The AI in Your Pocket and the AI in the Kill Chain: How Different Are They?

Commentary

The poster appeared on walls throughout the War Department in December 2025. Secretary of War Pete Hegseth, pointing directly at you, Uncle Sam style. The slogan read, “I want you to use AI.”

Within a week, 550,000 military and civilian defense personnel had logged on. Soon that number reached 1 million. The same tools that live on your phone—ChatGPT, Gemini, Grok—were repackaged and deployed across 3 million government desks.

Then, on May 1, the Pentagon went further. It announced agreements with eight companies to deploy artificial intelligence inside its most classified military networks. The eight companies are household names already woven into the daily lives of ordinary Americans: Google, OpenAI, Microsoft, Amazon, Nvidia, SpaceX, Reflection, and Oracle.

This raises a question that more people should be asking.

The military has a term for the sequence of steps from detecting a target to destroying it, calling it the “kill chain.” AI is now embedded inside it. The same companies whose tools answer your questions, draft your emails, access your files, and help your children with their homework are now contracted to operate inside the most classified kill chain infrastructure in the United States.

So how different is the AI in your pocket from the AI in the kill chain?

The AI most people use is called a large language model (LLM). It reads patterns in human language and produces responses that sound authoritative, fluent, and certain. It does not understand what it says. It generates what is statistically most likely to follow. The AI now embedded in military systems begins with the same foundational technology. It is built by the same companies, trained on the same class of architecture, and speaks in the same authoritative voice that most people have learned not to question.

Consumer AI was designed to feel helpful and familiar. Military AI inherited that same type of confidence that gives the sense that the system knows. One version tells you the best restaurant nearby. The other recommends which targets to strike and which munitions to assign. And the gap between those two purposes is narrower than most people imagine.

Project Maven is one system that most Americans have never heard of, but that has been running since 2017.

It began as a computer vision tool. It doesn’t read words, it watches. It processes drone footage, satellite imagery, and sensor feeds. It identifies objects, tracks movement, and flags what it finds. It was built because the military was collecting so much data and so much video that it didn’t have enough human eyes to watch it all.

Over time, Maven became an AI-assisted targeting and battlefield management system. It fuses the watching layer with the language layer—with computer vision identifying targets and LLMs synthesizing intelligence and recommending actions—and connects both to operational systems.

That fusion produced a new level of efficiency.

Before AI, analysts could process fewer than 100 potential targets per day. With computer vision, it grew to 1,000. And after LLMs were integrated, the number of potential targets per day increased to 5,000.

In March 2026, the Maven Smart System was elevated to an official Pentagon program of record—permanent, funded, and embedded in the architecture of American warfare for the foreseeable future.

So back to the comparison.

The AI on your phone goes beyond the apps you may be using to search or generate text. The AI on your phone also watches you. Face ID reads your face. Google Maps tracks your position in real time. Your camera identifies people, objects, and locations. Instagram’s camera identifies what you’re looking at to serve relevant ads. These are not language models; they are computer vision systems, the same category of AI as the watching layer inside Maven.

So the distinction is not what goes in. Both watch. Both read. Both track. The distinction is what the watching is connected to, and what it is authorized to do.

On your phone, the watching layer unlocks your screen, serves you a relevant ad, or suggests a caption. In Maven, the same category of watching fuses with language models, connects to weapons systems, and produces targeting recommendations sent to battlefield commanders.

Same capability. Different authorization. Different consequences.

What separates the AI in your pocket from the AI in the kill chain is not the technology. It is the instruction. It is the integration. It is what the system has been authorized to do with what it finds.

One company is notably absent from the Pentagon’s new list of AI partners.

Anthropic, maker of the popular Claude AI, refused to allow its technology to be used for autonomous weapons or to surveil American citizens. Recently, I wrote about Mrinank Sharma, who led Anthropic’s Safeguards Research Team, resigning with the words “the world is in peril.” The tension he signaled from inside that company became public months later when the Pentagon designated Anthropic a supply chain risk and the Trump administration ordered federal agencies to remove its products. A federal judge has since issued an injunction blocking those actions, for now.

Anthropic drew a line around two specific authorizations: autonomous weapons and the surveillance of American citizens. The fact that those two things had to be stated as conditions, and that stating them came at a cost, tells you what the architecture is capable of.

So when you next open your phone, remember that you are holding a device whose AI is supplied by the same companies now contracted to operate inside the most classified military networks in the United States. The question we deserve to know isn’t what those companies agreed to, but what they didn’t insist on refusing.

Views expressed in this article are the opinions of the author and do not necessarily reflect the views of The Epoch Times.