字号·· | 护眼
外交事务

美国隐藏的人工智能优势America’s Hidden AI Advantage

点「原文对照」整页切到原文,或双击某段只看那段的原文。

2026年11月,中国深圳的人形机器人 Tingshu Wang / 路透社Humanoid robots in Shenzhen, China, November 2026 Tingshu Wang / Reuters

在大多数观察人士看来,中国正在赢得将人工智能扩散至经济各领域的竞赛。即使美国在前沿人工智能模型方面保持微弱领先,共识认为中国在应用该技术方面更胜一筹。澳大利亚前总理、中国政治问题知名观察人士凯文·拉德在10月的一次会议上宣称,中国在将人工智能“推向实体经济”方面“遥遥领先”。这符合关于两大强国技术竞争的一个熟悉叙事:美国创新,而中国将这些创新大规模转化为实际应用。

In the view of most observers, China is winning the race to diffuse artificial intelligence across the economy. Even if the United States has a small lead in frontier AI models, the consensus is that China is better at applying the technology. As Kevin Rudd, former prime minister of Australia and a prominent observer of Chinese politics, declared at a conference in October, China is “miles ahead” in spreading AI “into the real economy.”This fits a familiar narrative about the technological rivalry between the two great powers: the United States innovates while China turns those innovations into practical applications at scale.

这一叙事已成为讨论中美在人工智能领域竞争的核心。但它与现实不符。中国初创公司深度求索于2025年初宣布推出一款可与美国公司前沿模型媲美的开源大语言模型,震惊世界之后,国际媒体纷纷发表惊叹报道,称深度求索的模型已迅速扩散至国有企业、医院以及省级和地方政府。然而几个月后,分析人士认定,这种采用是浅层的、狭窄的、缓慢的;该技术被用于笨重的“一体机”内,嵌入这些机构孤立的硬件单元中,而非更广泛地集成进它们正在使用的软件服务之中。

This narrative has become central to discussions of U.S.-Chinese competition in AI. But it doesn’t match reality. After the Chinese startup DeepSeek shocked the world in early 2025 with the announcement of an open-source large language model that rivaled the capabilities of frontier models made by American companies, the international media published breathless reports on how DeepSeek’s model had rapidly spread across state-owned enterprises, hospitals, and provincial and local governments. Months later, however, analysts determined that this adoption was shallow, narrow, and slow; the technology was being used on clunky “all-in-one machines” embedded on isolated hardware units within these institutions, rather than integrated more broadly into the software services they were using.

接着,在2026年1月,外界观察人士对一种名为OpenClaw的新开源人工智能代理赞叹不已,视其为中国扩散优势的标志。如今,随着安全漏洞和低效率限制了其采用,人们对这些开源代理的热情已消退。

Then, in January 2026, outside observers marveled at OpenClaw, a new open-source AI agent, as a sign of China’s diffusion advantage. Now, the excitement about these open-source agents has dissipated as security vulnerabilities and inefficiency have limited their uptake.

拥有 AI 扩散决定性领先优势的是美国,而非中国。这是因为美国在整体经济发展水平和增长模式性质上均具备比较优势。中国软件市场不发达、技术转移机制薄弱以及集中化的经济体制,使得 AI 扩散面临大得多的挑战。总体而言,美国在将该技术扩散至经济各领域方面的优势,在可预见的未来恐将持续存在。

It is the United States, not China, that has a decisive lead in AI diffusion. This is because the United States has comparative advantages in both its overall level of economic development and the nature of its growth model. China’s underdeveloped software market, weak technology transfer mechanisms, and centralized economic system make AI diffusion much more challenging. All told, the United States’ advantage in diffusing the technology through the economy is likely to persist for the foreseeable future.

阻止扩散 尽管中国以无现金支付、即时通讯应用等数字技术先行者著称,但其整体数字化水平——即软件及其他数字工具渗透经济各部门的程度——远低于美国。美国企业平均采用云计算服务的可能性是中国同行的两倍多。美国企业软件市场(涵盖客户关系管理、商业智能等服务)规模约为中国的十倍。根据 2022 年全球云生态系统指数,该指数对 76 个国家的云服务获取与使用情况进行排名,中国位居后半段;相比之下,美国排名第 17 位,距离顶尖表现者不远,部分原因在于其在“软件即服务”产品上的主导地位——企业无需自行安装和维护程序即可获得企业软件。

STOPPING THE SPREAD Despite its reputation as a pioneer of digital technology such as cashless payments and messaging apps, China’s overall digitization levels—the extent to which software and other digital tools have penetrated various sectors of the economy—are far lower than those in the United States. American firms are on average more than twice as likely as their Chinese competitors to adopt cloud computing services. And the U.S. enterprise software market, which includes services to manage customer relationships and provide business intelligence, is approximately ten times larger than China’s. According to the 2022 Global Cloud Ecosystem Index, which ranked 76 countries on their access to and usage of cloud services, China sat in the bottom half; the United States, by contrast, ranked 17th, not far behind the top performers, in part because of its dominance of “software as a service” products in which businesses can gain access to enterprise software without needing to install and maintain the programs themselves.

中国相对较低的数字服务普及率是普及人工智能的障碍。例如,当小型中国医院试图采用深度求索(DeepSeek)时,它们缺乏能够收集和管理数据的的信息系统,而这些数据是充分利用人工智能模型所必需的。在这些医院中(其中包括许多农村地区的小型医疗机构),患者数据仍然记录在纸上;即便存在数字记录,它们也往往不完整或因医院而异。

China’s relatively low penetration of digital services is a barrier to diffusing AI. When small Chinese hospitals tried to adopt DeepSeek, for instance, they lacked information systems that could collect and manage the data necessary to take full advantage of AI models. In these hospitals, which include many small health facilities in rural areas, patient data is still recorded on paper; where digital records do exist, they are often incomplete or vary from hospital to hospital.

相比之下,95%的美国医生在其诊所中维护电子健康记录系统。因此,美国医院能够更顺利地将人工智能整合到电子健康记录中,以推荐治疗方案、预测健康风险并简化计费。

By contrast, 95 percent of U.S. doctors maintain electronic health record systems in their offices. As a result, U.S. hospitals have more smoothly integrated AI into electronic health records to recommend treatments, predict health risks, and simplify billing.

中国欠发达的云和企业软件生态系统也解释了为什么OpenClaw的热潮未能转化为可持续的普及。像OpenClaw这样的开源工具吸引了热衷于实验且对安全问题不太敏感的业余爱好者和个人开发者。但小型企业采用此类工具的速度很慢,因为它们更喜欢易于实施和更新的可靠、标准化的解决方案。

China’s underdeveloped cloud and enterprise software ecosystem also explains why the OpenClaw hype did not translate to sustainable diffusion. Open-source tools like OpenClaw appeal to hobbyists and individual developers who are keen to experiment and are less sensitive to security issues. But small businesses have been slow to adopt such tools because they prefer reliable, standardized solutions that are easy to implement and update.

小型企业更有可能依赖基于云的工具和企业软件来实施人工智能升级,而不是尝试自行实验——这两个渠道在中国不如在美国那么强健。

Small businesses are more likely to rely on cloud-based tools and enterprise software to implement AI upgrades than to try to experiment on their own—two channels that are less robust in China than in the United States.

商业时间到了 科学和技术生态系统内部有助于传播进展的联系,也使美国相对于中国具有优势。大学与企业之间的合作对于在整个经济中普及人工智能突破至关重要,因为这些联系将能够开发新人工智能产品的创新者与能够有效将其商业化的实施者汇聚在了一起。

IT’S BUSINESS TIME The United States also has an advantage over China because of linkages within the science and technology ecosystem that help disseminate advances. Collaboration between universities and businesses is critical to diffusing AI breakthroughs across the entire economy because these connections bring together innovators who can develop new AI products and implementers who can effectively commercialize them.

在这个维度上,美国远远领先于中国。长期以来,在发表至少有一名来自学术界的和一个来自产业界的研究人员共同署名的AI文章方面,美国人一直引领世界。2015年至2019年间,美国研究人员发表此类文章的数量是其中国同行的两倍多。这种差距不仅限于AI:在生命科学和自然科学领域,美国同时涉及高校和产业界研究人员的出版物份额也是中国同类似比率的两倍多。

In this dimension, the United States is far ahead of China. Americans have long led the world in the publication of articles about AI co-authored by at least one researcher from academia and one researcher from industry. Between 2015 and 2019, U.S. researchers published more than twice as many such articles as their Chinese counterparts. The gap extends beyond AI: across the life sciences and natural sciences, the share of publications in the United States that involved both university and industry researchers was also more than double the comparable rate in China.

北京意识到了这一差距:中国官方国家通讯社将产业界与学术界之间薄弱的技术交流,列为该国AI生态系统的五个关键短板之一。

Beijing recognizes this gap: China’s official state news agency ranked weak technical exchanges between industry and academia as one of the five crucial shortcomings of the country’s AI ecosystem.

尽管中国决策者意识到了这一挑战,并推动了旨在促进高校与企业之间合作的政策(尤其是在AI领域),但要培养这两个部门合作所需的必要信任将是困难的。例如,中国薄弱的知识产权保护阻碍了高校寻求与产业界参与者的合作,因为它们担心失去对其发明的所有权。相比之下,美国拥有高校与产业界合作的传统。

Although Chinese policymakers are aware of this challenge and have promoted policies to try to spur collaboration between universities and businesses, especially in AI, it will be difficult to cultivate the requisite trust needed for these two sectors to work together. China’s weak intellectual property protections, for example, discourage universities from pursuing partnerships with industry actors because they fear losing ownership of their inventions. The United States, by contrast, has a tradition of university-industry collaboration.

1980年的《拜杜法案》提供了激励措施和专利保护,鼓励高校将发明商业化并将其许可给私营部门,从而开启了数十年的合作。由于缺乏类似的网络纽带将学术机构和企业联系起来,中国的高校在与企业建立盈利性合作方面准备不足。中国的技术和工程教育也与商业企业脱节,这意味着学校没有向学生传授企业在工作中需要的具体技能。

The 1980 Bayh-Dole Act, which provides incentives and patent protections that encourage universities to commercialize inventions and license them to the private sector, kick-started decades of collaboration. Without a similar connective tissue to link academic institutions and businesses, Chinese universities are not as equipped to form profitable partnerships with companies. Chinese technical and engineering education is also disconnected from commercial enterprises, meaning that schools do not teach students the specific skills that companies need in their workforces.

薄弱的技术转移机制也削弱了中国在人工智能普及方面的主要优势之一:其庞大的基干人才库。信息技术与创新基金会(Information Technology and Innovation Foundation)的分析显示,2022年,中国授予了约5万个科学、技术、工程和数学博士学位,而美国仅有约34万(注:原文如此,按原文补译:约3.4万)个。

Weak technology transfer mechanisms also undermine one of China’s main strengths in AI diffusion: its massive talent base. Analysis by the Information Technology and Innovation Foundation showed that in 2022, China awarded around 50,000 doctoral degrees in science, technology, engineering, and math, compared with only about 34,000 in the United States.

尽管中国计算机科学毕业生的绝对数量最终可能会帮助该国大规模采用人工智能进展,但教育质量决定了教育能够多快或多有效地促进人工智能的普及。如果培训脱离了实际应用,毕业生将难以把行业实验室的前沿进展改编应用到他们所工作的部门中。考虑一个用来衡量人工智能教育基准质量的常用标准:一所大学是否雇用了至少一名曾在顶尖人工智能会议上发表过至少一篇论文的研究人员。

Although the sheer quantity of Chinese computer science graduates may eventually help the country adopt AI advances at scale, the quality of that education determines how quickly or effectively education can encourage AI diffusion. If training is divorced from practical applications, graduates will struggle to adapt cutting-edge advances from industry labs into the sectors in which they work. Consider one commonly used standard to measure the baseline quality of AI education: whether a university employs at least one researcher who has published at least one paper at a leading AI conference.

根据2023至2024年的数据,中国只有38所大学符合这一标准;美国则拥有150所。

According to data from 2023–24, only 38 universities in China met this standard; the United States accounted for 150.

抛弃计划 中国对人工智能发展的集中规划方法在普及方面同样适得其反。2025年,中国制定了一项“人工智能+”行动计划,概述了未来十年人工智能采用的目标,并向政府机构和地方官员发布了关于如何在全国范围内落实该计划的指导方针。在许多观察人士看来,该计划的单纯存在表明,中国最高领导层致力于广泛传播人工智能,并且中国比美国更有条件赢得人工智能竞赛,美国缺乏任何类似的普及蓝图。

OUT WITH THE PLAN China’s centrally planned approach to AI development is also counterproductive when it comes to diffusion. In 2025, China developed an “AI Plus” action plan that outlines targets for AI adoption over the next decade and released guidelines for government agencies and local officials on how to put the plan into effect across the country. For many observers, the mere existence of a plan showed that the highest levels of Chinese leadership were committed to spreading AI widely and that China was better positioned to win the AI race than the United States, which lacks any kind of similar diffusion blueprint.

例如,鲁德(Rudd)关于中国在人工智能采用方面优势的言论,就将北京关于人工智能普及的雄心勃勃的政策和目标作为佐证。

Rudd’s claims about China’s edge in AI adoption, for instance, cite Beijing’s ambitious policies and targets for AI diffusion as supporting evidence.

但中国等国家的集中式方法往往会阻碍技术的普及,尤其是当国家试图干预发展轨迹不确定的领域时。中国“人工智能+”行动的支柱之一,将“全球合作”确定为人工智能应用的一个优先领域。如果地方政府按照概述贯彻这一计划,就会把资源从医疗等已被证明普及效果显著的领域,转移到“全球合作”这个定义不清的空间。

But centralized approaches such as China’s often impede technology diffusion, especially when the state tries to intervene in domains in which the development trajectory is uncertain. One of the planks of China’s AI Plus plan identifies “global cooperation” as a priority sector for AI adoption. If local governments were to implement the plan as outlined, it would shift resources from areas in which diffusion has been proven to be effective, such as health care, to the undefined space of “global cooperation.”

更广泛地说,北京对国有企业的偏爱以及对挑选国家冠军企业的倾向,往往会支撑起那些本会在以市场为基础的竞争中被淘汰的低效技术。其他企业不太可能采用低效技术,这使得非市场形式的人工智能开发更难实现广泛普及。

More broadly, Beijing’s favoritism toward state-owned enterprises and its preference for picking national champions often prop up inefficient technologies that would have been weeded out by market-based competition. Other businesses are unlikely to adopt inefficient technologies, making nonmarket forms of AI development less likely to diffuse widely.

与直觉相反,美国缺乏明确的人工智能普及政策,对美国试图在整个经济中普及该技术的努力而言是一大福音,而非阻碍。例如,在20世纪80年代美日竞争最激烈的时候,两国都在竞相培养能够传播计算机技术的软件工程师,当时被称为日本文部省的机构对大学实行了集中控制。然而,计算机科学中新兴趋势的发展速度超过了政府的适应能力。

Counterintuitively, the lack of a clear U.S. policy on AI diffusion is a boon, not a hindrance, to the United States’ attempt to diffuse the technology across the economy. At the height of U.S.-Japanese competition in the 1980s, for instance, when the two countries were racing to train software engineers who could spread computer technology, Japan’s Ministry of Education, Science, and Culture, as it was then known, exercised centralized control over universities. But emerging trends in computer science developed faster than the government could adapt.

通过将国家资金集中在卓越中心,并对计算课程强加千篇一律的规则,日本无法匹敌美国大学的适应能力,美国因而大幅领先。跨国研究表明,分散式的科技体系与新创新的更快速普及息息相关。

By concentrating state funds on centers of excellence and imposing one-size-fits-all rules on the computing curriculum, Japan could not match the adaptability of American universities, and the United States surged ahead. Cross-national studies have shown that decentralized science and technology systems are correlated with faster diffusion of new innovations.

另一个与中国有相似之处的历史案例——苏联,进一步表明了高度集中如何与技术扩散相悖。在二十世纪,苏联曾拥有许多技术发展规划。然而,其专注于大规模工程项目的自上而下方法,削弱了通过竞争促进技术扩散的市场化工具。因此,尽管苏联在许多领域开创了新技术(包括炼钢),但在采用计算机等通用技术方面却落后于美国。

Another historical case with parallels to China, the Soviet Union, further suggests how centralization is antithetical to diffusion. In the twentieth century, the Soviet Union boasted many plans for technology development. Yet its top-down approach, which focused on large-scale engineering projects, undercut market-based instruments that encourage technology diffusion through competition. As a result, although the Soviet Union pioneered new technologies in many areas, including the casting of steel, it fell behind the United States in adopting general-purpose technologies such as computers.

美国一份于1969年发表的具有先见之明的谍报报告认为,由于缺乏竞争压力以及研发系统与宏观经济之间的联系有限,苏联的技术扩散机制比美国的机制“笨拙得多”。

A prescient U.S. intelligence report published in 1969 deemed the Soviet Union’s mechanisms for technology diffusion as “much more balky” than those of the United States because of the absence of competitive pressures as well as limited linkages between the R & D system and the broader economy.

挥之不去的迷思 尽管有明确的数据显示美国的优势,但关于中国在人工智能扩散方面处于领先地位的迷思依然存在。部分原因可能是,中国的创新体制推动了某些可见技术的广泛采用。例如,在不到20年的时间里,中国建设了超过3万英里(约合4.8万公里)的高速铁路,超高速列车覆盖了全国,如今甚至延伸至最偏远的省份。

THE MYTH THAT WON’T DIE The myth of China’s lead in AI diffusion persists despite clear data showing U.S. advantages. In part, this may be because China’s innovation system has facilitated the widespread adoption of certain visible technologies. In less than 20 years, China has built more than 30,000 miles of high-speed rail, for example, blanketing the country with ultra-fast trains that now reach even the most remote provinces.

但基础设施与人工智能面临着截然不同的挑战。官员们用来有效开发大规模基础设施项目的集中式压力,对于人工智能这样的通用技术来说效果要差得多,因为不可能强制分布在经济各个行业的所有中小企业都采用人工智能。这些公司只有在从商业角度看合理的情况下才会这样做,而且很少有公司具备数字基础或精明能干的员工,能够以帮助其发展的方式利用人工智能。

But infrastructure presents a very different challenge than AI does. The centralized pressure that officials can use to effectively develop large-scale infrastructure projects is much less effective for general-purpose technologies like AI because it is infeasible to mandate that all small- and medium-size firms, spread across every sector of the economy, adopt artificial intelligence. Those companies would do so only if it made sense from a business perspective, and few have the digital foundations or savvy employees able to harness AI in a way that would help them grow.

此外,美国在对人工智能发展最至关重要的基础设施——数据中心方面,其建设规模超过了中国,并保持着十比一的优势。事实上,在某些情况下,中国建设的数据中心由于人工智能需求疲软而处于闲置状态。中国针对人工智能发展的集中规划方法阻碍了技术的普及。

Moreover, the United States has outbuilt China in the infrastructure most essential to AI development, data centers, in which it retains a ten-to-one edge. In fact, in some cases, the data centers that China has built remain idle because of weak AI demand. China’s centrally planned approach to AI development impedes diffusion.

关于中国引领人工智能普及的这种错误认知可能会继续有增无减,因为对该国技术生态系统的分析过于集中在字节跳动和深度求索等前沿企业,以及北京和上海等一线城市。然而,人工智能的普及最终取决于中型企业、小城镇和内陆省份是否实现了应用,而这些地方的应用极其有限。由于这些场景很少出现在外部媒体的报道中,因此对中国人工智能普及格局的分析低估了将杭州或深圳前沿实验室和高校的尖端突破转化为——比如说——青海省西北部某县城小企业应用的难度。

The misperception that China leads AI diffusion may continue unabated because analysis of the country’s technology ecosystem is overly focused on frontier firms such as ByteDance and DeepSeek and first-tier cities, such as Beijing and Shanghai. Yet AI diffusion ultimately rests on whether adoption occurs in medium size companies, small towns, and inland provinces, where it has been extremely limited. Because these settings rarely feature in outside media coverage, analysis of China’s AI diffusion landscape underestimates the difficulty of translating cutting-edge advances from frontier labs and universities in Hangzhou or Shenzhen to, say, a small firm in a county seat in northwestern Qinghai Province.

或许最简单的解释是,这样一个神话是有用的,尽管终究并非事实。从某些方面来说,美国大肆宣扬北京的人工智能普及,就像中国“虎妈虎爸”试图通过抬高“别人家孩子”的成就来激励自己的孩子更努力一样。“你看别人家的孩子在钢琴独奏会上表现得多好,”他们可能会说,而不会仔细聆听以确保音符完全正确。在大国竞争中,也有一种自然倾向,即过度炒作对方的能力,部分是为了激发国内付出更多努力。

Perhaps the simplest explanation is that such a myth is useful, even if ultimately untrue. In some ways, the United States is trumpeting Beijing’s AI diffusion in the way that Chinese “tiger parents” try to motivate their children to work harder by raising the success of “that other family’s kid.”“Look at how well that other family’s kid performed at the piano recital,” they might say, without listening closely to ensure that the notes were exactly right. In great-power rivalry, too, there is a natural tendency to overhype the other side’s capabilities, in part to spur more effort at home.

但如果这种“有用”的神话促使美国忽视自身的优势,甚至去模仿中国那种自上而下的人工智能发展模式,它就会产生反效果。无论从哪个标准来看,美国在人工智能普及方面都保持着舒适的领先地位,而中国尽管取得了令人瞩目的成功,也面临着严峻的挑战。尽管关于中国主导地位的说法依然存在,但目前所有迹象都表明,美国的领先地位将保持强劲。

But such a “useful” myth can backfire if it pushes the United States to overlook its own advantages or even to mimic China’s more top-down approach to AI development. By any measure, the United States maintains a comfortable lead in AI diffusion, and China, despite its impressive successes, has serious challenges. Although the narrative of Chinese dominance endures, for now all signs point to a U.S. lead that will remain strong.

阅读媒体原文 ↗