A Defense Department official knowledgeable about the issue indicated that the US military might leverage generative AI systems to create ranked lists of potential targets and provide recommendations—subject to human review—regarding which targets to prioritize for strikes. This revelation regarding the military’s potential use of AI chatbots coincides with ongoing scrutiny of a strike on an Iranian school that is still under investigation by the Pentagon.
Potential target lists could be input into a generative AI system being developed for classified environments by the Pentagon. The official, who requested anonymity to share sensitive information with MIT Technology Review, noted that humans may instruct the system to process this information and rank the targets, considering variables such as the current location of aircraft. Ultimately, it would be up to humans to verify and assess the outputs and suggestions. In theoretical terms, OpenAI’s ChatGPT and xAI’s Grok could serve as models for this scenario in the future, as both companies recently secured agreements for their models to be utilized by the Pentagon in classified situations.
The official provided this as a model of how operations might function but refrained from confirming whether it reflects the current use of AI systems.
Other sources have stated that Anthropic’s Claude has been incorporated into current military AI systems and utilized in missions in Iran and Venezuela, but the official’s remarks provide insight into the particular function chatbots may fulfill, especially in expediting target identification. They also reveal the military’s employment of two distinct AI technologies, each of which has its own constraints.
Since at least 2017, the US military has been pursuing a “big data” program named Maven. This initiative utilizes older AI techniques, particularly in computer vision, to process the vast amounts of data and imagery gathered by the Pentagon. For instance, Maven could process countless hours of aerial drone videos and algorithmically identify targets. A 2024 report from Georgetown University indicated that soldiers were using the system to choose and vet targets, which accelerated the approval process for these targets. Soldiers engaged with Maven via an interface featuring a battlefield map and dashboard, which could color-code potential targets and friendly forces differently.
The remarks from the official imply that generative AI is being introduced as an interactive chatbot layer—one that the military might employ to swiftly find and analyze data as it determines which targets to focus on.
Generative AI technologies, similar to those that form the basis of ChatGPT, Claude, and Grok, represent a fundamentally different approach from the AI that has primarily supported Maven. Underpinned by large language models, they are significantly less battle-tested. While Maven’s interface compelled users to carefully examine and interpret data depicted on the map, the results generated by generative AI models are more readily accessible but more challenging to verify.
The application of generative AI for such decision-making is decreasing the time needed for the targeting process, the official added, though further specifics were not provided when inquiring about how much additional speed could be achieved if humans needed time to confirm a model’s outputs.
The deployment of military AI systems is facing intensified public scrutiny following the recent attack on a girls’ school in Iran that resulted in the deaths of over 100 children. Several news sources have reported that the strike originated from a US missile, although the Pentagon has stated it is still under investigation. Furthermore, while the Washington Post has noted that both Claude and Maven have been involved in targeting decisions in Iran, there is currently no evidence clarifying the role of generative AI systems, if any. The New York Times reported on Wednesday that initial investigations found outdated targeting information contributed to the strike.
The Pentagon has been increasing its utilization of AI across various operations in recent months. Since December, it began providing non-classified access to generative AI models for tasks such as contract analysis or presentation writing to millions of service members through an initiative called GenAI.mil. However, only a limited number of generative AI models have been approved by the Pentagon for classified purposes.
The first of these was Anthropic’s Claude, which, in addition to its application in Iran, was reportedly employed during operations targeting Venezuelan leader Nicolas Maduro in January. However, due to recent disagreements between the Pentagon and Anthropic regarding restrictions on the military’s use of its AI, the Defense Department labeled the company as a supply chain risk, and President Trump urged on social media for the government to cease using its AI products within six months. Anthropic is currently challenging this designation in court.
OpenAI announced an agreement on February 28 allowing the military to utilize its technologies in classified contexts. Elon Musk’s xAI has similarly finalized a deal for the Pentagon to employ its model Grok in such environments. OpenAI has stated that its agreement with the Pentagon includes certain limitations, although the actual effectiveness of these limitations is still uncertain.
For anyone with information regarding the military’s application of AI, secure communication can be done via Signal (username jamesodonnell.22).
