The United Kingdom’s Military AI Ambitions Need Guardrails

Advancing military AI capability is a priority for the United Kingdom, and in recent weeks, the new government has doubled down on its ambitions for AI and defense via two major partnerships.

推进军事人工智能能力是英国的首要任务，最近几周，新政府通过两个主要合作伙伴关系加倍了其在人工智能和国防方面的雄心。

On Aug.

八月。

24, 2026, on his first overseas trip as the new prime minister of the United Kingdom, Andy Burnham signed a landmark partnership : The United Kingdom is the first international partner to receive access to Ukraine’s Avengers AI Labs.

2026 年 2 月 24 日，安迪·伯纳姆 (Andy Burnham) 作为英国新首相首次出访，签署了一项具有里程碑意义的合作伙伴关系：英国是第一个获得乌克兰复仇者联盟人工智能实验室使用权的国际合作伙伴。

The deal will allow the United Kingdom to train its military AI systems with battlefield data, which give crucial insights into warfighting in the 21st century.

该协议将使英国能够利用战场数据训练其军事人工智能系统，从而为 21 世纪的战争提供重要的见解。

In return, the United Kingdom will back Ukraine with the world’s third-largest AI ecosystem , as well as leading research and development for Ukraine to utilize for military technologies and solutions.

作为回报，英国将向乌克兰提供世界第三大人工智能生态系统的支持，并为乌克兰提供领先的研究和开发，以用于军事技术和解决方案。

On Sept.

九月。

22, Burnham announced a new AI and autonomy partnership with the United States, aimed at increasing military interoperability, using AI for defense, and protecting critical infrastructure.

2 月 22 日，伯纳姆宣布与美国建立新的人工智能和自主合作伙伴关系，旨在提高军事互操作性，利用人工智能进行防御，并保护关键基础设施。

The new government aspires to build a globally competitive edge in military AI.

新政府渴望在军事人工智能领域建立全球竞争优势。

There are windows of opportunity for doing so responsibly, but the country’s position is complicated.

负责任地这样做是有机会的，但该国的立场很复杂。

It relies on U.S. technology and partnership, and that reliance will continue.

它依赖美国的技术和合作伙伴关系，而且这种依赖将继续下去。

That reliance churns up fresh risks.

这种依赖会引发新的风险。

“Kill switch” anxieties circulate in Westminster, and some British policymakers are concerned that the country is deepening dependency on an untrustworthy ally.

“终止开关”的焦虑在威斯敏斯特蔓延，一些英国政策制定者担心该国正在加深对一个不值得信任的盟友的依赖。

If the new government does not act to mitigate risk, invest in security and safety guardrails such as mandatory testing and transparency around implementation, and shape governance efforts, rapid investment in military AI may undermine the security it is supposed to deliver.

如果新政府不采取行动降低风险，投资于安全和安全护栏，例如强制测试和实施透明度，以及塑造治理工作，那么对军事人工智能的快速投资可能会破坏其应提供的安全性。

The United Kingdom is one of many countries racing to invest in military AI.

英国是众多竞相投资军事人工智能的国家之一。

Members of NATO are scrambling to meet the target of 3.5 percent of gross domestic product on core defense and 1.5 percent on security-related spending by 2035 and boost their collective security.

北约成员国正在努力实现到2035年将国内生产总值的3.5%用于核心国防和1.5%的安全相关支出的目标，并加强其集体安全。

Many decision-makers see AI as crucial for competitiveness, resilience, efficiency, and lethality.

许多决策者认为人工智能对于竞争力、弹性、效率和杀伤力至关重要。

In the United Kingdom alone, the new Defence Investment Plan allocates over $6.6 billion to autonomous systems by 2030.

仅在英国，新的国防投资计划到 2030 年就会为自主系统拨款超过 66 亿美元。

The wars in Ukraine and Iran have highlighted the operational advantages of AI, particularly for target selection and some automated uncrewed systems.

乌克兰和伊朗的战争凸显了人工智能的作战优势，特别是在目标选择和一些自动化无人系统方面。

These speed up decision cycles through aggregating and analyzing massive amounts of data and presenting options for action and free up human capacity for more complex operations.

这些通过聚合和分析大量数据并提供行动选项来加快决策周期，并释放人力以进行更复杂的操作。

But they also demonstrate that warfare is becoming more asymmetric and volatile.

但它们也表明战争正变得更加不对称和不稳定。

Advantage is not guaranteed to the military with the most expensive and sophisticated systems: Far smaller militaries can hold their own using cheap, precise systems utilizing AI.

拥有最昂贵和最复杂系统的军队并不能保证获得优势：规模小得多的军队可以使用廉价、精确的人工智能系统来保持自己的优势。

Alongside the advantages, however, the grave costs of rapid integration are showing.

然而，除了这些优点之外，快速集成的巨大成本也逐渐显现出来。

Israel is accused by human rights organizations like Amnesty International and an independent United Nations commission of knowingly causing thousands of civilian deaths, in part through reported use of a “ Lavender ” AI-assisted targeting system in Gaza.

国际特赦组织和联合国独立委员会等人权组织指责以色列故意造成数千名平民死亡，部分原因是据报道在加沙使用了“薰衣草”人工智能辅助瞄准系统。

These allegations, which Israel denies, have resurfaced in the newly released NAZA documentary .

这些以色列否认的指控在新发布的 NAZA 纪录片中再次出现。

In July 2026, a fully autonomous Russian drone killed three Ukrainian civilians by a fully autonomous Russian drone, in reportedly the first documented case of civilian deaths in Ukraine by a weapon with no human involvement.

2026 年 7 月，一架完全自主的俄罗斯无人机杀死了三名乌克兰平民，据报道，这是乌克兰第一例有记录的无人参与的武器造成平民死亡的案件。

In the Iran war, the American and Israeli militaries are using AI for target identification and selection.

在伊朗战争中，美国和以色列军队正在使用人工智能进行目标识别和选择。

A U.S. strike on a school is alleged to have relied on data that was out of date.

据称，美国对一所学校的罢工依赖的是过时的数据。

An AI algorithm reportedly identified the target, and human oversight failed to catch the error.

据报道，人工智能算法识别了目标，但人类的监督未能发现错误。

The United Nations Human Rights Council’s independent fact-finding mission on Iran has “reasonable grounds to believe” this is a potential war crime .

联合国人权理事会对伊朗的独立实况调查团有“合理的理由相信”这是潜在的战争罪。

Beyond key conflict theaters, Anthropic’s recent threat report highlights how a range of actors are using AI to amplify their subthreshold activities, including cyber attacks, weapon development, and influence operations.

除了主要的冲突战场之外，Anthropic 最近的威胁报告还强调了一系列参与者如何利用人工智能来扩大他们的阈下活动，包括网络攻击、武器开发和影响行动。

Concerningly, these actors are improving and gaining new capabilities using widely available AI models.

令人担忧的是，这些参与者正在使用广泛使用的人工智能模型来改进并获得新的能力。

The United Kingdom is at a crossroads.

英国正处于十字路口。

There is strong pressure to invest in military AI and the strategic advantages of doing so.

投资军事人工智能面临着巨大的压力，而且这样做也具有战略优势。

The world urgently needs international cooperation to beat the rising tide of AI-related harms, develop safeguards, and improve reliability.

世界迫切需要国际合作来应对不断上升的人工智能相关危害、制定保障措施并提高可靠性。

Although governance efforts are ongoing, they have not resulted in meaningful regulation or developed common standards for the use of AI in the military.

尽管治理工作正在进行中，但尚未产生有意义的监管或制定军事中使用人工智能的通用标准。

Without movement on norms and regulation — to mitigate risks and build transparency on how AI is integrated — harm will proliferate.

如果规范和监管不采取行动——以降低风险并提高人工智能集成方式的透明度——危害将会激增。

Dark Infrastructure Different building blocks determine a competitive edge in AI — including its applications in the military.

黑暗基础设施 不同的构建模块决定了人工智能的竞争优势——包括其在军事中的应用。

Beyond the United States and China, most countries do not have fully sovereign control over these primary inputs, namely data, computing power, and frontier models.

除美国和中国之外，大多数国家对这些主要投入（即数据、计算能力和前沿模型）没有完全的主权控制。

Recognizing how central AI is for national security, most countries aspire to control this infrastructure, but few have the agency or resources.

认识到人工智能对于国家安全的重要性，大多数国家都渴望控制这一基础设施，但很少有国家拥有机构或资源。

However, agency is not determined by computing power, data, and models alone: It can be at least partially determined by talent, industrial policy, trust, and diffusion.

然而，代理并不仅仅由计算能力、数据和模型决定：它至少可以部分地由人才、产业政策、信任和扩散决定。

For middle powers like the United Kingdom — largely without sufficient sovereign computing power or guaranteed access to the most cutting-edge models and data — this presents a dependency dilemma.

对于像英国这样的中等强国来说——基本上没有足够的主权计算能力或无法保证获得最前沿的模型和数据——这带来了依赖困境。

To support its military AI buildout, the United Kingdom needs partnerships with foreign technology providers and allies, despite its existing strengths in talent and research and development.

为了支持其军事人工智能建设，尽管英国在人才和研发方面现有优势，但仍需要与外国技术提供商和盟友建立伙伴关系。

Partnership grants the country benefits, such as priority access to the world’s most powerful frontier models, and may even boost its national security.

伙伴关系会给国家带来好处，例如优先使用世界上最强大的前沿模型，甚至可能增强国家安全。

But relying on foreign partners for the infrastructure underlying military AI capability presents two types of vulnerability: risks arising from operational reliance on non-sovereign military technology, including accidental disruptions and intentional “kill switch” activities, and risks arising from a lack of information about or access to frontier models for evaluation.

但依赖外国合作伙伴提供军事人工智能能力的基础设施存在两种类型的脆弱性：因操作上依赖非主权军事技术而产生的风险，包括意外中断和故意的“终止开关”活动，以及因缺乏有关评估的前沿模型的信息或访问权限而产生的风险。

A leading concern for policymakers is the lack of British leverage over frontier labs across the Atlantic.

政策制定者最关心的一个问题是英国对大西洋彼岸的前沿实验室缺乏影响力。

Currently, access (for example, for the country’s AI evaluators to test models) relies on trust and goodwill.

目前，访问（例如，该国的人工智能评估人员测试模型）依赖于信任和善意。

Both are in short supply, reigniting concerns about a U.S.-mediated “ kill switch ,” losing access to frontier capabilities actively used in a military context, and the dearth of information about risks that models and systems may generate.

两者都供不应求，重新引发了人们对美国介导的“死亡开关”的担忧，失去了在军事环境中积极使用的前沿能力，以及缺乏有关模型和系统可能产生的风险的信息。

This is a live concern for many allies of the United States, with European countries like France pursuing sovereign AI and cloud computing capabilities, and states prioritizing national security over the larger scale and abilities that American companies deliver.

这是美国许多盟友的现实担忧，法国等欧洲国家追求主权人工智能和云计算能力，而各国则将国家安全置于美国公司提供的更大规模和能力之上。

Lack of Guardrails Global anxiety about AI is on the rise, with reports of unpredictable and deceptive model behavior from leading labs.

缺乏护栏 全球对人工智能的焦虑正在加剧，来自领先实验室的报告显示模型行为不可预测且具有欺骗性。

In September, Australian Prime Minister Anthony Albanese revealed that OpenAI agents reportedly hacked an Australian government website.

9 月，澳大利亚总理 Anthony Albanese 透露，据报道 OpenAI 特工入侵了澳大利亚政府网站。

OpenAI agents also reportedly escaped sandboxes to access data from websites like the FBI’s Crime Data Explorer.

据报道，OpenAI 特工还逃离了沙箱，访问了 FBI 犯罪数据浏览器等网站的数据。

There are few — arguably, if any — globally actionable technical interventions to clamp down on rogue models causing real-world harm.

可以说，全球范围内可采取行动的技术干预措施很少（如果有的话）来打击造成现实世界伤害的流氓模型。

This major risk is only amplified in conflict, where informational and operational limitations make threat identification technically tricky.

这种主要风险只会在冲突中放大，因为信息和操作的限制使得威胁识别在技术上变得棘手。

From a technical perspective, there is no automatic binary between the use of AI in military and civilian settings, as evidenced by the widespread use of general-purpose models for gathering threat intelligence and informing targeting decisions.

从技术角度来看，人工智能在军事和民用环境中的使用之间不存在自动二元关系，广泛使用通用模型来收集威胁情报和通知目标决策就证明了这一点。

The U.S. military reportedly used Anthropic’s Claude to support military action in Venezuela, leading to a furor.

据报道，美国军方利用 Anthropic 的克劳德支持在委内瑞拉的军事行动，引发了轩然大波。

This creates a three-pronged safeguards challenge.

这带来了三管齐下的保障挑战。

It is nearly impossible to predict model deployment environments, but essential to try.

预测模型部署环境几乎是不可能的，但必须尝试。

Rigorous and continuous safety and security testing is a necessity but will not preclude harm.

严格且持续的安全测试是必要的，但并不能消除伤害。

And guardrails adopted as a result may operationally constrain a military, putting it at a disadvantage compared to its adversaries without guardrails.

因此采用的护栏可能会在行动上限制军队，使其与没有护栏的对手相比处于劣势。

Verifying that a model is behaving as expected and not veering toward misalignment is challenging enough in civilian conditions, let alone a complex deployment setting like a war zone.

在民用条件下，验证模型的行为是否符合预期并且不会出现偏差已经足够具有挑战性，更不用说像战区这样复杂的部署环境了。

Inadvertent Escalation AI compresses decision-making timelines and can accelerate the pace of war.

无意中的升级人工智能会压缩决策时间并加快战争节奏。

One guardrail suggested is to keep a “human in the loop” where AI is involved in military decision-making.

建议的一个护栏是让人工智能参与军事决策时保持“人在循环中”。

But the danger is that this becomes merely a rubber-stamp exercise, with humans deferring to AI’s decision.

但危险在于，这只是橡皮图章练习，人类会听从人工智能的决定。

There is an important difference between this kind of nominal review and meaningful, informed human control.

这种名义上的审查与有意义的、知情的人类控制之间存在着重要的区别。

In either case, the effectiveness of a “human in the loop” in military contexts is under-reported.

无论哪种情况，“人在回路”在军事环境中的有效性都没有得到充分报道。

Reports from Gaza hint at the risks of nominal review, where military officers reported spending merely 20 seconds reviewing a target identified by AI algorithms.

来自加沙的报告暗​​示了名义审查的风险，军官报告称只花了 20 秒审查人工智能算法识别的目标。

Although the picture should sharpen as more information becomes available, this dynamic could also provide context for America’s strike on a school in Iran, where the humans in the loop may have approved the action based on outdated data.

尽管随着更多信息的出现，情况应该会更加清晰，但这种动态也可能为美国对伊朗一所学校的袭击提供背景，那里的人可能已经根据过时的数据批准了这一行动。

As AI enables and encourages soldiers to make operational decisions at an accelerating pace, and commanders increasingly rely on the “advice” given by these models, which risks inadvertent escalation during a military operation.

随着人工智能使并鼓励士兵加快做出作战决策，指挥官越来越依赖这些模型给出的“建议”，这可能会在军事行动中无意中升级。

There is also increasing evidence that AI may change the final decision itself.

越来越多的证据表明人工智能可能会改变最终决定本身。

During wargaming studies conducted by Stanford and King’s College London , AI models showed escalatory and unpredictable behavior.

在斯坦福大学和伦敦国王学院进行的兵棋推演研究中，人工智能模型表现出了升级和不可预测的行为。

Models in the King’s study threatened nuclear use in 95 percent of scenarios, and some models in the Stanford study executed nuclear strikes despite having had options for de-escalation.

国王研究中的模型在 95% 的情况下都威胁使用核武，而斯坦福大学研究中的一些模型尽管有降级选项，但还是执行了核打击。

If a human is taken out of the loop and militaries depend on increasingly agentic AI, this threat becomes all the more real, particularly if the escalation is occurring between nuclear powers.

如果人类被排除在外，而军队越来越依赖人工智能，那么这种威胁就会变得更加真实，特别是在核大国之间发生升级的情况下。

In this situation, a warring party could misconstrue “rogue” action by an AI model as intentional, prompting retaliation.

在这种情况下，交战方可能会将人工智能模型的“流氓”行为误解为故意行为，从而引发报复。

The United Kingdom is a nuclear-weapon state.

英国是一个核武器国家。

As the country integrates AI across its military, it confronts the very real risk of cross-domain escalation, with offensive actions across several environments launched quickly and without restraint.

随着该国将人工智能整合到军队中，它面临着跨域升级的非常现实的风险，跨多个环境的进攻行动迅速而不受限制地发起。

As a worst-case scenario, escalation could spiral to the nuclear level.

最坏的情况是，事态升级可能升级至核级别。

International guardrails and best practices to avoid escalation are undoubtedly emerging.

避免升级的国际护栏和最佳实践无疑正在出现。

The United States and China, for example, committed to an AI incident notification hotline in talks in September.

例如，美国和中国在 9 月份的会谈中承诺建立人工智能事件通知热线。

But there is an argument to be made that there are currently no sufficient, specific risk-reduction and response mechanisms to guard against AI-enabled miscalculation or unintended escalation, such as dedicated military-to-military AI hotlines.

但有一种观点认为，目前没有足够、具体的风险降低和响应机制来防止人工智能导致的误判或意外升级，例如专门的军队间人工智能热线。

There are patchwork — but insubstantial, and not specified to a military context — approaches to shared protocols (signaling where a system’s output may not be trustworthy), and transparency (namely, into where AI is used in decision-making).

有共享协议（在系统输出可能不可信的地方发出信号）和透明度（即人工智能在决策中的使用情况）的拼凑方法（但不是实质性的，并且不是专门针对军事环境的）。

One of the underlying dangers is that both relatively unsophisticated and cutting-edge AI capabilities can contribute to escalation risks.

潜在的危险之一是相对不复杂和尖端的人工智能功能都可能导致风险升级。

The strategic landscape is fraught, but London nonetheless has an opportunity to build reliable, responsible leadership on military AI.

战略形势充满挑战，但伦敦仍然有机会在军事人工智能领域建立可靠、负责任的领导地位。

Doing so demands that the new government tackle the above risks head-on, taking decisive steps through its research and development investments, multilateral engagements, and commercial and military partnerships.

为此，新政府必须正面应对上述风险，通过研发投资、多边参与以及商业和军事伙伴关系采取果断措施。

Underlying each should be the United Kingdom’s strategic interest in being a responsible state actor and upholding the rule of law.

每一项的基础都应该是英国作为负责任的国家行为体和维护法治的战略利益。

Build Reliability The United Kingdom is well placed to be a global leader on military AI risk identification and mitigation.

建立可靠性英国完全有能力成为军事人工智能风险识别和缓解方面的全球领导者。

It boasts strong technical authority on AI security and safety, with the AI Security Institute trailblazing on independent evaluations.

在人工智能安全保障方面拥有强大的技术权威，人工智能安全研究所在独立评估方面处于领先地位。

Its independent research ecosystem and civil society have a strong track record in open-source investigations and conflict risk monitoring.

其独立研究生态系统和民间社会在开源调查和冲突风险监测方面拥有良好的记录。

As the Ministry of Defence integrates AI across the military, it ought to work closely with independent evaluators, researchers, and civil society to develop context-specific benchmarks for testing and monitoring model performance in simulated conflict environments.

随着国防部将人工智能整合到军队中，它应该与独立评估者、研究人员和民间社会密切合作，制定针对特定背景的基准，用于在模拟冲突环境中测试和监控模型性能。

This is especially important as the military may rely on models it has not trained itself, although there is no transparency over whether foreign models are deployed in a military setting, or how they operate them.

这一点尤其重要，因为军方可能依赖于未经自身训练的模型，尽管外国模型是否部署在军事环境中以及如何操作这些模型尚不透明。

Technical reliability and shared standards are essential for trusted partnerships.

技术可靠性和共享标准对于值得信赖的合作伙伴关系至关重要。

For joint operations, London can work with allies on verification (in other words, ensuring models perform across a range of scenarios as expected) and develop and update benchmarks.

对于联合行动，伦敦可以与盟友合作进行验证（换句话说，确保模型按预期在一系列场景中执行）并开发和更新基准。

There may already be channels via NATO and between national AI Security Institutes, which governments can activate.

政府可以激活通过北约和国家人工智能安全研究所之间的渠道。

Bolster Existing Cooperation Full multilateral consensus on the regulation of military AI is politically infeasible.

加强现有合作就军事人工智能监管达成全面的多边共识在政治上是不可行的。

But pockets of agreement are possible and necessary, and there is growing appetite for international agreement.

但达成一些协议是可能且必要的，而且人们对国际协议的兴趣日益增长。

The U.K. government is also in a unique position to broker it: Doing so is in its interest, as an opportunity for bold but responsible and reliable international leadership on military AI.

英国政府在调解这一问题上也处于独特的地位：这样做符合英国的利益，也是在军事人工智能方面发挥大胆但负责任和可靠的国际领导作用的机会。

A first step can be to clarify the government’s “red lines” around military AI and to promote norms — including for responsible use and against civilian harm — with democratic partners.

第一步可以是澄清政府围绕军事人工智能的“红线”，并与民主伙伴一起推广规范，包括负责任地使用和防止平民伤害。

There is an opportunity now: AI security and safety concerns have re-entered the policy mainstream, with many rooted in anxieties about humanity and loss of life.

现在有一个机会：人工智能安全问题已重新进入政策主流，其中许多根源于对人类和生命损失的焦虑。

London should proactively engage with fellow travelers — democratic, middle-power governments — to discuss and agree on these “red lines.” There is a purpose-built summit for this which could benefit from a boost of diplomatic capital: The next Responsible Artificial Intelligence in the Military Domain meeting will take place in Kenya in April 2027.

伦敦应该积极与同行——民主的中等强国政府——讨论并就这些“红线”达成一致。为此有一个专门的峰会，可以从外交资本的增加中受益：下一次军事领域负责任的人工智能会议将于 2027 年 4 月在肯尼亚举行。

Having already co-hosted the 2024 Seoul summit and being one of the 45 states that endorsed the 2026 Pathways to Action , the U.K. government could consider committing additional resources to support Track 1.5 dialogues alongside it.

英国已经共同主办了 2024 年首尔峰会，并且是批准 2026 年行动路径的 45 个国家之一，因此英国政府可以考虑投入更多资源来支持 1.5 轨对话。

A larger gathering is also on the horizon: The Geneva AI Summit, scheduled for June 2027, could also be an opening for dialogue on AI and civilian harm, given Switzerland’s strong institutional foundation for advancing humanitarian consensus.

一场更大规模的聚会也即将举行：鉴于瑞士在推进人道主义共识方面拥有强大的制度基础，定于 2027 年 6 月举行的日内瓦人工智能峰会也可能成为有关人工智能和平民伤害对话的机会。

Take Big Diplomatic Bets Over the past month — an uncharacteristically intense one, even in the high-tempo AI governance space — the U.K. government has signaled willingness to facilitate some sort of global agreement on AI, although it has not signed last month’s middle-power declaration on controlling frontier AI.

在过去的一个月里，即使在快节奏的人工智能治理领域，这也是异常紧张的一个月，英国政府已表示愿意推动某种全球人工智能协议，尽管它尚未签署上个月关于控制前沿人工智能的中等强国宣言。

Military AI will not be on the United Kingdom’s 2027 Group of 20 agenda, which traditionally focuses on economic issues.

军事人工智能不会出现在英国 2027 年 20 国集团议程上，该议程历来关注经济问题。

But AI security more broadly will be on the agenda .

但更广泛的人工智能安全将提上议程。

Any agreement, such as one on mandatory international testing, will have knock-on effects for the use of AI in the military.

任何协议，例如关于强制性国际测试的协议，都将对人工智能在军队中的使用产生连锁反应。

The United Kingdom should task legal experts to explore how emerging frameworks and norms can apply to AI systems that are effectively dual-use.

英国应该责成法律专家探索新兴框架和规范如何适用于有效的双重用途人工智能系统。

Anxieties about artificial “superintelligence” — popularly understood as the point at which system capabilities fully outpace human intelligence — are closely tied to concerns about autonomous weapons systems.

对人工智能“超级智能”的担忧——通常被理解为系统能力完全超越人类智能——与对自主武器系统的担忧密切相关。

This is because the use of totally autonomous, artificial superintelligence for military gain could be catastrophic and have unpredictable impacts.

这是因为使用完全自主的超级人工智能来获取军事利益可能会带来灾难性的后果，并产生不可预测的影响。

Global calls for an international treaty banning artificial superintelligence are gaining volume, with a cross-party group of U.K. policymakers urging the government to support domestic legislation to this effect.

全球呼吁制定禁止人工智能的国际条约的呼声越来越高，英国政策制定者的跨党派团体敦促政府支持这方面的国内立法。

A treaty may be a policy pipe dream: U.S.-Chinese competition dynamics have a “gridlock” effect on prospects for international agreement.

条约可能是一个政策白日梦：美中竞争动态对国际协议的前景产生“僵局”效应。

But hunger for internationally binding rules is rising, and the United Kingdom has the opportunity to position itself at the center of these nascent efforts.

但对具有国际约束力的规则的渴望正在上升，英国有机会将自己置于这些新生努力的中心。

Relatedly, in August 2026, the United Nations secretary-general and the International Committee of the Red Cross renewed calls for another diplomatic effort: a treaty prohibiting unpredictable and antipersonnel autonomous weapons.

与此相关的是，2026 年 8 月，联合国秘书长和红十字国际委员会再次呼吁做出另一项外交努力：缔结一项禁止不可预测的杀伤人员自主武器的条约。

Over 70 states in a U.N.

联合国70多个国家

working group — the Group of Governmental Experts on Lethal Autonomous Weapons Systems under the Convention on Certain Conventional Weapons — committed to negotiating a legally binding instrument from November.

工作组——《特定常规武器公约》下的致命性自主武器系统政府专家组——承诺从 11 月起谈判一项具有法律约束力的文书。

Disappointingly, the U.K. government was not one of them, arguing that existing international law is sufficient to regulate lethal autonomous weapons despite having participated in the working group for the past three years.

令人失望的是，英国政府并不是其中之一，尽管英国政府在过去三年中一直参加该工作组，但他们认为现有的国际法足以监管致命性自主武器。

London should revisit this position ahead of the November negotiations.

伦敦应该在 11 月谈判之前重新审视这一立场。

The United Kingdom should continue to invest in new partnerships such as the one with Ukraine, which will provide invaluable insights for modernization and streamlining of the military to prepare for the new realities of conflict.

英国应继续投资于新的伙伴关系，例如与乌克兰的伙伴关系，这将为军队现代化和精简提供宝贵的见解，为应对新的冲突现实做好准备。

The government must approach its AI partnership with the United States, however, from a position of informed dependency and caution.

然而，政府必须以知情依赖和谨慎的立场来对待与美国的人工智能伙伴关系。

Contractually, partnerships must be designed to avoid any lock-in to U.S. technology providers that generates unacceptable levels of risk.

根据合同，合作伙伴关系的设计必须避免与美国技术提供商的任何锁定，从而产生不可接受的风险水平。

London ought to actively implement diplomatic measures and technical safeguards to protect against the possibility of a “kill switch” and continuously assess where national security concerns outweigh the advantages of access to cutting-edge American technology.

伦敦应积极实施外交措施和技术保障措施，防止出现“终止开关”的可能性，并不断评估国家安全担忧超过获得美国尖端技术优势的情况。

The United States aside, the United Kingdom should diversify its military AI partnerships, collaborating with government and industry from European and other allied partners like Canada and Japan.

除了美国之外，英国应该使其军事人工智能伙伴关系多样化，与欧洲以及加拿大和日本等其他盟国的政府和工业界合作。

These partnerships are already forming between other U.S. allies, and there are many prospects for the United Kingdom to both strengthen existing collaboration and pursue new opportunities.

美国其他盟友之间已经形成了这些伙伴关系，英国有很多前景可以加强现有合作并寻求新机遇。

Transparency, control, and safeguards must be front and center in each of these partnerships.

透明度、控制和保障措施必须成为这些伙伴关系的首要和核心。

Autonomy is possible amid informed dependence, and investing in both sovereign capabilities and middle-power coalitions may signal to Washington that it should more carefully reconsider the costs of its anti-regulatory stance on AI.

在知情依赖的情况下，自治是可能的，对主权能力和中等强国联盟的投资可能会向华盛顿发出信号，表明它应该更仔细地重新考虑其对人工智能的反监管立场的成本。

Where national security is paramount, London must support homegrown alternatives and can do so via procurement policy.

在国家安全至关重要的地方，伦敦必须支持本土替代方案，并且可以通过采购政策来做到这一点。

It can adopt supportive measures through investment decisions via the Sovereign AI Fund: Launched earlier this year, the $680 million fund supports British AI startups and has committed to funding those working on trust, safety, and assurance, stating that the United Kingdom’s comparative advantage is in “defining the standard .

英国可以通过主权人工智能基金的投资决策来采取支持措施：该基金于今年早些时候推出，价值 6.8 亿美元，支持英国人工智能初创企业，并致力于资助那些致力于信任、安全和保证的企业，并指出英国的比较优势在于“定义标准……”。

.

。

.

。

As AI moves into healthcare, finance, legal, and defense, trust becomes the question everything else depends on.” What this looks like in practice is shaping defense technology markets by championing the use of British-led trust and safety practices, such as verification and monitoring.

随着人工智能进入医疗保健、金融、法律和国防领域，信任成为其他一切所依赖的问题。”在实践中，这看起来像是通过倡导使用英国主导的信任和安全实践（例如验证和监控）来塑造国防技术市场。

Lead by Example While most legal experts and states agree that international humanitarian law applies to military AI, there is a lack of agreement in the interpretation of its application to specific use cases.

以身作则 虽然大多数法律专家和国家都同意国际人道法适用于军事人工智能，但在解释其在特定用例中的应用方面却缺乏共识。

Waiting for the courts to adjudicate these issues is far too reactive, particularly while AI targeting systems are already in use and contributing to civilian casualties.

等待法院对这些问题做出裁决太过被动，尤其是在人工智能瞄准系统已经投入使用并造成平民伤亡的情况下。

The United Kingdom should lead responsibly and ethically by clarifying London’s position on how states can use military AI in compliance with international humanitarian law, particularly the obligations of distinction, proportionality, and precaution, which pose the greatest potential issue.

英国应该以负责任和道德的方式发挥领导作用，阐明伦敦对各国如何根据国际人道主义法使用军事人工智能的立场，特别是区分、相称和预防的义务，这些义务构成了最大的潜在问题。

It can back this position with immediate and unilateral action by being more transparent about how its own military uses AI, so that humanitarian organizations and other independent parties can assess compliance with international humanitarian law.

它可以通过立即采取单边行动来支持这一立场，方法是提高本国军队如何使用人工智能的透明度，以便人道主义组织和其他独立各方能够评估国际人道主义法的遵守情况。

Full transparency, particularly where it reveals sensitive information about capabilities, may never be possible.

完全透明，特别是在披露有关能力的敏感信息的情况下，可能永远不可能。

But national security concerns do not preclude reporting on, for example, the extent of AI integration with the broader decision-making architecture, how the military is implementing operational risk mitigation measures, and how it is revising command accountability principles to account for the use of AI.

但国家安全担忧并不妨碍报道人工智能与更广泛的决策架构的整合程度、军队如何实施作战风险缓解措施，以及如何修订指挥问责原则以考虑人工智能的使用。

In more sensitive areas, the U.K. government can collaborate with a trusted evaluator like the AI Security Institute, and publicize that the partnership is occurring without detailing information about sensitive evaluations.

在更敏感的领域，英国政府可以与人工智能安全研究所等值得信赖的评估机构合作，并在不提供有关敏感评估的详细信息的情况下公开合作伙伴关系的发生。

The government should also implement safeguards that preserve meaningful human control where it is ethically or legally required, such as systems that keep decision-makers engaged and comprehensive data collection of an AI model’s “reasoning” alongside human decisions.

政府还应实施保障措施，在道德或法律要求的情况下保留有意义的人类控制，例如让决策者参与的系统以及人工智能模型“推理”与人类决策的全面数据收集。

Taking these actions builds public trust and sets a precedent for other countries to follow, which lays the groundwork for global standards for the integration and reliability of military AI.

采取这些行动可以建立公众信任，并为其他国家树立先例，为军事人工智能集成和可靠性的全球标准奠定基础。

The United Kingdom’s desire to obtain a strategic advantage through investment in military AI is reasonable, but speed cannot outweigh security.

英国希望通过投资军事人工智能获得战略优势是合理的，但速度不能超过安全。

As countries race ahead with AI integration while touting the benefits, they also need to build mechanisms that address the risks: rigorous testing, safeguarded and diversified partnerships, clear red lines, and transparency about integration.

随着各国在人工智能融合方面大肆宣传其好处的同时，他们还需要建立应对风险的机制：严格的测试、受保障和多元化的合作伙伴关系、明确的红线以及融合的透明度。

The United Kingdom has the technical authority and global status to promote these requirements.

英国拥有推动这些要求的技术权威和全球地位。

Governments and militaries cannot afford to wait for severe harm before they act.

政府和军队不能等到严重损害才采取行动。

Georgia Cole is a research fellow in the International Security Programme at Chatham House.

乔治亚·科尔是查塔姆研究所国际安全项目的研究员。

Her work focuses on emerging technologies in conflict, arms control, and nuclear weapons policy.

她的工作重点是冲突、军备控制和核武器政策中的新兴技术。

Isabella Wilkinson is a senior research fellow in the Digital Society Programme at Chatham House.

伊莎贝拉·威尔金森 (Isabella Wilkinson) 是查塔姆研究所数字社会项目的高级研究员。

Her work focuses on the geopolitics of global technology governance (specifically AI), new centers of decision-making power on technology, and disinformation.

她的工作重点是全球技术治理（特别是人工智能）的地缘政治、技术决策权的新中心以及虚假信息。