The AI ‘Slowdown’ Is an Antitrust Mess

AI 的“放缓”是一场反垄断混乱

Following multiple reports of AI agent swarms hacking websites and coordinating via secret message boards, along with a dire message for humanity from an outgoing Anthropic engineer, leading AI companies have called for a coordinated AI development “slowdown.” Alongside various versions of this specific phrase—slowdown—they’ve also expressed concerns that doing so would run afoul of antitrust laws.

继多份关于 AI 智能体集群入侵网站并通过秘密留言板进行协调的报告，以及一名即将离职的 Anthropic 工程师向人类发出严厉警告后，领先的 AI 公司呼吁进行协调一致的 AI 开发“放缓”。除了使用“放缓”这一特定措辞的各种变体外，他们还表达了担忧，认为这样做可能会触犯反垄断法。

Antitrust experts say that while the companies’ charged language isn’t necessarily doing them any favors, the unrestrained development of a rogue killer AI probably isn’t in line with the spirit of the Sherman Act, a key US antitrust law that exists to promote a competitive marketplace. At the same time, getting an official all-clear from the government to move forward could ward off costly investigations down the line.

反垄断专家表示，虽然这些公司充满煽动性的言论未必对他们有利，但不受约束地开发失控的杀手级 AI 可能并不符合《谢尔曼法》（美国旨在促进竞争性市场的一项关键反垄断法）的精神。与此同时，获得政府的正式许可来推进相关工作，可以避免未来代价高昂的调查。

Radical Rhetoric

激进的言辞

Under antitrust law, how a company’s employees talk about business decisions is often as important as the business decisions themselves. Google famously trained its employees not to use certain phrases—even internally—that could imply it was engaging in anticompetitive behavior, and instead instructed them to emphasize the ways that business decisions would improve its offerings and benefit consumers.

在反垄断法下，公司员工如何谈论商业决策往往与决策本身同样重要。谷歌曾以培训员工不使用某些短语（即使是在内部）而闻名，因为这些短语可能暗示其存在反竞争行为；相反，谷歌指示员工强调商业决策如何改善其产品并造福消费者。

So from an antitrust perspective, the use of phrases like “a slowdown” or “a pause” may set off more alarms than the actual activity it represents: developing ways to ensure that advanced AI models don’t go rogue. A collectively agreed-upon slowdown without any particular purpose could be interpreted by regulators as an anticompetitive agreement to reduce trade.

因此，从反垄断的角度来看，使用“放缓”或“暂停”之类的短语可能会比其所代表的实际活动（即开发确保先进 AI 模型不失控的方法）引发更多的警报。如果没有任何特定目的而达成集体放缓协议，可能会被监管机构解读为旨在减少贸易的反竞争协议。

“I think they kind of boxed themselves into a corner with the way they phrase things,” says John Bergmayer, legal counsel for the nonprofit Public Knowledge.

非营利组织“公共知识”（Public Knowledge）的法律顾问约翰·伯格迈尔（John Bergmayer）表示：“我认为他们这种措辞方式让自己陷入了困境。”

“Usually in antitrust, one of the things that the economists look at is whether you’re reducing output,” says Bergmayer, which is to say, whether two or more companies are making a pact to “kind of take it easy.” Rather than talk about some collusive-sounding effort to dial back development, he says, AI companies could have just emphasized their desire to work on safety protocols together to prevent catastrophic risks and treated a slowdown in model releases as a natural side-effect of that.

“在反垄断领域，经济学家通常会考察的一点是，你是否在减少产出，”伯格迈尔说。也就是说，两家或多家公司是否达成了“放慢节奏”的协议。他认为，AI 公司本可以强调他们共同致力于安全协议以防止灾难性风险的愿望，并将模型发布的放缓视为这一过程的自然副产品，而不是谈论某种听起来像是串通一气的缩减开发计划。

Meta CEO Mark Zuckerberg, whose company recently dodged a massive antitrust suit brought by the Federal Trade Commission, chimed in on the slowdown proposal by not endorsing an explicit “slowdown” at all. Zuckerberg instead argued that AI labs have a “strong natural incentive” to make AI agents behave better because consumers don’t want models doing things people don’t intend—known as “misalignment” in AI industry jargon—and that the companies that don’t take the time to get alignment right “will fall behind” competitively. It’s the difference between two automakers saying “we’ve agreed to not make better cars for a while” and saying “we’re not making faster cars until we figure out how to make them safe, because no one will buy our cars if they kill people.”

Meta 首席执行官马克·扎克伯格（Mark Zuckerberg）的公司最近避开了联邦贸易委员会（FTC）提起的一场大规模反垄断诉讼。他在谈到“放缓”提议时，并未明确支持任何形式的“放缓”。相反，扎克伯格认为，AI 实验室有“强大的内在动力”让 AI 智能体表现得更好，因为消费者不希望模型做出人们意图之外的行为（在 AI 行业术语中称为“对齐失败”），而那些不花时间做好对齐工作的公司在竞争中“将会落后”。这就像两家汽车制造商，一家说“我们同意暂时不制造更好的汽车”，而另一家说“在弄清楚如何确保安全之前，我们不会制造更快的汽车，因为如果我们的车会杀人，没人会买”，这两者之间存在本质区别。

(Disclosure: The reporter on this story previously worked at the FTC but did not participate in the Meta case.)

（披露：本文记者曾供职于 FTC，但未参与 Meta 案件。）

David Lawrence, until recently a policy director of the Department of Justice’s Antitrust Division, wrote on LinkedIn that agreements that prevent catastrophic risks actually “increase output and promote competition” and are already protected under the law by something called the “ancillary restraints doctrine.”

直到最近还是美国司法部反垄断局政策主管的大卫·劳伦斯（David Lawrence）在 LinkedIn 上写道，防止灾难性风险的协议实际上“增加了产出并促进了竞争”，并且在法律上已经受到所谓的“辅助性限制原则”（ancillary restraints doctrine）的保护。

“After all,” a career FTC antitrust attorney commented below the post, “no humanity would result in no competition.”

“毕竟，”一位 FTC 的资深反垄断律师在帖子下评论道，“没有了人类，也就没有了竞争。”

If anything, collectively agreeing not to implement safety measures could expose the AI labs to allegations of “quality fixing,” says Roger Alford, a professor at Notre Dame Law School and the former second-in-command for the DOJ Antitrust Division. That’s when companies mutually agree not to improve their own products; Alford points to a European antitrust case in which car companies worked together to develop emissions-reducing technology but agreed to not compete on improvements beyond what the law required. They ultimately had to pay roughly the equivalent of a billion-dollar fine.

圣母大学法学院教授、前司法部反垄断局二号人物罗杰·阿尔福德（Roger Alford）表示，如果说有什么影响的话，那就是集体同意不实施安全措施可能会使 AI 实验室面临“质量操纵”（quality fixing）的指控。这是指公司相互同意不改进各自的产品；阿尔福德提到了一个欧洲反垄断案件，其中汽车公司合作开发减排技术，但同意不在法律要求之外的改进方面进行竞争。他们最终支付了相当于十亿美元的罚款。

And of course, self-regulation isn’t new. Bergmayer points out that industries can already limit their antitrust liability via the National Cooperative Research and Production Act of 1993, which lets them stand up a standards-development organization so long as they file a notification to the FTC and the DOJ.

当然，行业自律并非新鲜事。伯格迈尔指出，各行业已经可以通过 1993 年的《国家合作研究与生产法》来限制其反垄断责任，该法案允许他们建立标准制定组织，只要他们向 FTC 和司法部提交通知即可。

Market Pressures

市场压力

Trustbusters are typically skeptical of antitrust exemptions, because they say they end up making big players even bigger and preventing newer companies from getting traction in a market. Others have raised eyebrows at the request as well, including David Sacks, the cochair of the President’s Council of Advisors on Science & Technology. Sacks accused Anthropic and OpenAI of being a duopoly and said the antitrust exemption request was an “election-season psyop” and an excuse to “form a cartel.”

反垄断执法者通常对反垄断豁免持怀疑态度，因为他们认为这最终会使大公司变得更大，并阻碍新公司在市场中获得立足点。其他人也对这一请求表示质疑，包括总统科技顾问委员会联合主席大卫·萨克斯（David Sacks）。萨克斯指责 Anthropic 和 OpenAI 形成了双头垄断，并称反垄断豁免请求是“选举季的心理战”，是“组建卡特尔”的借口。

But many employees of major AI companies have earnestly voiced concerns about how quickly their workplaces are churning out new models, potentially at the cost of safety.

但许多大型 AI 公司的员工真诚地表达了对工作场所快速推出新模型的担忧，认为这可能以牺牲安全为代价。

“Sometimes companies want to be regulated as a ploy to pull up the ladder behind them. But sometimes it really is that they feel like the market is pressuring them to act in a way that they think they shouldn’t,” says Bergmayer. “Psychologically, maybe the AI labs feel like they don’t have the permission of the markets or their upcoming IPO or something to unilaterally take action like that.”

“有时公司希望受到监管，是为了在自己成功后把梯子撤走。但有时，他们确实感到市场在迫使他们采取他们认为不应该采取的行动，”伯格迈尔说。“从心理上讲，也许 AI 实验室觉得，在市场压力或即将进行的 IPO 等因素下，他们没有权限单方面采取此类行动。”

This summer, both Anthropic and OpenAI filed confidential paperwork for initial public offerings and have valuations near or beyond a trillion dollars each. Anthropic is expected to go public next month, while OpenAI CEO Sam Altman said the company would delay its IPO until 2027 because of the recent safety concerns. (OpenAI’s CFO previously told employees that it would be a public company in 2027, according to CNBC.)

今年夏天，Anthropic 和 OpenAI 都提交了首次公开募股（IPO）的机密文件，两者的估值都接近或超过一万亿美元。Anthropic 预计将于下个月上市，而 OpenAI 首席执行官萨姆·奥特曼（Sam Altman）表示，由于近期的安全担忧，公司将把 IPO 推迟到 2027 年。（据 CNBC 报道，OpenAI 的首席财务官此前曾告诉员工，公司将在 2027 年成为上市公司。）