私有化的收益，社会化的损失：国会必须强硬反对大型人工智能公司一家仍处于起步阶段、盈利前景 questionable 的前沿技术公司，正处在一场轰动的首次公开募股(IPO)之前的几周时间。它的联合创始人是这项大多数人和政府都不理解的技术的代名表。
Privatized gains, socialized losses: Congress must forcefully say no to Big AI A frontier technology company still in its infancy, of questionable profitability, is weeks away from a blockbuster initial public offering. Its co-founder is the poster child for a technology most people and governments do not understand.
最近的新闻充满了一种颓废主义的公共服务公告，宣扬这项技术的风险，尤其是对儿童的影响。华盛顿对此分歧已见。一位民主党参议员希望对任何在儿童可能看到的情况下上传“猥亵”内容的人处以监禁。而一对跨党派的国会议员则持相反观点，希望为用户发布的内容提供诉讼豁免权，这样公司就可以自行管理和监管自己的产品。
Recent news has been one long, nihilism-inducing public service announcement about this technology’s risks, particularly for children. Washington is split on a cure. A Democratic senator wants prison time for anyone who uploads “indecent” material where a child might see it. A bipartisan pair of congressmen want the opposite: immunity from lawsuits over what users post, so companies can police themselves and their products.
硅谷的策略简单明了：不要过于仔细审查你的产品或你的用户正在做什么，只管尽快吸引尽可能多的用户。强硬的政府监管威胁到了维持美国竞争优势的根本要素：自由、主权和追求利润。
Silicon Valley’s playbook is simple: Don’t look too closely at what your product, or your users, are doing; just get as many users as fast as possible. Heavy-handed government regulation threatens the essence of what will keep America winning: freedom, sovereignty and the pursuit of profit.
但上述描述并不是关于人工智能、Anthropic 和 Dario Amodei 的今天——而是关于1995年的互联网、Netscape 和 Marc Andreessen。那位民主党参议员是 James Exon（D-内布拉斯加）。那些共同发起《通信法》中臭名昭著的第230条的国会议员是 Chris Cox（R-加利福尼亚）和 Ron Wyden（D-俄勒冈）。
But the vignette above is not about artificial intelligence, Anthropic and Dario Amodei today — it is about the internet, Netscape and Marc Andreessen in 1995. The Democratic senator was James Exon (D-Neb.). The congressmen Chris Cox (R-Calif.) and Ron Wyden (D-Ore.), who co-sponsored the infamous Section 230 of the Communications Act.
这项法律条文看似并无危害：“任何互动式计算机服务的提供者或使用者，均不应被视为其他信息内容提供者所发布信息的发布者或发言人。” 科克斯与怀登此举旨在解决一个现实的商业难题：当时，拨号上网服务提供商即便只是对信息进行适度管控，也会遭到起诉。他们未曾料到，这一原本旨在鼓励行业自律、规避诉讼风险的规定，日后竟会成为社交媒体时代的法律与商业基石，使得各大平台得以吸纳数十亿用户，却无需为算法所传播的内容负责——即便这些内容可能导致用户伤害自身或他人。
The law’s words sound harmless: “No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider.”Cox and Wyden were solving a real business problem: Dial-up services were being sued for moderating at all. They could not foresee a litigation liability shield meant to encourage disciplined self-regulation becoming the legal and business foundation of the social media era, letting platforms grow to billions of users without answering for what their algorithms amplify, including when users harm themselves or others.
2026年，硅谷再度提出新的诉求：要求获得反垄断豁免权，以及针对其系统与用户所生成内容的诉讼责任保护。本月，阿莫迪发表了一篇广为流传的文章，呼吁实施前者，以便各前沿实验室能够同步放缓研发步伐。9月15日，财政部长斯科特·贝森特在众议院委员会上表示，这些实验室不应享有责任豁免权。
In 2026, Silicon Valley is back with new demands — an antitrust waiver and liability protection from lawsuits over what their systems and their users generate. This month, Amodei published a viral essay asking for the former, so that frontier labs can slow down together. On Sept. 15, Treasury Secretary Scott Bessent told a House committee these labs should not get a liability exemption.
“确保安全的最佳方式，就是让技术的创造者为其研发与生成的内容承担相应责任，”他说道。
“The best way to guarantee safety is that the creators are liable for what they build and generate,” he said.
据悉，Anthropic计划于10月上市，估值有望达到2万亿美元，这一数值大约是网景公司市值的700倍。
Anthropic reportedly plans a $2 trillion October IPO, roughly 700 times the value of Netscape’s.
值得庆幸的是，近期反对这些诉求的声音获得了两党支持。乔·拜登政府负责反垄断事务的官员乔纳森·坎特对CNBC表示，各前沿实验室即便没有反垄断豁免权，也能自行制定安全标准。特朗普政府前任人工智能事务负责人戴维·萨克斯则在接受CBS新闻采访时指出：“确保产品安全是他们的责任，如果无法做到这一点，那就不应推出相关产品。” 萨克斯的观点是正确的，历史也印证了这一点：1996年时美国政府未能制定出最优政策，如今同样难以做到。
Recent pushback has, thankfully, been bipartisan. Jonathan Kanter, Joe Biden’s antitrust chief, told CNBC frontier labs can adopt safety standards without an antitrust waiver. David Sacks, President Trump’s former AI czar, told CBS News “it is on them to make their products safe,” adding, “If you can’t control it, then don’t do it.”Sacks is right, and history shows why. Washington could not write the best policies in 1996, and it cannot do so now.
上一次具有颠覆性意义的重大技术诞生于曼哈顿计划时期，当时政府掌控着所有原子能实验室、相关技术知识以及使用权限的分配权。而如今的各类人工智能实验室均为私营企业，其估值甚至超过了多数国家的经济总量；在自由市场环境下，政府根本无法掌控这些企业。
The last existential technology came out of the Manhattan Project, where the government owned the atomic labs, the knowledge and the decision about who got access. Today’s AI labs are private and worth more than most economies; no free market would hand them to the state.
众议院已休会，直到中期选举结束，而那些酝酿了数月之久的人工智能相关法案仍迟迟未能通过。此时要求国会来主导相关监管工作，就好比指望一个从未开过车的乘客来防止校车发生碰撞。对于这种人们尚且知之甚少的技术而言，我们需要的是能够随技术发展而不断调整的规则，而非仅仅作为竞选口号或承诺的空话。
The House has gone home until after the midterms with months-old AI bills still stalled. Asking Congress to take the wheel is therefore like asking a passenger who has never driven to keep a school bus from crashing. A technology this poorly understood needs rules that evolve, not a campaign slogan or promise.
这绝不是反对对人工智能实施监管的理由。相反，这是在反对在该技术尚未完全融入社会、其实际作用尚未明确之时就将其监管规则以法律形式固定下来。30年来，《通信规范法》第230条仅被修订过一次；当初为保护少数拨号上网服务而设立的“保护伞”，如今却依然在庇护着那些当初并未存在的企业。
None of this is an argument against regulating AI. It is an argument against carving the rules into statute in the industry’s early years, before its full role in our society has settled. Section 230 has been amended just once in 30 years. The shield written to protect a few dial-up services outlived the companies it was drafted for.
责任豁免或反垄断豁免措施也可能出现类似情况——它们最终只会惠及当前规模最大的企业。那些寻求获得许可以自行协调安全事宜的私营人工智能实验室，实际上也是在寻求垄断性特权；而真正有能力放慢发展步伐的企业，恰恰就是那些处于技术前沿的公司。我们应当尽早启动监管，但制定的规则必须能随技术演进而灵活调整，绝不能赋予企业不可撤销的豁免权。
A liability exemption or antitrust waiver could harden the same way, around whoever is largest now. Private AI labs asking permission to self-coordinate on safety are also asking permission to coordinate, and the only firms that can afford to slow down are already at the frontier. Begin to regulate now but write rules that can change as the technology does. Don’t hand over an immutable immunity.
阿莫迪现年43岁。与1995年时年仅24岁的安德森不同，阿莫迪拥有可供借鉴的历史经验。今年秋季，芝加哥大学禁止在核心社会科学课程中使用人工智能，目的是让学生们重新学会阅读、写作与独立思考。那些经典著作其实能给如今的新机器带来启示：保护机制的存在时间往往远超其最初的保护对象，而后来留下的使用者则必须承受那些被保护下来的后果。
Amodei is 43 years old. Unlike Andreessen, who was 24 in 1995, he has history to draw and learn from. This fall, the University of Chicago barred AI from its core social sciences courses so that students learn to read, write and think. Those old books hold a lesson for these new machines: A shield outlasts whomever it was built to protect, and those who remain must live with whatever was protected.
追求利润就必须承担相应责任。企业不能只把收益据为己有、把损失转嫁给社会，还指望对发展过程中可能出现的所有风险都获得豁免许可。倘若那些处于技术前沿的人工智能实验室真的认为自身技术存在巨大风险，读完本文后他们完全可以自行放慢发展步伐——他们并不需要我们的许可才能踩下“刹车”。
Accountability is the price of pursuing profit. You cannot privatize gains, socialize losses and ask for a permission slip for everything that could go wrong in between. If frontier AI labs believe their technology is that dangerous, they can slow down after reading this article. They don’t need our permission to hit the brakes.
赞德·考恩曾是致力于研发科学级超级智能的人工智能实验室“莉拉科学”的产品部门员工。
Zander Cowan is a former employee in product at Lila Sciences, an AI lab pursuing scientific superintelligence.