Survival of the quickest: Military leaders aim to unleash, control AI

PARIS — Artificial intelligence is massively accelerating military decision making, and armed forces that don’t keep up risk being outmatched, the NATO commander in charge of strategic transformation at the alliance said at the AI Action Summit in Paris this week.

Alliance members are now using AI in the decision-making loop of observe, orient, decide and act, NATO Supreme Allied Commander Transformation Adm. Pierre Vandier, said at a conference focused on military AI. Analysis that previously took hours or days, such as processing large amounts of sensor data, can now be done in a matter of seconds, he said.

“The speed of operations will dramatically change,” Vandier said at a press briefing on Monday. “You see that in Ukraine. If you do not adapt at speed and at scale, you die.”

The major powers have identified AI as a key enabler for future warfare, with the U.S. spending billions on AI for defense, while trying to limit China’s access to enablers such as hardware from Nvidia. Meanwhile, summit host France says it plans to become the leader in military AI in Europe.

AI brings “a huge acceleration of the speed of decision,” Vandier said. “A huge acceleration that overtakes a lot of things in our system, and the system of the enemy we intend to outpace.”

Vandier made a comparison to the movie The Matrix, where the main character Neo dodges bullets by having learned to move faster than his opponent’s projectiles. “The question for us is, are we already dead? So it’s a question of speed of change.”

The speediness of AI raises questions about whether having a human in the control loop improves the quality of decision making, said Jeroen van der Vlugt, chief information officer at the Netherlands Ministry of Defence. He said AI can make decisions based on amounts of data that would be impossible for humans to manage, with analysis brought down to milliseconds.

A group of 25 countries at the Paris summit signed a declaration on AI-enabled weapon systems , pledging they won’t authorize life-and-death decisions by an autonomous weapon system operating completely outside human control. Summit co-chair India didn’t sign the declaration, nor did the U.K. or the United States.

“We already have militaries full of intelligent, autonomous agents – we call them soldiers or airmen or Marines,” said Gregory Allen, the director of the Wadhwani AI Center at the Center for Strategic and International Studies, a Washington-based think tank. “Just as military commanders are accountable, states are also responsible for the actions of their military forces, and nothing about the changing landscape of artificial intelligence is going to ever change those two facts.”

Germany’s Helsing and France’s Mistral AI on Monday announced an agreement to jointly develop AI systems for defense. Google owner Alphabet last week dropped a promise not to use AI for purposes such as developing weapons, while rival OpenAI in December announced a partnership with military technology company Anduril.

Frontier AI models will be useful in summarizing large intelligence reports and for war gaming and “red teaming,” said Ben Fawcett, product lead at Advai, which tests AI systems for vulnerabilities. “These kind of models will have a real utility in order to test commanders on how their plan will survive contact, especially if they’re able to update that based on what is the latest situation.”

The first AI-based simulation tools are arriving that allow commanders to test and refine plans before putting them into action, according to Vandier. He said AI doesn’t mean fewer human decisions but faster and better ones, at least in theory.

“AI is not a magic bullet,” Vandier said.”It gives solutions to go faster, better, more accurate, more lethal, but it won’t solve the war itself, because it’s a race between us and our competitors.”

Vandier and Van der Vlugt mentioned the importance of AI for autonomy and robotics, particularly swarming technology, which relies on AI to work. “The scalability and autonomy part of it is really changing our landscape at this moment,” Van der Vlugt said.

Success of AI depends on adoption, and Vandier has introduced a monthly learning package with required reading for officers at Allied Command Transformation in Norfolk, Virginia, after finding out his top brass didn’t know “that much” about AI.

“The technology goes so fast that ultimately, we realize that managers are not necessarily up to speed,” Vandier said. “So there is really a training challenge. If you want a head of capacity development, someone who defines the capacities of tomorrow, to be good, they need to have understood what is at stake with these technologies.”

Large language models over the past decade have been getting roughly 13 times better every year, and that trend is not expected to stop, meaning models might be more than 1,000 times better in three years and more than 1 million times better in 10 years, according to Allen at CSIS.

“What we aren’t seeing right now is large language models generating unique insights that would be relevant to say, planning a campaign of war, fighting operations,” Allen said. “Just because they are very far away from that level of performance today doesn’t mean that they are very far away in terms of time, because performance is improving so rapidly.”

Large language models will be transformative for national security capabilities, which helps explain why the U.S. stopped selling AI chips to China, according to Allen. “We see in the not too distant future, genuinely transformative AI capabilities, and it’s important that that is a party that China is not invited to.”

When asked at the press briefing whether machines will take control, Vandier said he didn’t know. He mentioned the 1983 movie WarGames, in which a computer decides to trigger nuclear war, and the Terminator series of movies, whose premise includes an AI launching a nuclear attack against humanity, saying “it could happen.”

The NATO commander said while fears around AI are understandable, citizens already have the technology in their pockets with smart phones. He said new technology is not inherently good or bad, what matters is the use case.

“What people want when they fight is not to be all destroyed, they want to win,” Vandier said. “As it has been for nuclear arms, one day we will have to find ways to control the AI, or we will lose control of everything.”

Rudy Ruitenberg is a Europe correspondent for Defense News. He started his career at Bloomberg News and has experience reporting on technology, commodity markets and politics.

巴黎——北约负责战略转型的指挥官本周在巴黎举行的人工智能行动峰会上表示，人工智能正在大幅加快军事决策，跟不上人工智能发展的武装部队将面临被超越的风险。

北约最高盟军转型司令皮埃尔·范迪​​耶海军上将在一次以军事人工智能为主题的会议上表示，联盟成员现在正在将人工智能应用于观察、判断、决策和行动的决策循环中。他指出，过去需要数小时甚至数天才能完成的分析工作，例如处理大量传感器数据，现在只需几秒钟即可完成。

“作战节奏将发生巨大变化，”范迪尔周一在新闻发布会上说。“乌克兰的局势就是如此。如果你不能迅速且大规模地做出调整，你就死定了。”

各大强国已将人工智能视为未来战争的关键推动因素。美国在国防人工智能领域投入数十亿美元，同时试图限制中国获取英伟达等公司的硬件等技术。与此同时，峰会东道主法国表示，计划成为欧洲军事人工智能领域的领导者。

范迪尔说，人工智能带来了“决策速度的巨大提升”。“这种巨大的提升超越了我们系统中的许多方面，也超越了我们打算超越的敌方系统。”

范迪尔将此比作电影《黑客帝国》，片中主角尼奥通过学会比对手的子弹更快移动来躲避子弹。“对我们来说，问题是，我们是否已经死了？所以，关键在于变革的速度。”

荷兰国防部首席信息官杰伦·范德弗卢格特表示，人工智能的超高速度引发了人们的疑问：在控制回路中引入人类是否真的能提高决策质量？他指出，人工智能能够基于人类根本无法处理的海量数据做出决策，分析速度更是快至毫秒级。

在巴黎峰会上，25个国家签署了一份关于人工智能武器系统的宣言，承诺不会授权完全不受人类控制的自主武器系统做出关乎生死的决定。峰会联合主席国印度、英国和美国均未签署该宣言。

“我们的军队中已经充满了智能自主的个体——我们称他们为士兵、飞行员或海军陆战队员，”华盛顿智库战略与国际研究中心瓦德瓦尼人工智能中心主任格雷戈里·艾伦表示。“正如军事指挥官要承担责任一样，国家也对其军队的行为负责，人工智能领域的任何变革都无法改变这两个事实。”

德国的Helsing公司和法国的Mistral AI公司周一宣布达成协议，将联合开发用于国防的人工智能系统。谷歌母公司Alphabet上周放弃了此前承诺不将人工智能用于武器研发等用途，而其竞争对手OpenAI则在去年12月宣布与军事技术公司Anduril建立合作关系。

Advai公司产品负责人本·福塞特表示，前沿人工智能模型将有助于总结大型情报报告，并应用于战争推演和“红队演练”。Advai是一家专门测试人工智能系统漏洞的公司。“这类模型对于测试指挥官的作战计划在遭遇战中的有效性非常实用，尤其是在他们能够根据最新情况更新计划的情况下。”

范迪尔表示，首批基于人工智能的模拟工具即将问世，指挥官们可以在行动前测试并完善计划。他指出，人工智能并不意味着减少人为决策，而是至少在理论上，能够做出更快、更好的决策。

“人工智能并非万能灵药，”范迪尔说。“它能提供更快、更好、更准确、更致命的解决方案，但它无法解决战争本身，因为这是我们与竞争对手之间的一场竞赛。”

Vandier 和 Van der Vlugt 都提到了人工智能对自主系统和机器人技术的重要性，尤其是依赖人工智能才能运行的群体智能技术。“目前，人工智能的可扩展性和自主性正在真正改变我们的格局，”Van der Vlugt 说。

人工智能的成功取决于其应用，范迪尔在发现他的高级将领对人工智能“了解不多”之后，为弗吉尼亚州诺福克盟军转型司令部的军官们推出了每月学习包，其中包含必读材料。

“技术发展日新月异，最终我们发现管理者未必能跟上步伐，”范迪尔说。“因此，培训确实是一个挑战。如果你想要一位优秀的能力发展负责人，一位能够定义未来能力的人，他们就必须了解这些技术所带来的风险。”

过去十年，大型语言模型每年大约改进 13 倍，而且这种趋势预计不会停止，这意味着模型在三年内可能会改进 1000 倍以上，在 10 年内可能会改进 100 万倍以上，CSIS 的 Allen 表示。

艾伦说：“我们目前还没有看到大型语言模型能够产生独特的洞见，而这些洞见对于制定战争计划、执行作战任务等都至关重要。虽然它们目前的性能还远未达到那种水平，但这并不意味着它们在时间上也遥遥无期，因为它们的性能提升速度非常快。”

艾伦认为，大型语言模型将对国家安全能力产生变革性影响，这也解释了为什么美国停止向中国出售人工智能芯片。“我们预见，在不久的将来，人工智能将真正具有变革性的能力，而重要的是，中国不应参与其中。”

在新闻发布会上被问及机器是否会控制世界时，范迪尔表示他不知道。他提到了1983年的电影《战争游戏》，其中一台电脑决定引发核战争；他还提到了《终结者》系列电影，该系列电影的前提是人工智能对人类发动核攻击。范迪尔说：“这种情况有可能发生。”

北约指挥官表示，人们对人工智能的担忧可以理解，但民众的智能手机已经将这项技术装进了口袋。他指出，新技术本身无所谓好坏，关键在于它的应用场景。

“人们发动战争的目的不是为了被彻底摧毁，而是为了胜利，”范迪尔说。“就像核武器的发展历程一样，总有一天我们必须找到控制人工智能的方法，否则我们将失去对一切的控制。”

鲁迪·鲁伊滕贝格是《防务新闻》的欧洲记者。他的职业生涯始于彭博新闻社，拥有报道科技、大宗商品市场和政治方面的经验。