内容来源：https://www.geekwire.com/2026/etzioni-on-ai-does-ai-bolster-or-undercut-democracy/

内容总结：

美国250周年：AI是民主的助力还是威胁？

美国刚刚迎来建国250周年。开国元勋们为小册子和市政会议的世界设计了自治制度，而如今，我们却用人工智能来运行这套政治架构。

这个生日提出的核心问题是：AI究竟是巩固民主，还是削弱民主？严肃的思想家们站在对立双方，各持充分论据。

以下是我基于五本书和七篇文章梳理出的“评分卡”，以及双方都未提出的问题：对AI的控制权和获取权，哪个增长得更快？

监视：AI让专制成本降至“分文不取”

尤瓦尔·赫拉利在《纽带》一书中指出，民主是具备自我纠错机制的分布式信息网络——自由媒体、反对党和法院负责发现并修正错误；而专制则是压制纠错的集中化网络。两个世纪以来，集中化有其内在成本，因为全面监视需要庞大的告密者队伍，而“人海战术”代价高昂。AI消除了这一成本——它以“分文计价”的成本全天候监视所有人。这已非假设。《经济学季刊》的研究记录了中国的地方反馈循环：地方出现动荡，政府采购人脸识别AI，而采购行为又抑制了后续动荡。论文标题直接命名为《AI专制》。

经济：AI瞄准整个劳动力市场

过去的技术取代的是特定岗位——接线员、收费员——同时为操作新机器的人创造了就业。但AI的野心覆盖整个劳动力大军。2024年诺贝尔奖得主阿西莫格鲁在《权力与进步》中写道：“当前AI的发展路径既不利于经济，也不利于民主。”今年2月他在《财富》杂志进一步警告，按照目前的就业摧毁和贫富分化趋势，“美国民主将难以存活”。

自治机制：AI工业化地生产“假货”

我早在2019年就曾在《哈佛商业评论》发出警告：AI将使高清伪造视频、音频和文件变得廉价且自动化，对民主造成灾难性后果。伪造古已有之，而AI将其产业化。安全专家施奈尔预测，AI将优化游说能力、草拟“微立法”——那些悄悄为特定群体谋利的小条款。他观察到，这项技术主要让强者更强。沙克在《科技政变》中提供了制度层面的结论：未经选举的公司现在承担着本属于政府的职能。

辩护方：AI让稀缺变得丰裕

7月4日，计算机科学家科勒以参观沙斯塔大坝纪念美国250岁生日。她提出，美国的标志性成就是“将稀缺变为丰裕”——水变成电，电变成人人可接入的电网，计算设备装进口袋。她联合创立的Coursera已让1.5亿人获得精英教育。AI则是下一章，“让世界上最稀缺的资源之一——强大推理能力——变得丰裕”。过去属于持证专家的判断力，现在属于任何能提出正确问题的人。

经济反论：AI可能重塑中产阶层

阿西莫格鲁的MIT同事奥托尔在《Noema》杂志中论证，AI可以将专业知识延伸至缺乏精英证书的工人，从而重建空心化的劳动市场中段。早期证据支持这一观点：一家财富500强公司为客服人员配备AI助手后，生产力平均提升15%，收益最大的是最缺乏经验的新手，速度和品质双双提升。如果这一模式成立，AI可能压缩阿西莫格鲁所担心的那些鸿沟。

最有趣的发现：双方争论的从未交汇

悲观者在争论“谁控制AI”，乐观者在争论“谁使用AI”。权力和获取是不同的问题，双方可能同时正确。

科勒的大坝从物理层面说明了这一点：发电是集中的——少数几台涡轮机由少数人拥有；电网是分布的——任何人都能接入。同一个机器同时完成两件事。AI具有同样的结构：任何人花20美元就能接入一个前沿模型，而模型权重和训练它们的数据中心掌握在六家公司手中。

古腾堡印刷术则展示了时间维度：印刷术打破了罗马对圣经的垄断，四个世纪后它又建立了赫斯特的传媒帝国——获取权和支配权在同一台机器上交换了位置。两种力量都是真实的。悬而未决的问题是：哪个跑得更快？当前关于开放权重、芯片出口和模型所有权的争论，正是在帮助回答这个问题。

科勒以一段寄语结束她的帖子，这段话恰如其分地适用于美国建国250年：任何获得了超出自己份额之利益的人，都有责任确保下一个稀缺之物不会长久稀缺。智能是下一个稀缺之物。大坝已经建成，前沿模型和数据中心也已建成。摆在我们面前的选择是：我们是否也要建设电网——为所有美国人提供广泛、廉价的人工智能接入？

中文翻译：

美国刚刚迎来250岁生日。开国元勋们当初为 pamphlets（小册子）和 town meetings（镇民大会）的世界设计了自治制度，而如今我们却在用人工智能运行这套政治架构。
这个生日提出的问题是：人工智能究竟是巩固民主，还是削弱民主？严肃的思想家们已分列两派，各自提出了有力的论证。
以下是我基于五本书和七篇文章提炼出的评判表，以及双方都未提出的问题：对人工智能的控制权与使用权，哪个增长得更快？

先从 surveillance（监控）说起。
尤瓦尔·赫拉利在《Nexus》中指出，民主制是一个具有自我纠错机制的分布式信息网络：自由媒体、反对党和法院能发现错误并加以修正。而独裁制则是压制纠错的中央集权网络。两个世纪以来，集权制一直带有内在成本，因为全面监控需要大量人类告密者，而养一支告密者军队费用高昂。人工智能消除了这一成本。它能以极低的代价全天候监控每一个人。相关证据已不再是假设。《Quarterly Journal of Economics》中的一项研究记录了这一反馈循环：地方动荡导致政府购买面部识别人工智能，而这些采购又压制了后续的动荡。该论文作者将这篇论文命名为《AI-tocracy》。

第二个论点是经济层面的。
过去的技术取代了特定工人——电话接线员、收费员——同时为操作新机器的人创造了就业岗位。而人工智能的野心瞄准了整个劳动力市场。达龙·阿西莫格鲁和西蒙·约翰逊在专著《Power and Progress》中专门论述了这一担忧，写道：“当前的人工智能发展路径既不利于经济，也不利于民主。”2024年诺贝尔奖得主阿西莫格鲁今年2月在《财富》杂志上进一步强调了这一点，警告称，按照目前工作岗位被摧毁且不平等加剧的趋势，“美国民主将无法存活”。

第三个论点直指自治制度本身。
我早在2019年的《哈佛商业评论》中就发出过这一警告，指出人工智能将使高保真伪造视频、音频和文件变得廉价且自动化，可能对民主造成灾难性后果。伪造并非新鲜事。但人工智能将其工业化。安全技术专家布鲁斯·施奈尔预测，人工智能将优化游说活动并起草“微立法”——那些悄无声息地让某个群体受益的微小条款。他观察到，这项技术主要让强者更强。当他与内森·桑德斯看到一封由人工智能撰写、反对人工智能监管的信件在《纽约时报》上发表时，他们开始真正担忧起来。玛丽杰特·沙克在《The Tech Coup》中提供了制度层面的总结：未经选举产生的公司如今正在履行原本属于政府的职能。

控方陈述完毕。现在轮到辩方。
7月4日，计算机科学家达芙妮·科勒在参观沙斯塔水坝时，既庆祝美国建国250周年，也纪念自己作为移民来到这个国家37周年。在当天发布的一篇感想中，她认为美国的标志性成就在于将稀缺变为富足：在沙斯塔将水变为电力，将电变为任何人都能接入的电网，将计算能力放入口袋。她本人就做到了这一点；她共同创立的Coursera让超过1.5亿学习者获得了精英教育。她写道，人工智能是下一个篇章，“让世界上最稀缺的资源之一——强大的推理能力——变得富足”。过去只有获得认证的专家才能做出的判断，如今任何能提出正确问题的人都能获得。律师和医生按小时收费。人工智能按秒回答。

经济方面的反驳来自阿西莫格鲁在麻省理工学院的同事大卫·奥特尔。他在《Noema》中论证道，人工智能可以将专业知识延伸到没有精英资质的工人手中，从而重建劳动力市场中已空心化的中间层。早期证据支持他的观点。当一家财富500强公司为其客服人员配备人工智能助手后，生产率平均提高了15%，而收益绝大部分流向了最新手、技能最低的工人——他们在速度和品质上都有提升。这项发表在《Quarterly Journal of Economics》上的研究发现，最有经验的员工几乎没有获益。如果这一模式成立，人工智能或许能弥合阿西莫格鲁所担心的那些不断扩大的差距。

里德·霍夫曼和格雷格·贝亚托在《Superagency》中概括了乐观派的观点：人工智能如此广泛地放大了个人自主性，以至于真正的危险在于民主国家将其发展拱手让给不那么仁慈的行动者。在《Plurality》一书中，台湾首位数字事务官员唐凤和经济学家格伦·韦尔描述了十年来数字工具如何在两极分化的公众中，围绕实时立法——从网约车规则到防疫政策——达成共识。一项受控实验支持了他们的观点。谷歌DeepMind的研究人员构建了一个人工智能调解员，对5734名英国人进行了测试，让他们讨论英国脱欧和移民等问题。研究结果发表在《科学》杂志上，显示参与者更倾向于人工智能生成的团体声明而非人类调解员的声明，认为其更清晰、偏见更少。这些团体最终的分歧也更小。镇民大会从未能容纳一百万人。但现在或许可以。

我将两列观点并列，注意到一个奇怪的现象：它们从未交锋。悲观派争论的是谁控制人工智能。乐观派争论的是谁使用人工智能。控制权与使用权是不同的问题，而两派可能同时正确。

科勒的水坝直观地说明了这一点。发电是集中的——少数人拥有几台涡轮机。电网是分布式的——任何人都可以接电。同一台机器同时完成这两件事。人工智能也有类似的结构：任何人每月花20美元就能接入一个前沿模型，而前沿模型的权重以及训练它们的数据中心，却属于六家公司。

古腾堡印刷术则增加了时间维度。印刷术打破了罗马对经文的垄断，四个世纪后又建起了赫斯特的帝国；使用权和控制权在同一部机器上互换位置。两种力量都是真实的。悬而未决的问题是哪个增长得更快，而当前围绕开放权重、芯片出口和模型所有权的斗争，正是将有助于解答这一问题的博弈。

开国元勋们曾面临类似的权力集中问题，他们的答案是分配投票权——起初范围很窄，后来几乎推广到所有人。科勒在帖文结尾提出了一项符合美国250岁生日的义务：任何获得超出自己份额的人，都有责任确保下一种稀缺资源不会长久稀缺。智力就是下一种稀缺资源。科勒的水坝已经建成，前沿模型和训练它们的数据中心也已就绪。摆在我们面前的选择是：我们是否也建设电网——为所有美国人提供广泛、廉价的人工智能接入？

英文来源：

America just turned 250. The founders designed self-government for a world of pamphlets and town meetings, and we now run their political architecture on AI.
The birthday question is whether AI bolsters democracy or undercuts it. Serious thinkers have lined up on both sides with substantial arguments.
Here is my scorecard, distilled from five books and seven articles, and then the question neither side asks: which is growing faster, power over AI or access to it?
Start with surveillance.
Yuval Noah Harari argues in Nexus that a democracy is a distributed information network with self-correcting mechanisms: a free press, opposition parties, and courts that catch mistakes and fix them. A dictatorship is a centralized network that suppresses correction. For two centuries, centralization carried a built-in cost, because total surveillance required armies of human informants, and armies are expensive. AI removes the cost. It watches everyone, all the time, for pennies. The evidence is no longer hypothetical. A study in the Quarterly Journal of Economics documented the feedback loop in China: local unrest leads to government purchases of facial-recognition AI, and those purchases suppress subsequent unrest. The authors titled their paper “AI-tocracy.”
The second argument is economic.
Past technologies replaced particular workers, the switchboard operator, the toll collector, while creating jobs for the people who ran the new machines. AI’s ambition targets the entire workforce. Daron Acemoglu and Simon Johnson devoted a book, Power and Progress, to this worry, writing that “the current path of AI is neither good for the economy nor for democracy.” Acemoglu, a 2024 Nobel laureate, sharpened the point in Fortune this February, warning that on the current path of job destruction and rising inequality, “U.S. democracy is not going to survive.”
The third argument targets the machinery of self-government itself.
I sounded this alarm in Harvard Business Review back in 2019, warning that AI was poised to make high-fidelity forgery of video, audio, and documents cheap and automated, with potentially disastrous consequences for democracy. Forgery is ancient. AI industrializes it. Security technologist Bruce Schneier predicts that AI will optimize lobbying and draft “micro-legislation,” tiny provisions that quietly benefit one group, and he observes that the technology mostly makes the powerful more powerful. He and Nathan Sanders began worrying in earnest when an AI-written letter opposing AI regulation ran in the New York Times. Marietje Schaake supplies the institutional capstone in The Tech Coup: unelected companies now perform functions that once belonged to governments.
The prosecution rests. Now comes the defense.
On July 4, computer scientist Daphne Koller marked the country’s 250th birthday, and her own 37th anniversary as an immigrant, with a visit to Shasta Dam. In a reflection posted that day, she argued that America’s signature achievement is taking what was scarce and making it abundant: water into power at Shasta, electricity into a grid anyone could plug into, computation into a pocket. She has done it herself; Coursera, which she co-founded, put an elite education in front of more than 150 million learners. AI, she wrote, is the next chapter, “making abundant one of the world’s scarcest resources: powerful reasoning.” The judgment once reserved for credentialed specialists now belongs to anyone who can frame the right question. Lawyers and doctors bill by the hour. AI answers by the second.
The economic counter comes from Acemoglu’s MIT colleague David Autor, who argues in Noema that AI can extend expertise to workers without elite credentials and thereby rebuild the hollowed-out middle of the labor market. Early evidence points his way. When a Fortune 500 firm gave its customer-support agents an AI assistant, productivity rose 15% on average, and the gains went overwhelmingly to the newest and least skilled workers, who improved in both speed and quality. The study, published in the Quarterly Journal of Economics, found that the most experienced agents gained little. If the pattern holds, AI could compress the very gaps Acemoglu fears it will widen.
Reid Hoffman and Greg Beato’s Superagency states the optimistic case in general form: AI amplifies individual agency so broadly that the real danger lies in democracies ceding its development to less benevolent actors. In Plurality, Taiwan’s first digital minister Audrey Tang and economist Glen Weyl describe a decade of digital tools that found consensus across a polarized public on live legislation, from ride-sharing rules to pandemic policy. A controlled experiment backs them up. Google DeepMind researchers built an AI mediator, tested it on 5,734 Britons deliberating questions like Brexit and immigration, and reported in Science that participants preferred the AI’s group statements to a human mediator’s, rating them clearer and less biased. The groups also ended up less divided. A town hall has never fit a million people. It might now.
I set the two columns side by side and noticed something odd: they never meet. The pessimists are arguing about who controls AI. The optimists are arguing about who gets to use it. Power and access are different questions, and both camps can be right at the same time.
Koller’s dam makes the point physically. Generation is concentrated, a handful of turbines owned by a few. The grid is distributed, and anyone can plug in. One machine does both at once. AI shares that anatomy: anyone can plug into a frontier model for $20 a month, while the frontier weights and the data centers that train them belong to a half-dozen companies.
Gutenberg adds the time dimension. The press broke Rome’s monopoly on scripture, and four centuries later it built Hearst’s empire; access and power traded places on the same machine. Both forces are real. The open question is which one moves faster, and the current fights over open weights, chip exports, and model ownership are fights that will help settle this question.
The founders faced a similar question about concentrated power and answered it by distributing the vote, narrowly at first, and later to nearly everyone. Koller ended her post with an obligation that fits the country’s 250th year: anyone given more than their share owes the work of making sure the next scarce thing does not stay scarce for long. Intelligence is the next scarce thing. Koller’s dam is already built, along with the frontier models and the data centers that train them. The choice in front of us is whether we also build the grid, providing broad, cheap access to AI for all Americans.