U.S. intelligence agencies are intensifying their efforts to integrate artificial intelligence (AI) to handle the overwhelming influx of data from advanced surveillance technologies. They also aim to stay competitive with global rivals utilizing AI for sophisticated tasks, such as spreading deepfakes on social media.
Although AI holds significant potential, it is still an emerging and fragile technology. Officials acknowledge that generative AI, despite its power, is not yet ideal for intelligence work that involves high stakes and deception.
CIA Director William Burns emphasized the need for advanced AI models capable of processing vast amounts of both open-source and clandestine information. However, Nand Mulchandani, the CIA's inaugural chief technology officer, cautioned that generative AI models can produce unreliable outputs, likening them to a "crazy, drunk friend" with insightful but occasionally misleading information.
Security and privacy remain major concerns. Adversaries could compromise these AI systems, potentially poisoning the data or stealing sensitive information. AI is seen as a valuable assistant for identifying critical data points but not a replacement for human analysts.
Thousands of analysts across 18 U.S. intelligence agencies now use a CIA-developed generative AI called Osiris. This AI processes unclassified and commercially available data to create annotated summaries and includes a chatbot for follow-up questions. Osiris utilizes multiple commercial AI models without committing to a single vendor, indicating the experimental phase of this technology.
Experts foresee AI's role in predictive analysis, war-gaming, and scenario brainstorming as crucial for intelligence operations.
Before the rise of generative AI, intelligence agencies already used machine learning and algorithms for tasks like alerting analysts to critical developments during off hours. Companies like Microsoft and Primer AI are competing to provide AI solutions tailored for intelligence work, capable of sifting through vast amounts of data from diverse sources to detect emerging threats.
The immediate challenges include countering adversaries' use of AI for penetrating U.S. defenses, spreading disinformation, and undermining intelligence capabilities. The White House is concerned about the security of generative AI models used by U.S. agencies, particularly the risk of infiltration and data poisoning.
Ensuring the privacy of individuals whose data may be embedded in AI models is another significant issue. Intelligence agencies are cautious about adopting a "move-fast-and-break-things" approach due to these risks.
The adoption of AI varies across intelligence agencies based on their specific missions. For example, the National Geospatial-Intelligence Agency (NGA) focuses on extracting geospatial intelligence from imagery and sensors, which could benefit greatly from AI. In contrast, the FBI faces legal restrictions on domestic surveillance that complicate AI integration.
Despite AI's advancements, it won't easily replace human analysts who work with incomplete and often contradictory information. Quick, decisive actions based on human judgment remain crucial, especially for intelligence briefings to the president. Former CIA deputy director of analysis Linda Weissgold asserts that human insight and experience are irreplaceable in the decision-making process.
In summary, while AI offers transformative potential for intelligence operations, its current limitations and risks necessitate a careful and measured approach. The integration of AI will complement but not replace the critical role of human analysts in the intelligence community.
