we can't socially engineer people, and even if it were feasible we wouldn't get there immediately. There are things we may reach at some point in the next 300 years that aren't feasible in the near future, and we have to take those as constraints.
02
The Industrial Revolution analogy is a smokescreen
People forget its brutal first 80 years. It was slow and local; AI is fast and broad, and several big sectors laying off at once would be far worse.
08:12 · Does change come from inside or outside?
8:15Sinead Bovell: I read your book, every single page, and it looked as though change consistently happened outside the house. Labor organizing started, and institutions responded by formalizing labor relations. People were getting sick from scam medicines in the age of mass production, and the FDA formed. When we see people organizing around data centers, is that the origin story of a new institution, change from the outside in? Or does it happen inside the house?
8:57Daron Acemoglu: Both. In all your examples the agitation also came from inside, but they had to create new regulatory bodies. And sometimes people galvanize around a cause that is secondary in the grand scheme of things but merges into something else. In US history the People's Party, the original populist movement, had mostly rural, agricultural and economic grievances, and it merged with the progressive movement decades later. To me data centers are not where the big fight is. But if that's what exercises people, perhaps it's an opportunity for organization. That's where leadership comes in: keep your eye on the big prize. The big prize is not to stop AI; that would be a waste of such a promising technology. Nor is it small changes to social insurance so people don't starve when they lose their jobs to AI. It's finding a more socially beneficial direction for AI.
10:25 · Why the Industrial Revolution is the wrong comparison
10:23Sinead Bovell: You posted on X, in response to a petition some economists signed, that you don't like comparing AI's impact on the economy to the Industrial Revolution, that it feels like apples to oranges. Why isn't it a worthwhile comparison, given how poorly that went for the people living through it? And what's your forecast for the labor market?
11:09Daron Acemoglu: The specific reason is that when people in the tech conversation compare AI to the British Industrial Revolution, they are incidentally or purposefully forgetting the first 80 years and referring to all the good things it delivered later. I wrote Power and Progress with Simon Johnson to emphasize precisely that point, and it still gets ignored. So I fear that comparisons to the Industrial Revolution are often a coded way of saying this is all going to be wonderful.
12:02Daron Acemoglu: The second reason is that, despite all the hardship economic historians have documented, the Industrial Revolution was very slow and extremely localized. Two sectors define its first 80 years: textiles and coal. Not much happened in the rest of the economy, and that's the context in which the hardship transpired. Now imagine the speed with which AI is evolving. I'm the first to say we have not yet seen huge job displacement; whether and how we will is very uncertain. But imagine several major sectors laying off workers at the same time. That would be cataclysmic compared with the Industrial Revolution. We can still learn a lot from the past, absolutely.
13:33Sinead Bovell: So the telltale sign isn't the thriving we have now, it's that 70-to-80-year nightmare, localized. Imagine going through that at a wider scale, maybe over a decade. People get caught up in whether there will be no jobs at all, which probably won't happen, instead of asking what happens over that 10-to-15-year transition. That's where the nightmare could be.
Nobody knows what an AI economy looks like; we choose the direction
Everyone understood the internet's business model; not AI's. Pro-worker AI expands what people can do, rather than rewarding early users and then commodifying a whole profession's skills.
13:46 · Nobody knows what an AI economy looks like
14:12Sinead Bovell: What could an AI-first economy look like? We're in an internet-first economy that still has the rhythms of the industrial age: schedules, weekends, mass schooling. Before the Industrial Revolution nobody had weekends; we invented all that. Is the AI-first economy as differently shaped as the post-industrial economy was from the pre-industrial one?
14:51Daron Acemoglu: Absolutely, and you're asking exactly the right questions. I think the transition from the industrial to the post-industrial economy, which happened very slowly, was hugely disruptive; that's what my book was about. What happened to liberal democracy? In one word: post-industrial society. If we lose liberal democracy, and I think we're at the cusp of doing so, that is a disaster far worse than I could have imagined 10 years ago. Liberal democracy is the apex of our achievements. Given our history of warfare, selfishness, conflict, inequality and hierarchy, we built societies of millions of people with very complex interactions that live peacefully, create shared prosperity, and have all-powerful states that by and large provide public services and useful regulation, with everybody having a voice. That's amazing, and we're at the cusp of losing it. So if we get something as disruptive as the move to a post-industrial economy, we need a much better road map and much better guardrails.
16:52Daron Acemoglu: I'm a huge fan of the internet, though we've misused some of it; social media and online communication have been terrible. But by and large, and I was there as an early user, everybody understood what the internet would do. Even the companies that went bust in the dot-com crash had the right business model. Everybody laughs about pets.com, but it had the right model: use the internet as a new platform to provide more services and more varied goods more cheaply, competing with existing stores by giving customers more choice and better information. That model endured. Compare that to AI: we still have no idea what an AI economy will look like. That uncertainty really complicates the transition.
18:03Daron Acemoglu: I would love to tell you what the AI economy will look like in 20 years. I have no idea. I suspect there will be workers in it, but I don't know what they will be doing. Social tasks? Will they all be engineers and technicians looking after AI models? Will the models still be jagged, so they need a lot of hand-holding from human workers? Will workers find new tasks that contribute to innovation and productivity? There's so much uncertainty, including about what our social system will be like in 20 years. If we generate a few more billionaires and displace a lot of workers in the process, good luck to us.
19:22 · We decide which direction AI takes
19:29Sinead Bovell: AI isn't a gift from the universe; we're building it, so there's no reason we can't steer it. But don't you have to understand the shape of the future market and economy before you can understand jobs?
19:57Daron Acemoglu: I say this very often: forecasting the effects of AI is not like forecasting the weather, because we control AI. It depends on what we do with it. You also ask what democratic governance looks like in the age of AI. I don't know. My thinking is conventional in some sense: I think de Tocqueville had it right. Democracy is intertwined with local and national associations where people participate, form cross-cutting memberships, and provide information, public services and civic duties. We've destroyed them over time; there isn't one guilty party, but social media contributed. Perhaps with AI we'll build new ones. That would be amazing, but I don't see any plans for it. The principle you emphasize is central: there isn't a single direction for AI. We have to shape it, through our institutions and the democratic process, and we need the companies at the frontier on board. A bureaucrat in the White House or in Ottawa can't decide the direction of AI; it will be whatever the leading companies decide. But we have huge influence over these companies, which so far we have not exercised.
21:32 · What pro-worker AI means
22:05Sinead Bovell: So what is your thesis for pro-worker AI? How does it work in practice, and what would it mean for a company to deliver on it?
22:23Daron Acemoglu: Let me answer at three levels. First, contrary to the emphasis on AGI and superintelligence, which makes it sound like AI models are very similar to humans and should naturally do everything humans can do, artificial intelligence is very different from human intelligence. There are a lot of possibilities for complementarity. In particular, AI is a very powerful technology for providing context-dependent, reliable, useful information to humans, so they can solve problems, perform new tasks, and develop new expertise and capabilities. We have to see the whole spectrum of things we can do with AI.
23:35Daron Acemoglu: Second, by pro-worker AI I mean very specifically the end of that spectrum where AI becomes a tool for humans to expand what they can do and gain new expertise. That is critical for the human contribution to production to increase rather than for humans to be sidelined.
24:17Daron Acemoglu: Third, and this needs care: OpenAI, Anthropic and others tell us AI is useful to workers, but what they mean is very different from pro-worker AI and would have very different economic consequences. Their example is that if you're one of the first journalists or authors to use AI, your productivity rises, you do things faster and more of them, and you gain an advantage over competitors. That's true, but it isn't pro-worker AI. For journalists or authors as a whole, once most of them use AI, it commodifies their skills. That's very different from creating new expertise and new capabilities. So when tech companies claim their models are already helping workers: no, they're not. That's a transitional phase that helps a few people at the expense of others, and everybody's going to wake up and smell the napalm at some point.
04
A valuable model isn't a profitable one
Two or three models plus comparable open-source ones can't charge much, so profits move elsewhere. Hundreds of billions invested with no path to payback; if the money dries up, recession.
25:40 · A valuable model isn't a profitable one
25:43Sinead Bovell: Who bears the cost of that design choice? Our market structure is built on profits, with fiduciary duties to shareholders. And most new companies in America, Canada and Europe lean on open source. If I'm a startup founder today, I'm building AI-first with some strange business model, maybe a fluid org chart with nobody full-time. It's not even jobs; it's evolving, fluid companies and people.
26:59Daron Acemoglu: All of that is true and very confusing. I'm happy to blame the market for many things, but I'm not sure I'd blame it for the current direction of AI. These companies are losing more money than you can imagine, so the market isn't actually rewarding them. And here's another uncertainty. Even if what Anthropic, Google, OpenAI or Meta promise turns out approximately true, that these models become very valuable production tools, that does not guarantee they will make money.
27:49Daron Acemoglu: If there are two or three models, plus a few open-source models with roughly the same capabilities, they won't be able to charge that much. Some other part of the economy, some other part of the AI stack, will get the profits. So it's particularly jarring that venture capitalists, private wealth and sovereign funds are investing hundreds of billions of dollars without a clear path to getting that money back. That adds to the uncertainty, and it creates a risk I wouldn't want to see realized: at some point the money dries up and we have a recession.
28:39Daron Acemoglu: But the key is your question: where do these design choices come from, and who made them? My sense is that a very close-knit group, who read the same science fiction, share the same sensibilities, and were trained or socialized in the same milieu, is leading this charge. The Google founders, Elon Musk, Sam Altman, Dario Amodei, some of their leading engineers: they were all part of the same milieu, read the same books, and were all friends at some point before becoming enemies. I've never seen a period in human history where such a close-knit group has been in charge of so much. You have to recognize the ideology of AI.
29:30 · Many possible futures
29:51Sinead Bovell: Take a future close to what you expect: AI drives productivity and some GDP growth, nothing astronomical, maybe like the computer age where we waited a long time to see it in the numbers. And we start from wealth inequality that isn't looking great. What levers would you pull?
30:35Daron Acemoglu: If AI goes in a pro-worker direction, I forecast, though I can't be sure, that it will drive both productivity growth and wage growth, an engine of shared prosperity, and then we'd want it to spread rapidly. But if the developing world falls behind in infrastructure and education, pro-worker AI drives growth in Canada and the US and not in Mexico or Paraguay; we have to worry about that too. And there isn't one future but a multi-dimensional continuum. There's the Chinese future, better than the US in some respects: they're much more insistent on integrating AI into production, so they may get productivity gains faster, though they copy a lot of US technology, which isn't feasible in the long run. On the downside, AI is a powerful tool for surveillance that has pacified the population.
32:00Daron Acemoglu: If we go down the Elon Musk path, private companies dominate AI as a centralizing technology, possibly leading to a two-tier society in which a lot of workers are sidelined, or to a very non-democratic society, either because the two-tier structure shakes democracy's foundations or because inequality reaches levels where the powerful decide they need repression à la China. Or, as I mentioned, investment slows sharply, some companies go bankrupt, and like the AI springs and winters of the past, this becomes the mother of all winters. Some of these are under our control, which is where the democratic process comes in; some are not. And our democracy is already ailing; that's why I wrote the book. We have to be realistic about what we can achieve, and act before it's too late.
05
UBI isn't the answer: steer, don't stop
UBI can't be funded, amounts to giving up, and can't prevent a two-tier society. Stop pouring another trillion into the AGI race; tap the brake and redirect talent and money to applications that help workers.
33:41 · Even with a fast takeoff, UBI isn't the answer
33:40Sinead Bovell: If there's no fast takeoff and we shape AI to be pro-worker, workers could benefit as they have historically, with rising wages and shared productivity, and we wouldn't even need new redistribution mechanisms like UBI. But there's a scenario B: a breakthrough, maybe Transformers aren't the end, a faster takeoff and a wobbly transition for a decade or more. Is there a different economic structure we need, in how income tax and the whole system work?
35:01Daron Acemoglu: You're asking fantastic questions, and I wish I had better answers. Again, so much uncertainty. Even pro-worker AI: I've been advocating it and think it's our best chance, but it may not work; it may not be feasible. If we transition very quickly to something like comprehensive AGI, where AI models are better than us at everything, you can't really have meaningful pro-worker things.
35:38Daron Acemoglu: And even with pro-worker AI, we still need a lot of public support for workers. The labor market will almost certainly be different. Many people are still in routine jobs, and AI will do those. When I talk about pro-worker AI, I don't mean AI shouldn't be used for automation. It will be, and it should be; we welcome automation if it's coupled with things that create jobs at the same time. So people need different skills, from vocational programs but mostly from schools, which need to adapt. Many jobs will be more dynamic, so workers will need to be more flexible, which is a very difficult skill to teach, especially in low-income schools and neighborhoods. And flexibility isn't enough: we also need better social insurance. How exactly our institutions adapt depends on how those uncertainties resolve.
36:50Daron Acemoglu: Yes, UBI is very linear thinking, I love that term. It's a very simplistic solution. First, it won't work: can you imagine our current political economy funding a decent UBI? I can't. Second, it's throwing in the towel, saying AI is out of control and all we can do is create a modern version of bread and circuses. Third, even if it worked and was generous, it would not prevent what I've called a two-tier society. Everybody would understand that 80%, 70%, however many, live on the crumbs of the tech billionaires, and that steep status hierarchy would be very inconsistent with liberal democracy.
37:47 · Steer, don't stop
37:53Sinead Bovell: For some people UBI comes from the best intentions, but it takes us to a state of disempowerment and then asks, now what? The future hasn't happened yet. And "stop this technology" is actually an easy scapegoat for companies.
38:17Daron Acemoglu: I've spent as much of my career studying political economy and its history as studying technology, and I'd be the first to tell you that almost every attempt to stop or block technology in the past has been disastrous. So be very careful when asking people to resist technology; you need a positive, proactive plan. The image I try to communicate isn't slowing or stopping AI. We're in a fast car, driving 200 miles an hour toward a steep cliff, and we need to steer away from it. You can't swerve a car at 200 miles an hour; you may need to tap the brake a little and turn gently. That's the kind of slowdown to think about: let's not put another trillion dollars into the race for AGI. Let's use our engineering talent, which is very scarce, our funding and our entrepreneurial energy to develop applications more useful for society and for workers. That's the redirection.
39:53Sinead Bovell: And we lack long-term vision from leaders. We think in election cycles; we don't hear a leader say, here's why this technology will be worth it in 12 to 14 years if we build it this way.
40:14Daron Acemoglu: This is what I emphasize in almost every conversation: our approach to regulation is completely wrong. We are dealing with the most powerful corporations humanity has ever seen. First we give them complete carte blanche; they have lawyers to let them do whatever they want. Then in four or five years, if some of what they did turns out sufficiently disastrous, we add backward-looking, reactive regulation. That's a complete recipe for disaster. We need two realizations at once. First, proactive regulation: start with where we want to go and how to set up the institutions and governance to get there. Second, recognize you're dealing with the most powerful corporations imaginable. They're not our enemies, and there's a lot of innovative talent there, but they have their own agenda and they're very powerful. You have to take that into account.
06
Regulate ahead of time, and globally
Carte blanche first and reactive rules later is a recipe for disaster. We need a framework for auditing AI models and an international alliance; the AI race has wrecked cooperation with China.
42:03 · Regulation can be inventive, and it has to be global
42:01Sinead Bovell: Not all of it has to be regulation. With accounting, governments essentially outsourced oversight to private companies that make money keeping people in check. We could have regulatory markets; the economist Gillian Hadfield talks about this a lot. People can make money steering this in the right direction.
42:30Daron Acemoglu: The details really matter. Am I in favor of the military-industrial complex? Hell no; we spend too much and so much goes wrong there. But in the first two decades after World War II, DARPA, which came out of defense, was very forward-looking. It had a clear vision of technology and how to steer it, and it was extremely open-minded, venturing far beyond killing machines and laying the foundations of many later technologies. So how you support different kinds of technology, even in defense, can have very different consequences.
43:55Daron Acemoglu: And AI is a global technology; its applications will be global. It's insane to think we can content ourselves with national policy. The framing of a race toward AGI has had another pernicious effect: instead of enabling us to work with China, it created the impression of a zero-sum race and closed off collaboration with China, not just on AI. I'm a very vocal critic of the Chinese system; it has a lot of problems. But on AI, climate change, nuclear nonproliferation and pandemics we have to communicate and collaborate with China, and the AI race has ruined that.
45:14 · What we should start measuring
45:34Sinead Bovell: In the late 1800s and early 1900s we started measuring employment, unemployment, product safety. As we move into this new era, what should institutions start measuring?
45:56Daron Acemoglu: Some labor market outcomes we already measure remain central: the wages people receive and their employment options and mobility. We probably also need to understand job satisfaction, where people find dignity, a bit better. But there's much more to measure about AI models. We need an auditing framework for AI models, so we know what they can be used for that's damaging to society. Sometimes these things are exaggerated, but I don't doubt that Mythos, for example, had very powerful capabilities that could have been used for ill. How can regulators and governments understand those things and build the right guardrails ahead of time? That's part of the measurement problem.
47:37 · Advice for CEOs, civil society and policymakers
47:40Sinead Bovell: We have a diverse audience, from key decision-makers in governments and alliances to civil society activists. What message should policymakers take away, and how can the general public think non-linearly and see how much power we have?
48:13Daron Acemoglu: Let me give advice to CEOs, civil society and policymakers. To CEOs, and in my experience CEOs who know their business sympathize with this: don't think of labor as a cost to be cut. Especially in our age, when we need more innovation and more new goods and services in a changing world, labor is your most important resource. Start with that mindset and you can articulate a demand for technology that's much more about raising productivity and using your workers well than just automating, which is neither that easy nor often that profitable.
49:01Daron Acemoglu: For civil society, first, we all need to get informed. Second, recognize that civil society as a whole has much more power than I could have imagined 10 or 15 years ago. If today there are no regulations, no guardrails, no effort to steer AI, it's because the narrative has supported a vision where everybody will benefit, AI is inevitable, and geniuses are leading the charge. If we want a better future from AI, we need a different narrative, and that starts with civil society.
49:47Daron Acemoglu: For policymakers, here's the difficulty. It wouldn't be a massive exaggeration to say AI has two beating hearts, one in China and one in the US. I'd have very clear advice for US policymakers, but what about Canada? You can do things locally: encourage your companies to use AI the right way, and your talent to go into the right fields of AI. But that's small. The bigger thing policymakers can do is work toward an international alliance. At the end of the day we need international policy. Countries like Canada and the UK have more clout than first meets the eye, and expertise, and more power still if they bring many countries together around well-defined aims for steering AI the right way.
51:33Sinead Bovell: Your book says labor could organize more easily in manufacturing because everyone was in one place, 9 to 5, doing the same thing. The irony of social media is that we're also all there.
51:47Daron Acemoglu: It's not just proximity; it's solidarity. Workplaces were repositories of organized resistance because they built solidarity. So the question is whether we can find online spaces that build solidarity. I don't think the answer is no, but right now I know of no example, and social media kills solidarity rather than building it. For solidarity we need community; that's why a large part of my book is about community and its role in freedom.
52:41Sinead Bovell: One small example: in Tokyo's latest election, the first candidate under 40 to gain a meaningful share came out of nowhere with 2.4% or 2.5% of the vote and finished fifth. He followed Audrey Tang, Taiwan's former digital minister, and her idea of broad listening: using technology to hear what people are saying and what they need, with nuance.
53:32Daron Acemoglu: I didn't know about Tokyo. In Power and Progress I give Audrey and her achievements in Taiwan as an example of how digital technologies, and AI more generally, could be used in a more pro-democracy way. There's a lot more that can be done; we're just not doing it.
07
Audience questions: inequality, careers, population, a crash
AI will raise capital's share, so tax all income the same; students should learn AI, civic responsibility and flexibility; a crash is a risk, but the clock hasn't started.
54:00 · Audience questions
54:19Sinead Bovell: Our audience sent questions. Girish Megalani asks: will AI enable us to improve wealth distribution and overcome the extreme concentration at the top?
54:32Daron Acemoglu: No. The future is very hard to know, and it may be that these AI investments go bust so badly that some trillionaires become millionaires. But here's the problem: this year probably about a trillion dollars is being invested. If it has anything like the returns people expect, that's even more return to capital and less to labor, so we're already on a path to wider inequality. The pro-worker direction would ameliorate that; pro-worker models would be more domain-specific and less centralized. But AI's centralizing tendencies would persist even in a somewhat more pro-worker direction, so we need other policies, and antitrust is really important. The wealth inequality question should be bundled with how centralized we want our economy to be. Some people think centralization means the Soviet Union. But if one company or one AI model plays a very important role, that's also centralizing, and it will come with inequality and imbalances of power.
56:03Sinead Bovell: But can't we adjust how returns are split between capital and labor?
56:08Daron Acemoglu: Yes, that's why we need policy. One policy I've advocated for a long time, on fairness and economic efficiency grounds, is getting rid of the distortion in our tax system whereby we tax labor and don't tax capital. We may need to tax capital returns much more. A feasible, effective, fair and efficient model would tax all income the same, whether or not we label it capital income, and make sure capital income can't go to the Cayman Islands or be hidden one way or another. That would really change things.
56:54Sinead Bovell: Natasha Bowman asks: what's the best measure for reducing wealth inequality?
57:02Daron Acemoglu: Keep three levels in mind and reduce all three: labor income inequality, income inequality and wealth inequality. Labor income inequality drove the huge increase in US inequality in the '80s, '90s and 2000s. Income inequality rose further as capital returns increased. Then, because we didn't tax capital or exercise antitrust, wealth inequality ballooned. I should add that labor income and income inequality are measured very well, so I'm 99% sure they rose in the US. Wealth inequality is badly measured, so there's more debate, though it seems pretty obvious it rose. It would be far easier to deal with labor income and income inequality by redirecting AI to create more jobs, opportunities and wage growth. Once wealth inequality is as huge as it is today, you can't rectify that that way. Elon Musk's wealth will stay at $1 trillion unless he loses it or his companies go bust. That's where taxation and anti-monopoly measures come in.
58:41Sinead Bovell: John Polo asks: what careers should high school and college graduates go into?
58:59Daron Acemoglu: Three principles for students. First, learn AI; there's no way you'll do okay in the future if you've never worked with it or don't understand it. Second, develop norms adapted to the AI age, including our civic responsibilities: using AI responsibly and participating in debates and democratic politics. Democracy doesn't exist independent of people's participation, and we've neglected civic education, either imposing values top-down or ignoring them. That matters even more with AI. Third, flexibility. No area is completely immune to AI, but many parts of the skills an occupation requires will still need human input. Which parts isn't always clear: not the most routine ones, but perhaps those needing more experience, expertise, social interaction or technical understanding. So you need a holistic enough grounding to shift across tasks and find where your contribution can be greatest.
1:02:04Sinead Bovell: Rukob asks: what are the plans for dealing with a growing human population?
1:02:13Daron Acemoglu: Depending on your perspective, the news is good or bad: world population will stop growing in about a decade or two, and it already has in most of the developed world. Some worry that declining population creates macroeconomic and innovation problems and a lack of young workers; some celebrate it. I think it creates risks, but my research also shows that labor scarcity, when entering cohorts are small, is quite good for wage growth and for how we use technology. When labor is scarce, companies must economize on it, which makes automation more productive but also gives them incentives to use existing labor well. That's why German companies didn't lay off blue-collar workers when they introduced robots: labor was so valuable, and they had trained it, so they retrained workers as technicians. We need a different mindset, but it can work very well, and historically it has.
1:03:58Sinead Bovell: Charlotte Holland asks: will the $40 trillion debt, high energy costs, high inflation and a stagnant job market cause a crash?
1:04:08Daron Acemoglu: Possibly. I don't have a crystal ball. Stock markets are very forward-looking; if there were a general consensus of a crash next year, it would crash today. I'd say there is certainly some hype in AI, and the investments are amazingly large. But there's also a lot of private wealth and money in sovereign wealth funds, SoftBank, the UAE, Saudi Arabia, US billionaires, so they can keep financing hundreds of billions of dollars of losses for quite a while. It's hard to know whether there'll be a slowdown or a hard stop, but the risk is there. If you're squeamish about risk, take it into account in your portfolio.
1:05:16Sinead Bovell: Finally, Nikki asks: what can everyday people do to shape how we use AI?
1:05:25Daron Acemoglu: Become informed. Turn up to discussions. Be part of your community. Try to improve the quality of the conversation. Perhaps I'm naive, but if in the late 2010s, instead of the unproductive false dichotomy of AI as our savior led by geniuses versus killer robots coming for us, we had started the conversations we're having today, we would be in a much better place. It's the responsibility of all of us, not just journalists, to raise the conversation to that level.
1:06:17Sinead Bovell: The future is much more nuanced than jobs or no jobs, AI is amazing or AI is terrible. That's not how society or history has ever worked. To reach a different future, we have to meet the moment with the nuance it requires.
Where Indigo landsFurther
Indigo's conclusion
The weightiest opponent of the Industrial Revolution analogy, because he wrote that history. And from labor economics he independently reaches two conclusions: the path to returns is unclear, and profits will leave the model layer.
What to remember
Trust the technology, doubt the claims: AI is real, but “it will all be wonderful” mostly isn't; social media and past automation are the track record.
The Industrial Revolution analogy is a smokescreen: AI is fast and broad, so simultaneous layoffs would be worse. The nightmare is the 10-to-15-year transition, not the end state.
Returns are the biggest uncertainty: two or three models plus open source can't charge much, so profits leave the model layer; hundreds of billions invested with no clear payback.
Steer, don't stop: no more trillions into the AGI race; carte blanche first and reactive regulation later is a recipe for disaster.
Claims you can check later
Claim
Who
When we will know
How firm
The Industrial Revolution analogy misleads: AI is fast and broad, and several big sectors laying off at once would be worse
Acemoglu
10-to-15-year transition
First-hand; an economic historian's judgment
However valuable, models aren't guaranteed to make money; profits will leave the commodified model layer
Acemoglu
Ongoing
First-hand; reached independently
With no clear path to returns, money may dry up and cause a recession (though private and sovereign wealth can keep it going for a while)
Acemoglu
No date
First-hand; a risk judgment
Pro-worker AI is the best chance, but fails if comprehensive AGI arrives quickly
Acemoglu
Conditional
First-hand; he admits the gap
UBI can't be funded and can't prevent a two-tier society
Acemoglu
First-hand; a political-economy judgment
World population stops growing within one or two decades; labor scarcity is good for wages and for good uses of technology
Acemoglu
Around the 2040s
First-hand; backed by his research
Back on the long-running theses
confirms
The AI jobs apocalypse Acemoglu is the economist this view names for “technology doesn't automatically benefit workers”; this is his latest and fullest first-hand statement.
confirms
AI capability is a bounded exponential: paradigm shifts don't come from hill-climbing Economic testimony for the slower path: returns arrive more modestly than claimed, and the ceiling also comes from economic and institutional friction.
