Patricia Brennan
The Washington Institute for Near East Policy
Dimasheq Editorial Team
The conflict has demonstrated how AI-enabled warfare can accelerate the “kill chain.” However, human analysts must still know when to question machine recommendations to avoid harmful operational and strategic consequences.
During Operation Epic Fury, US forces reportedly struck more than 1,000 targets within the first 24 hours and more than 13,000 targets over 38 days. According to US Central Command Commander Admiral Brad Cooper, the United States achieved its military objectives in fewer than 40 days by “undermining Iran’s ability to project power beyond its borders and threaten the region and our interests.” He also noted that integrating artificial intelligence into CENTCOM operations helped reduce tasks that previously took hours or days to mere seconds, highlighting the potential operational advantages of AI-enabled systems.
These advantages extended beyond accelerating target identification to rapidly delivering critical intelligence about threats to US aircraft operating in hostile airspace. Major General Michael Conley, commander of US Air Force Special Operations, praised AI’s role in reducing the time required to deliver intelligence to the cockpit from approximately 20 or 30 minutes to just two or three seconds, enhancing the safety of US crews facing Iranian threats.
For decades, the US military has invested in technologies designed to reduce the human effort and time required to identify targets and determine their geographic locations, thereby accelerating the “kill chain.” The war with Iran has demonstrated the potential operational returns of these investments, as AI-enabled systems integrated information from hundreds of intelligence sources to generate targeting recommendations at unprecedented speed and scale.
Palantir’s Maven Smart System was central to these capabilities, bringing together intelligence from various sources and platforms to identify targets and prioritize them. Anthropic’s Claude added further reasoning capabilities within the Maven system, while Anduril’s Lattice integrated sensor data to support battlefield situational awareness and autonomous operations.
The use of artificial intelligence in warfare is not new to the Middle East. Israel has relied for years on systems such as Gospel, Lavender, and other AI-enabled targeting tools to gather intelligence and propose targets. More recently, reports indicated that US Central Command had used a tool similar to Maven to help Saudi Arabia identify Houthi force targets in Yemen and share intelligence on them. CENTCOM has also announced the establishment of its first joint military task force with the United Arab Emirates focused on artificial intelligence.
These developments demonstrate that AI is being used increasingly in military operations across the region, underscoring the importance of applying the lessons learned from Operation Epic Fury more broadly.
Nevertheless, one of the most important lessons is the need for a more rigorous examination of AI’s role in the targeting process before expanding the integration of these advanced capabilities into US and partner operations.
It remains unclear how accurately AI performed during the most intense phases of the war, how effective the strikes were, or how heavily human operators relied on AI. Regardless of the specific outcome of this conflict, conducting targeting operations at machine speed carries a general risk: human decision-makers may become either bottlenecks that slow the process or mere rubber stamps that formally approve machine recommendations. Both scenarios could have negative operational and strategic consequences.
This dilemma is not limited to individual targeting decisions; it may also extend to the way Washington conducts its wars. AI may give US forces the ability to strike thousands of targets at an accelerated pace, but an increased capacity to target does not necessarily translate into greater strategic impact.
If available targets begin to determine what should be struck, the targeting process could end up shaping strategy rather than serving it. This issue is particularly important during sustained military campaigns, when limited munitions must be prioritized for targets that directly contribute to achieving US objectives.
The Risks of Excessive Reliance on Artificial Intelligence
As US Central Command expands its cooperation with regional countries in the field of artificial intelligence, the challenge will extend beyond determining which technologies should be transferred to partners and when. It will also involve ensuring that US forces and partner forces are prepared to critically assess AI-generated recommendations.
This issue is particularly important because the speed of AI information processing leaves analysts with less time—and potentially less motivation—to scrutinize machine recommendations. Under the pressures of war, delays caused by additional human review may encourage personnel to approve AI recommendations automatically rather than treating human oversight as an additional step of verification and scrutiny.
One way to prevent this is to ensure that analysts do not confuse two distinct concepts: machine confidence and confidence in the intelligence underlying a target assessment.
Under traditional methods, analysts build confidence in a target by reviewing previous and current intelligence and requesting additional collection to develop a comprehensive understanding of it. AI-generated recommendations, however, change this approach. Instead of building confidence in the target themselves, analysts receive a ready-made recommendation and are asked to verify its accuracy.
Building confidence in a target and subsequently verifying the reliability of a machine’s assessments are two different cognitive processes. The danger lies in confusing confidence in the machine with having sufficient and reliable information about the target itself.
This distinction matters because AI may process information correctly yet produce an incorrect recommendation if the underlying data is outdated or deliberately manipulated.
Consider, for example, the US strike on a girls’ school in Iran in February, which killed more than 160 people. Reports indicate that the site remained listed as a military target in the Defense Intelligence Agency’s database, and that imagery may not have clearly shown that part of the site had been converted to civilian use.
In addition to the human casualties, the strike undermined the US strategic message and gave Tehran an opportunity to claim that US forces had committed a war crime.
Although the investigation remains ongoing and the role of AI in this tragedy has not yet been established, the incident highlights a potential vulnerability: AI-enabled targeting could amplify errors by accelerating the kill chain before analysts have an opportunity to detect outdated or conflicting intelligence.
Achieving the right balance between human scrutiny and the use of artificial intelligence is therefore essential to ensuring the reliability of target identification without sacrificing the speed advantage AI provides.
Striking the Right Balance in Human-Machine Collaboration
Achieving this balance requires ensuring that trust in AI-enabled systems is proportionate to their actual capabilities.
In research on human interaction with automated systems and artificial intelligence, the concept of “calibrated trust” refers to the extent to which human operators rely on a system’s outputs based on their assessment of its capabilities and reliability.
In the military context, this means knowing when to rely on AI-generated targeting recommendations, when to question its outputs, and when to override them.
Overreliance may lead to targeting errors that undermine the United States’ broader political objectives, while insufficient reliance may result in missed opportunities.
In September, for example, US forces nearly intercepted a Chinese commercial vessel in the Middle East after an analyst relied on an incorrect assessment generated by an AI system, which concluded that the vessel’s cargo was linked to nuclear weapons. US aircraft had already been dispatched to the location, and personnel were preparing to board the vessel before more experienced analysts identified the error.
Accordingly, the level of human scrutiny must be proportionate to the nature of the target and the potential consequences of misidentifying it.
For example, an excavator at an Iranian nuclear facility whose status has recently been verified may require less scrutiny than a target located in a populated area for which the underlying intelligence has not been reviewed recently.
The United States has invested heavily in accelerating the kill chain, but recent conflicts in the Middle East demonstrate the need to devote equal attention to training the human personnel involved in this process so that they can keep pace with technological developments and exercise appropriate oversight over AI systems.
Although faster targeting cycles increase the speed and scale of military operations, human analysts must still understand the sources underpinning AI recommendations, quickly identify conflicting or fabricated information, and recognize when additional intelligence collection or verification is necessary.
These lessons also apply to any potential conflict with China, which possesses the capabilities needed to challenge US capabilities, sensors, and networks.
This could include deceiving AI-enabled targeting systems on a broader scale and through more sophisticated methods, such as creating false tracks, spoofing signals, deploying decoy targets, and using deepfake technologies.
Anthropic recently noted that AI models remain weak at detecting spoofed signals. This means adversaries do not necessarily need to hack US targeting systems directly if they can continuously manipulate the information those systems receive.
In light of these concerns, the standard for evaluating the effectiveness of AI-assisted targeting should not be limited to whether a human is involved in the decision-making process. It should also assess whether that person can distinguish between situations in which AI recommendations can be trusted and those that require skepticism.
Successful human-machine collaboration depends on preserving human judgment without unnecessarily slowing operations. This requires, among other things, understanding when the potential consequences of an error warrant slowing down the kill chain and when they do not.
Policy Recommendations
Drawing on the experience gained by US forces during Operation Epic Fury and its aftermath, US defense officials should develop a comprehensive strategy for AI adoption that strengthens human judgment, advances technical safeguards, continuously evaluates system performance, and incorporates lessons learned from partner experiences.
This strategy should focus on the following priorities:
1. Assessing AI Performance in the War with Iran
The US Department of Defense should commission independent external reviews of AI systems’ performance in combat operations. These reviews should examine the accuracy of AI-generated targeting recommendations and the extent to which human personnel relied on or challenged their outputs.
2. Strengthening AI Literacy Across the Armed Forces
The Department of Defense should incorporate AI literacy into formal training programs for analysts and provide recurring opportunities to work with this technology under realistic conditions.
Training exercises should include outdated intelligence, adversarial deception attempts, degraded sensor performance, and algorithms with varying levels of reliability. Analysts should be required to assess their performance and determine the appropriate degree of reliance on AI.
3. Developing Models for Cross-Validation Between Machines
Future targeting systems should incorporate independently developed models and sensors to detect inconsistencies or unusual patterns in AI-generated recommendations.
These models should be tested, and their results compared with human performance and human-machine teams to determine whether independent automated verification improves the ability to detect errors.
Such verification should reinforce human scrutiny rather than replace it, providing an additional layer of assurance.
4. Expanding Dialogue with Allies and Partners
US Central Command should organize regular dialogues across the Middle East to exchange lessons learned and build on existing AI cooperation with Israel, Saudi Arabia, and the United Arab Emirates.
The Department of Defense should also expand these discussions globally to develop common approaches to AI-enabled targeting, grounded in established best practices for human-machine collaboration and consistent with the law of armed conflict.
5. Integrating Additional Safeguards into AI Systems
The Department of Defense should adopt technical safeguards capable of detecting data that may be outdated or fabricated, clearly identifying the portions of final intelligence reports generated by automated systems, and requiring independent verification of target locations and the sensor data underpinning AI-generated targeting recommendations.
Patricia Brennan is the US Air Force’s 2026–2027 Fellow at The Washington Institute. Before joining the Institute, she commanded the 303rd Intelligence Squadron at Osan Air Base in South Korea.
Note: The views, conclusions, and recommendations expressed in this article are those of the author and do not represent the official positions or policies of the US Air Force or any other US government agency.







