Questions for Congress: AI in Targeting and Civilian Protection

At Palantir’s 2026 AIPCon, Chief Digital and Artificial Intelligence Officer of the Department of Defense, Cameron Stanley, asked, “How do you get better decisions faster than your adversary?” This statement reflects the underlying philosophy of speed and scale that motivates the development of AI and targeting technology. Autonomous weapons technologies are very real, and algorithmic warfare has already transformed the landscape of global armed conflict. But what protections are in place to ensure that these systems are used in a manner that is consistent with existing legal obligations? And what efforts are being made to anticipate and address the dangers that these systems may pose to civilians in conflict zones around the world?

Here we suggest questions for members of Congress to ask Department of Defense (DoD) officials to better understand the use of AI in targeting operations and how these technologies relate to the overarching strategic and moral imperative to avoid civilian harm in armed conflict.

What role do humans play in targeting decisions?

In U.S. military operations, AI capabilities generally can be divided into two camps: autonomous weapons systems (AWS) and AI Decision Support Systems (AI-DSS). An AWS is a “weapon that is able to select and apply force to targets without human control.” AWS have capabilities outside of agentic targeting, including “navigation, route tracking, flight/cruise control, intelligence, surveillance and reconnaissance” and often operate with “human in the loop” capabilities. Autonomous Weapons Watch, a database that compiles evidence and DoD communications regarding AI targeting systems, notes five U.S. systems that are “advertised or suspected to possess autonomous capabilities specifically in target selection and target engagement.” For example, U.S Central Command (CENTCOM) has noted that SpektreWorks’ LUCAS unmanned aerial munition system has “extensive range” and is “designed to operate autonomously.” AI-DSS systems, alternatively, use “AI techniques to collect and analyze data, provide information about the operational environment as well as actionable recommendations, with the aim of aiding military decision-makers.” The primary AI-DSS used in U.S. operations is Maven Smart System (MSS). 

In March 2026, DoD designated MSS as a formal program of record, cementing it as the flagship AI-DSS for the department. MSS’s usage is nearly ubiquitous in U.S. operations, enabling analysts to process immense volumes of data into actionable intelligence faster than humans could. The system functions as a single platform for commanders and military intelligence officers to view battle plans, surveillance, and ultimately conduct targeting operations. MSS is essentially a layering system that integrates numerous individual platforms, software, data sources, and third-party tools into a cohesive system where AI can detect and recommend targets, determine courses of action, and select munitions for the user to fire. With a single click, MSS can transfer coordinates from a tactical data link to a specific weapons platform, drastically reducing the time it takes to move the kill chain from detection to strike.

DoD employs both AWS and AI-DSS in its targeting operations. However, a black box surrounds which systems are used, how these systems and corresponding targeting decisions are made, what role humans play in those decisions, and how AI systems prosecute targets without human intervention. Key questions include:

1. 1. A loophole for decision support systems: DoD Directive 3000.09 requires testing, senior review, and appropriate human judgment for autonomous and semi-autonomous weapons. However, expert analysis suggests that AI-DSS falls outside the directive’s scope. Does the Department consider AI-DSS as subject to 3000.09? 
   2. Embedded human judgment prior to target approval: Human judgment is important at several stages prior to operational targeting. What processes are in place within DoD for determining the legal parameters governing strikes in the software design and development process? At the command level, how does DoD establish, for example, what geographic areas AWS are authorized to be used in, whether they may be used only against specific categories of targets or military objectives, time periods in which they may be used, or placing other constraints on use upstream of ultimate targeting decisions?
   3. Time for human review of nominated targets: In the Iran War, how much time does a human reviewer spend on each AI-nominated target before approval? How much time is spent per target in the lethal strikes against suspected drug traffickers at sea? For counterterrorism operations in Somalia? In other areas of operation? Does the Department track this data? Do these figures change for static versus dynamic targeting? How many targets is a single reviewer or cell expected to process in a shift?
   4. Information available for human review: What information does the reviewer see throughout the targeting process? Is a human reviewer seeing the underlying sources, the model’s confidence, and the age of the data, or only the recommendation? Does this change for static versus dynamic targeting?
   5. Human review outcomes: How often do human reviewers reject or modify AI-generated target nominations? Does the Department track that rate? Has it changed since the hostilities began in Iran?
   6. Independent verification: Are any AI-nominated targets independently checked by a person or unit that did not rely on the same system or data? If so, for which categories of targets?
   7. Policy constraints on compressed review: Is there a point at which the volume or tempo of AI-generated targets is judged too high for meaningful human review? Does this change based on the operational environment (e.g., urban areas), target class, or system employed? Who decides, and on what basis?
   8. Guardrails against cognitive offloading and deskilling: When AI tools generate target assessments and recommended courses of action, what programs or policies are in place within DoD to ensure human operators maintain the capacity to conduct target assessments independently? How will DoD ensure human commanders and their units can still generate recommended courses of action and evaluate targets against legal and operational standards without relying on AI outputs, and how is that capacity measured and maintained over time?
   9. Policies throughout the targeting cycle: To the extent AI systems are used, for example, for intelligence fusion, target nomination, target validation, collateral damage estimation, weaponeering, or battle damage assessment, what policies are in place to ensure appropriate use in each of these contexts? To what extent is human verification required at each of these stages, and what standards govern that verification?

What kinds of AI systems have been used during the recent conflict in Iran and what efforts are in place to mitigate and learn from instances of civilian harm?

In the first 24 hours of the U.S. attack on Iran in February 2026, U.S. forces carried out over 1,000 targeted strikes. To put that number in context, an official at the National Geospatial Intelligence Agency (NGA), the former home of U.S. AI targeting technology, reported that prior to AI systems, analysts could generate under 100 targets a day. With the integration of large-language models, provided by services such as Claude or ChatGPT, that number has risen to around 5,000 and is only increasing. 

The Iran War has become what the Chief Technology Officer of Palantir calls the first major AI-driven conflict. On February 28, 2026, the opening day of the war, two Tomahawk missiles struck an elementary school in the Iranian town of Minab. The strike killed over 150 people, including 123 children. The decision to strike the Minab school was reportedly the result of flawed intelligence inputs in Maven Smart System. This faulty intelligence moved along the kill chain unchecked amid a system that incentivizes speed and, as a result, overreliance on AI technologies. 

Reporting on the Minab school strike raises concerns about outdated targeting information and how it can move through the kill chain unverified, leading to a failed strike. A separate incident in which a falsely generated intelligence report spotted a Chinese ship transporting nuclear components raises serious questions about errors in AI-generated intelligence. When targeting errors involve several people and systems, it is unclear how the Department assigns responsibility and investigates failures. Key questions include:

1. What has the Department’s investigation into the Minab school strike found about the role of human-machine teaming, operator understanding of the AI tools they are employing, independent verification of AI nominated targets, the age of the imagery used, and the reliability of information in the targeting database? Will the findings be provided to Congress in full? Will the findings be publicly released?
2. How many targets struck in operations in Iran were first nominated or prioritized by an AI system? What share of approved targets does that represent?
3. How are no-strike lists and protected-site data (schools, hospitals, places of worship) integrated into AI targeting tools? How often are they updated, and by whom?
4. Does the Department ensure there is a common operating picture (COP) ensuring relevant operational information integrating data used in MSS (or other DSS or AWS systems) tracks non-combatants, critical infrastructure, and humanitarian activities alongside enemy activity and other targeting data? 
5. When an AI-assisted strike mistakenly or erroneously causes civilian harm, who in the chain of command is held responsible for the targeting decision: the approving officer, the commander, the system’s program office, or others?
6. Do civilian harm investigations record whether an AI tool contributed to the targeting decision (including intelligence and assumptions feeding into that decision) and how? Is that recorded in a way that allows patterns to be identified across incidents?
7. Do AI targeting systems keep logs sufficient to reconstruct, after the fact, what the system recommended, what data it used, what the human reviewer saw, and what the human reviewer approved?
8. Does the Department view systemic failures, for example, a database that does not receive updates, as a matter for command accountability, and if so, at what level?
9. How are lessons from after-action reviews of AI-assisted strikes fed back to system developers and to the units using them? How long does that take?
10. Does the Department publish or share with Congress its methodology for assessing civilian casualties in AI-assisted operations, including how it treats reports from outside sources?

What role do private commercial partners play?

U.S. targeting systems depend on private development and continuing involvement from companies like Palantir and Anduril that build AI-DSS. Private, third-party integrations within Palantir’s MSS, such as its use of Anthropic’s Claude, have been critical to the system’s development. 

1. What roles do private companies and their employees play in operating, maintaining, or updating targeting systems? How well do DoD operators, commanders, and operational lawyers understand the AI-DSS they employ – how much training is required and what testing, red-teaming, or other preparatory measures are taken prior to deployment?
2. It has been widely reported that U.S. targeting systems depend on private development and continuing involvement from companies like Palantir and Anduril. What access do such third-party vendors have to operational and targeting data, and what safeguards govern that access?
3. Following the designation of Anthropic as a supply chain risk in March 2026, to what extent can other AI companies that contract with DoD impose usage restrictions or technical safety limits on the systems the military deploys?

What role can international cooperation play in providing safeguards?

1. Recent reporting describes U.S. and Russian objections to proposed international safeguards for autonomous weapons during the Convention on Certain Conventional Weapons (CCW) Group of Governmental Experts process. Why did the United States make such objections to including language on the predictability and reliability of AI targeting systems?
2. What is the U.S. position on requiring human review of AI-generated targets before a strike? If the United States opposed that requirement during CCW deliberations in Geneva, why, and is that position consistent with Department policy?
3. What constraints on the use of AI in targeting would the Department support including in a joint statement or international agreement? 

What can we expect from the AI Acceleration Strategy and announcement of the Autonomous Warfare Command?

1. What does the Department’s January 2026 AI Acceleration Strategy, declaring the U.S military an “AI-first warfighting force,” mean for targeting? 
2. How does the recent announcement of DoD’s effort to create a new Autonomous Warfare Command impact systems development and control for AI targeting operations?
3. How will the new Autonomous Warfare Command incorporate civilian protection in its training, doctrine, and command structure from the ground up?
4. Which new AI capabilities for target development or engagement are planned over the next five years, and what safeguards will accompany them?