Spy Vs. AI

U.S. Diplomacy


Since its founding in 1922, Foreign Affairs has been the leading forum for severe discussion of American foreign policy and worldwide affairs. The publication has featured contributions from lots of prominent international affairs professionals.


More Resources


- Feedback

- Institutional Subscriptions

- Gift a Membership

- About Us

- Events

- Issue Archive

- Advertise

- Audio Content

- Account Management

- FAQs


Spy vs. AI


ANNE NEUBERGER is Deputy Assistant to the President and Deputy National Security Adviser for Cyber and Emerging Technology on the U.S. National Security Council. From 2009 to 2021, she served in senior operational functions in intelligence and cybersecurity at the National Security Agency, including as its very first Chief Risk Officer.


- More by Anne Neuberger


Spy vs. AI


How Artificial Intelligence Will Remake Espionage


Anne Neuberger


-.

Copy Link Copied.
Article link: https://www.foreignaffairs.com/united-states/spy-vs-aihttps://www.foreignaffairs.com/united-states/spy-vs-ai.
Copy



Gift Link Copied.
This is a subscriber-only function. Subscribe now or Check in.



Create Citation Copied.
Chicago MLA APSA APA.
Chicago Cite not available at the minute.
MLA Cite not available at the moment.
APSA Cite not available at the moment.
APA Cite not available at the moment



Download PDF.
This is a subscriber-only feature. Subscribe now or Sign in.



Request Reprint.
Request reprint permissions here.






In the early 1950s, the United States faced a vital intelligence challenge in its blossoming competitors with the Soviet Union. Outdated German reconnaissance photos from The second world war might no longer supply enough intelligence about Soviet military capabilities, and existing U.S. monitoring abilities were no longer able to permeate the Soviet Union's closed airspace. This shortage spurred an adventurous moonshot initiative: the development of the U-2 reconnaissance aircraft. In just a few years, U-2 missions were delivering crucial intelligence, capturing images of Soviet missile setups in Cuba and bringing near-real-time insights from behind the Iron Curtain to the Oval Office.


Today, the United States stands at a similar juncture. Competition in between Washington and its competitors over the future of the international order is intensifying, and now, much as in the early 1950s, the United States should take advantage of its first-rate personal sector and ample capability for development to outcompete its adversaries. The U.S. intelligence community need to harness the country's sources of strength to deliver insights to policymakers at the speed of today's world. The integration of expert system, especially through big language designs, uses groundbreaking opportunities to improve intelligence operations and analysis, allowing the shipment of faster and more appropriate support to decisionmakers. This technological transformation features substantial drawbacks, nevertheless, especially as enemies make use of similar advancements to uncover and counter U.S. intelligence operations. With an AI race underway, the United States must challenge itself to be first-first to gain from AI, first to secure itself from enemies who may utilize the technology for ill, and first to utilize AI in line with the laws and values of a democracy.


For the U.S. national security neighborhood, fulfilling the guarantee and handling the danger of AI will require deep technological and cultural changes and a willingness to change the method companies work. The U.S. intelligence and military neighborhoods can harness the capacity of AI while alleviating its inherent threats, guaranteeing that the United States maintains its one-upmanship in a rapidly evolving worldwide landscape. Even as it does so, the United States need to transparently convey to the American public, and to populations and partners around the world, how the country intends to fairly and safely use AI, in compliance with its laws and worths.


MORE, BETTER, FASTER


AI's potential to revolutionize the intelligence community depends on its capability to procedure and examine vast amounts of information at unprecedented speeds. It can be challenging to analyze big quantities of gathered information to create time-sensitive warnings. U.S. intelligence services might utilize AI systems' pattern recognition abilities to determine and alert human experts to potential dangers, such as missile launches or military motions, or important worldwide advancements that experts know senior U.S. decisionmakers have an interest in. This capability would ensure that vital cautions are timely, actionable, and relevant, permitting more effective responses to both quickly emerging dangers and emerging policy chances. Multimodal models, which incorporate text, images, and audio, enhance this analysis. For instance, utilizing AI to cross-reference satellite imagery with signals intelligence could offer a detailed view of military motions, allowing much faster and more precise risk assessments and possibly new ways of delivering details to policymakers.


Intelligence analysts can likewise offload repetitive and lengthy jobs to makers to focus on the most fulfilling work: generating original and deeper analysis, increasing the intelligence community's total insights and performance. A fine example of this is foreign language translation. U.S. intelligence firms invested early in AI-powered capabilities, and the bet has paid off. The capabilities of language designs have grown increasingly sophisticated and accurate-OpenAI's recently launched o1 and o3 designs showed substantial progress in accuracy and reasoning ability-and can be used to even more quickly equate and sum up text, audio, and video files.


Although challenges remain, future systems trained on higher quantities of non-English information could be efficient in discerning subtle distinctions between dialects and comprehending the meaning and cultural context of slang or Internet memes. By depending on these tools, the intelligence community could concentrate on training a cadre of extremely specialized linguists, who can be hard to find, typically struggle to make it through the clearance process, and take a long time to train. And naturally, by making more foreign language products available across the best firms, U.S. intelligence services would be able to faster triage the mountain of foreign intelligence they receive to choose the needles in the haystack that really matter.


The value of such speed to policymakers can not be undervalued. Models can swiftly sort through intelligence information sets, open-source details, and standard human intelligence and produce draft summaries or preliminary analytical reports that experts can then confirm and fine-tune, guaranteeing the last products are both detailed and accurate. Analysts could coordinate with an advanced AI assistant to work through analytical problems, test ideas, and brainstorm in a collective fashion, enhancing each iteration of their analyses and providing finished intelligence faster.


Consider Israel's experience in January 2018, when its intelligence service, the Mossad, covertly burglarized a secret Iranian facility and took about 20 percent of the archives that detailed Iran's nuclear activities in between 1999 and 2003. According to Israeli officials, the Mossad collected some 55,000 pages of files and a further 55,000 files kept on CDs, consisting of pictures and videos-nearly all in Farsi. Once the archive was obtained, senior officials positioned immense pressure on intelligence specialists to produce detailed evaluations of its content and whether it pointed to a continuous effort to build an Iranian bomb. But it took these professionals several months-and numerous hours of labor-to translate each page, examine it by hand for appropriate content, and include that details into assessments. With today's AI abilities, the first 2 steps in that procedure might have been accomplished within days, possibly even hours, allowing experts to comprehend and contextualize the intelligence quickly.


One of the most intriguing applications is the way AI could transform how intelligence is consumed by policymakers, enabling them to communicate straight with intelligence reports through ChatGPT-like platforms. Such capabilities would allow users to ask particular questions and get summarized, appropriate details from countless reports with source citations, assisting them make notified decisions quickly.


BRAVE NEW WORLD


Although AI uses numerous benefits, it also positions substantial brand-new risks, especially as foes establish similar innovations. China's improvements in AI, particularly in computer system vision and surveillance, threaten U.S. intelligence operations. Because the nation is ruled by an authoritarian regime, it does not have personal privacy constraints and civil liberty protections. That deficit makes it possible for large-scale information collection practices that have actually yielded information sets of tremendous size. Government-sanctioned AI models are trained on large quantities of individual and behavioral data that can then be utilized for different purposes, such as security and social control. The presence of Chinese business, such as Huawei, in telecoms systems and software application around the world could offer China with ready access to bulk information, significantly bulk images that can be utilized to train facial acknowledgment models, a specific issue in nations with large U.S. military bases. The U.S. national security community must think about how Chinese models built on such comprehensive data sets can offer China a tactical benefit.


And it is not simply China. The expansion of "open source" AI designs, such as Meta's Llama and those developed by the French business Mistral AI and the Chinese company DeepSeek, is putting powerful AI abilities into the hands of users around the world at fairly economical expenses. Many of these users are benign, however some are not-including authoritarian routines, cyber-hackers, and criminal gangs. These malign actors are utilizing large language designs to rapidly produce and spread out false and malicious content or to conduct cyberattacks. As seen with other intelligence-related technologies, such as signals obstruct abilities and unmanned drones, China, Iran, and Russia will have every incentive to share a few of their AI developments with customer states and subnational groups, such as Hezbollah, Hamas, and garagesale.es the Wagner paramilitary company, consequently increasing the danger to the United States and its allies.


The U.S. military and intelligence community's AI models will become appealing targets for adversaries. As they grow more powerful and main to U.S. nationwide security decision-making, intelligence AIs will end up being critical nationwide properties that must be protected against enemies looking for to compromise or manipulate them. The intelligence community need to invest in establishing safe and secure AI models and in establishing requirements for "red teaming" and continuous assessment to safeguard against potential risks. These groups can utilize AI to replicate attacks, discovering possible weak points and developing strategies to reduce them. Proactive procedures, including cooperation with allies on and financial investment in counter-AI technologies, will be important.


THE NEW NORMAL


These obstacles can not be wished away. Waiting too long for AI innovations to totally mature carries its own dangers; U.S. intelligence capacities will fall behind those of China, Russia, and other powers that are going complete steam ahead in developing AI. To guarantee that intelligence-whether time-sensitive cautions or longer-term tactical insight-continues to be a benefit for the United States and its allies, the nation's intelligence community needs to adapt and innovate. The intelligence services must quickly master making use of AI and make AI a foundational element in their work. This is the only sure way to ensure that future U.S. presidents receive the best possible intelligence support, remain ahead of their enemies, and secure the United States' sensitive abilities and operations. Implementing these changes will need a cultural shift within the intelligence neighborhood. Today, intelligence experts mainly construct items from raw intelligence and information, with some support from existing AI models for voice and images analysis. Progressing, intelligence officials need to explore consisting of a hybrid technique, in line with existing laws, utilizing AI designs trained on unclassified commercially available information and improved with categorized details. This amalgam of innovation and standard intelligence gathering could result in an AI entity providing instructions to imagery, signals, open source, and measurement systems on the basis of an integrated view of typical and anomalous activity, automated images analysis, and automated voice translation.


To accelerate the transition, intelligence leaders must champion the advantages of AI integration, stressing the enhanced abilities and performance it offers. The cadre of newly designated chief AI officers has been developed in U.S. intelligence and defense to function as leads within their firms for promoting AI development and getting rid of barriers to the innovation's implementation. Pilot jobs and early wins can construct momentum and self-confidence in AI's abilities, encouraging broader adoption. These officers can utilize the proficiency of national laboratories and other partners to check and fine-tune AI designs, guaranteeing their efficiency and security. To institutionalise modification, leaders need to produce other organizational incentives, including promotions and training chances, to reward inventive approaches and those workers and units that show reliable usage of AI.


The White House has created the policy needed for using AI in nationwide security agencies. President Joe Biden's 2023 executive order relating to safe, safe and secure, and credible AI detailed the assistance required to fairly and safely utilize the innovation, and National Security Memorandum 25, issued in October 2024, is the country's foundational strategy for harnessing the power and handling the dangers of AI to advance nationwide security. Now, Congress will need to do its part. Appropriations are needed for departments and companies to produce the facilities required for innovation and experimentation, conduct and scale pilot activities and evaluations, and continue to purchase assessment capabilities to make sure that the United States is constructing reliable and high-performing AI technologies.


Intelligence and military neighborhoods are devoted to keeping humans at the heart of AI-assisted decision-making and have developed the frameworks and tools to do so. Agencies will require standards for how their analysts ought to use AI models to make certain that intelligence products fulfill the intelligence community's requirements for dependability. The federal government will likewise need to maintain clear assistance for managing the information of U.S. citizens when it pertains to the training and usage of big language designs. It will be essential to stabilize the use of emerging innovations with safeguarding the privacy and civil liberties of residents. This suggests augmenting oversight systems, updating appropriate frameworks to reflect the capabilities and dangers of AI, and cultivating a culture of AI development within the nationwide security device that utilizes the capacity of the innovation while securing the rights and freedoms that are foundational to American society.


Unlike the 1950s, when U.S. intelligence raced to the forefront of overhead and satellite imagery by developing a number of the crucial technologies itself, winning the AI race will need that community to reimagine how it partners with private market. The private sector, which is the main ways through which the government can recognize AI progress at scale, is investing billions of dollars in AI-related research, information centers, and calculating power. Given those business' improvements, intelligence companies need to focus on leveraging commercially available AI models and fine-tuning them with categorized data. This approach allows the intelligence community to rapidly expand its capabilities without needing to start from scratch, permitting it to remain competitive with foes. A recent collaboration between NASA and IBM to create the world's largest geospatial structure model-and the subsequent release of the model to the AI neighborhood as an open-source project-is an exemplary demonstration of how this type of public-private collaboration can operate in practice.


As the nationwide security neighborhood integrates AI into its work, it must ensure the security and resilience of its models. Establishing requirements to release generative AI firmly is crucial for maintaining the stability of AI-driven intelligence operations. This is a core focus of the National Security Agency's brand-new AI Security Center and its cooperation with the Department of Commerce's AI Safety Institute.


As the United States deals with growing rivalry to form the future of the international order, it is urgent that its intelligence agencies and military profit from the nation's development and leadership in AI, focusing especially on large language models, to provide faster and more appropriate details to policymakers. Only then will they gain the speed, breadth, and depth of insight required to browse a more complicated, competitive, and content-rich world.