Pentagon and Anthropic dispute Claude amid plans to use AI for target selection

The Pentagon is discussing a more active role for generative AI in the military decision loop: such systems could help rank targets and suggest which ones to attack first. Simultaneously, the US Department of Defense has entered into a tough conflict with Anthropic, which refuses to remove restrictions from Claude for surveillance and fully autonomous weapons.

How This Could Work

According to MIT Technology Review, a Pentagon official familiar with the matter described a scenario in which a list of potential targets is loaded into a generative system operating in a classified environment. Military personnel could then ask the model to analyze the data, take context into account—such as aircraft locations—and provide priorities: what to consider an urgent target and what can be delayed. The key caveat is that final verification and decision remain with humans.

It's important to note that the official was discussing a possible scheme and did not confirm that such a process is already being used in exactly this form. But the description itself shows where US military infrastructure is heading. This is no longer just about recognizing objects in video or satellite imagery, but about a conversational interface layered on top of combat analytics, which can accelerate data retrieval and prepare recommendations.

From Maven to Chatbots

This story didn't emerge from nowhere. Since 2017, American military forces have used Project Maven—a system that analyzes large volumes of images and video, including drone footage, and helps identify potential targets. Previously, operators had to work through maps, dashboards, and visual markers.

Now, large language models like Claude, ChatGPT, or Grok can be layered on top of such infrastructure, turning data search and sorting into a dialogue. This transition has both benefits and problems. Generative models are more convenient for humans: you can simply ask which targets seem most prioritized and why.

But their responses are harder to verify than a map or table with raw data. Against the backdrop of strikes where questions later emerge about the source data and verification procedures, this shift becomes particularly sensitive.

• Project Maven has worked with computer vision and intelligence data for many years
• A generative layer adds a conversational interface and text recommendations
• In December 2025, the Pentagon launched GenAI.mil for millions of military personnel in non-combat and non-classified tasks
• For classified systems, only a few models have been approved so far

Why the Dispute Over Claude

The sharpest conflict emerged around Anthropic. According to US media reports, Claude could be used in conjunction with military systems in operations related to Iran and Venezuela. After this, the Pentagon demanded broader use conditions, and Anthropic refused to agree to vague formulations that, according to the company, would not prevent the model from being used for mass surveillance of Americans or for fully autonomous weapons systems.

"We cannot in good conscience agree to this," is how

Dario Amodei described the Pentagon's requirements.

The Pentagon's response was harsh: the company was classified as a supply chain risk, which could eliminate Claude from defense contracts. Against this backdrop, OpenAI publicly announced an agreement with the Pentagon on February 28, 2026, to work in classified environments, and xAI also received approval for Grok. Formally, everyone speaks about human control, but the dispute is no longer about whether AI is needed in the military decision chain, but about who exactly sets the restrictions and how binding they are.

What This Means

The main news here isn't that the military wants a "chatbot for war," but that generative AI is becoming another layer on top of already existing targeting and analysis systems. This accelerates the decision-making cycle, but at the same time blurs the line of accountability: if a model suggests target priority, it becomes increasingly difficult for humans to prove that they truly independently verified the recommendation rather than simply approved it.