Analytical Essay: Intelligence Obligations: Artificial and Traditional
Artificial intelligence (AI) is changing the way many people around the globe do their jobs. It should be expected that intelligence analysts are then also using this new technology to make their jobs simultaneously easier and their scope more expansive. Although the ability to analyze data sets is not a novel concept, the speed at which AI can draw trend lines and connect disparate pieces of information as actionable intelligence goes beyond what a single intelligence analyst could do, historically (Bellaby 2024; Lowenthal 2025).[cite: 16] While AI’s ability to derive intelligence and make judgments can scale at the speed of compute, the mechanisms for transparency, accountability, and oversight may not scale in comparison unless policy changes chart the map ahead.[cite: 16]
The ethical implications of AI are being sorted out in real time across a myriad of impacted industries. Software engineers in the tech industry are able to create more code and solve bigger problems faster, but ultimately they are responsible for their own output, regardless of who wrote it. Intelligence agencies and their analysts are facing a similar dilemma with regard to AI as a new tool in the intelligence toolkit. U.S. persons are protected by the Fourth Amendment of the U.S. Constitution when it comes to searches and seizures, but it has limits around digital footprints that are not settled. But a gap exists between information that is willingly given away, information that is obtained, and information that is understood by context clues, which is where intelligence analysts live. Bellaby (2024) recognizes that many people are unaware that the information that they give out may be then harvested in categorical terms. Vacation photos could be mined for geolocation data, facial recognition, or social networks. Intelligence agencies have the ability to collect this data, wholesale, through various scraping tools or application programming interfaces. This overwhelming collection and analysis of data may have technically been possible previously, but the speed, cost, and breadth to draw individualized, pattern-of-life routines and behaviors is novel (Lowenthal 2025), giving intelligence agencies the ability to peer far further into individual lives without updated oversight or Fourth Amendment guardrails.[cite: 16]
In Paul Scharre’s Army of None (2018), he relays a story about his time in Afghanistan of encountering a girl, unsure if the person was perhaps in communication with malign actors as a lookout. Scharre recounts that he was able to get close enough to hear the girl singing, a humanizing, contextual detail that put the soldiers at ease. Although the book is about autonomous weapons, intelligence analysts leaning on artificial intelligence for insights face similar dilemmas about model training and context. A hypothetical example would be how an AI model might see a man carrying a Kalashnikov rifle slung over his shoulder. In many circumstances, it could be an overt threat. However, AI needs to take into account local customs and norms, as well as the security posture of the surveilled area (Scharre 2018; Bellaby 2024). In Kabul, perhaps it’s not uncommon to see a Kalashnikov. But in Atlanta, GA, the context changes.[cite: 16]
When human analysts are not up to par, they are held accountable via a chain of command. However, when an AI agent incorrectly identifies a target, who is to blame (Lowenthal 2025)?[cite: 16] In many cases, there are still humans in the loop, acting as a circuit breaker when an agent or model incorrectly proposes a solution. In a “human on the loop” future, the human will sit outside of the decision tree, intervening ideally before the execution of an incorrect task. In the Russo-Ukrainian war, fully-autonomous weapons have already been unleashed in a first-of-their-kind manner (Mako 2026).[cite: 16] Intelligence is not the same as drone warfare, although the decision-making aspects of each can have catastrophic consequences if misaligned.[cite: 16]
A concrete example of the decision vulnerabilities facing intelligence analysts in the future happened in Iran in 2026. Bartenstein and Karra (2026) report that some of the intelligence that contributed to the bombing of an elementary school was gathered by AI.[cite: 16] Unfortunately, the compounding effects of outdated geospatial intelligence, site classification, and compressed timelines resulted in the deaths of 123 children.[cite: 16] However, AI is not solely to blame for this mistake. Humans at every level who acquiesced with the decision failed to break a deference of “automation bias.”[cite: 16] It is not that humans are infallible or incapable of making bad decisions, but when that happens, humans should also be able to articulate the train of thought that went into the decision-making process, even if it is illogical.[cite: 16] Probabilistic AI models, by contrast, given the same inputs and constraints may return a response with different recommendations and very little evidence of its line of reasoning (Kahn, Probasco, and Kinoshita 2024).[cite: 16] Humans, then, as a matter of policy and oversight, must remain not as passive button-pushers, but active participants in the decision-making.[cite: 16]
What mechanisms, then, should the U.S. Intelligence Community institute as a way to preserve both oversight and accountability? Perhaps the lowest bar to clear would be that all clandestine activities that involve kinetic components have a minimum human confirmation on the facts on the ground, in alignment with Lowenthal’s (2025) ethical frameworks.[cite: 16] AI provides insights and recommendations at fantastic speeds, but even the best special forces need time to get into physical position with the appropriate materiel.[cite: 16] Time and space are constraints that we have not yet overcome, and so even if the decision is lightning fast, the speed advantage forfeited for a human review is minimal.[cite: 16] This does not guarantee a future that is free from violent or tragic mistakes; humans will still kill humans.[cite: 16] But it should not be an ambiguous screen making the decisions.[cite: 16] And maybe it increases the likelihood that a human being with at least a modicum of geospatial intelligence will fact-check to ensure that an elementary school is not an IRGC barracks.[cite: 16]