The Pentagon's AI War Machine: How Claude and ChatGPT Ended Up on the Front Lines
The use of artificial intelligence in warfare has transitioned from science fiction to reality. In recent, high-profile instances, the US military deployed AI, notably Anthropic's model Claude, for critical operations, including the capture of Venezuelan President Nicholas Maduro and during conflicts involving Israel, Iran, and the United States.
However, this partnership was fraught with tension. The US government and Anthropic clashed over a proposed contract that would have permitted autonomous AI weapon control and mass surveillance of US citizens. Anthropic's refusal led to their exclusion from government contracts, creating a vacuum that a surprising competitor rushed to fill. Just hours after the fallout, Sam Altman announced that his company, OpenAI, would accept the military's terms, marking a definitive shift from its non-profit origins to a for-profit entity deeply integrated with military operations.
This move triggered a significant public boycott of ChatGPT, as hundreds of thousands of users unsubscribed in protest. This story uncovers how AI became a tool of war just three years after entering the mainstream, how it functions in a military context, and the profound implications for society.
Military-Grade AI vs. Your Chatbot
A common question is how an AI like ChatGPT, which can struggle with simple concepts, could be reliable for identifying military targets. The answer is simple: the AI available to the public is not the same as the one used by the military.
According to Anthropic CEO Dario Amodei, the company was the first to create custom models for national security purposes. The key differences lie in the model itself and the hardware it operates on.
- Customized Models: Anthropic developed a unique model specifically for the military, running it within a secure data center filled with classified information.
- Internal Controls: Researchers at Anthropic discovered how to peer inside the "black box" of their neural networks. This allowed them to understand and control the AI's ethics and behavior mathematically, essentially placing a "speed limit" on personality changes to ensure stability.
- Dedicated Hardware: Military AI runs on dedicated hardware, with 100% of the available computing power allocated to a single customer. This can mean an entire data center is used for a single request, offering performance that is five or six orders of magnitude greater than what consumer AI, which shares resources among millions of users, can achieve.
Despite these advantages, the technology is not perfect. Paul Scharre, Executive Vice President at the Center for a New American Security, emphasizes, "AI gets it wrong. We still need humans to check the output of generative AI when the stakes are life and death." Even Amodei concedes that the technology for fully autonomous weapons is "not ready." Whether OpenAI can replicate the performance of Anthropic's custom model remains to be seen, but the risk of error in life-or-death situations is immense.
How AI Works in Modern Warfare
The Maven Smart System, built by Palantir and powered by a custom version of Anthropic's Claude, was instrumental in a 24-hour operation in Iran where 1,000 targets were struck. This system demonstrates the practical application of AI in war.
Maven gathers and processes vast amounts of classified data from satellites, surveillance, and other intelligence sources. For example, the Israeli government reportedly hacked nearly all traffic cameras in Tehran to monitor movement. The system synthesized this information from 179 different data sources, using Claude to interpret the data and provide actionable insights.
In the opening days of the war, the system organized and prioritized Iranian targets in real-time, issuing "precise location coordinates and prioritized those targets according to importance." A Georgetown University study found that the system enabled a single artillery unit of 20 people to perform the work of 2,000 staff, turning weeks-long battle planning into a real-time operation.
This same Maven-Claude hybrid was allegedly integral to the capture of President Maduro in Venezuela, an operation Anthropic claimed it was not aware its software was being used for. This event became the flashpoint in the conflict between the company and the US government.
The Fallout Between Anthropic and the Pentagon
The trouble began in July 2025, when Anthropic signed a $200 million contract with the Pentagon via Palantir. The Pentagon demanded that Claude be usable for "all lawful purposes" without any private-sector restrictions. This was a critical point of contention, as existing laws like the Patriot Act could legally permit mass domestic surveillance.
Anthropic pushed back, demanding two safeguards be written into the contract:
- A prohibition on the mass surveillance of Americans.
- A requirement for a human to be in the loop for any lethal decisions made by autonomous weapons.
On February 24th, the government issued an ultimatum: comply by 5:01 p.m. that Friday or face consequences. Dario Amodei refused, stating, "We cannot in good conscience acquiesce."
The government's response was swift and severe. It banned its agencies from using Anthropic's tools, giving them six months to phase them out, and labeled the company a "supply chain risk to the United States"—a designation never before used for an American company.
OpenAI's Opportunism and the Public Backlash
In a move that stunned observers, Sam Altman announced just hours later that OpenAI would take the deal Anthropic had rejected. This decision effectively secured OpenAI's financial future with government funding, insulating it from a declining enterprise market share.
Altman claimed on X that OpenAI's deal included the same safeguards Anthropic had fought for, but given the deal was negotiated in just two days, many were skeptical. The public reaction was intensely negative, coalescing into the Quit GPT movement. Organizers claim that over 2.5 million people have taken action by canceling subscriptions, sharing boycott messages, or signing petitions.
The Larger Threat: AI as a Surveillance Tool
The conflict highlights a disturbing potential use case for AI: mass surveillance. The proposed contract would have allowed Claude to analyze geolocated data, web browsing history, and personal finance information purchased from data brokers.
This practice—government agencies buying data on citizens from private firms—is not illegal, but it was never practical on a mass scale before AI. This threat is not limited to government contracts.
- Palantir has already installed surveillance systems in Australian supermarkets.
- Meta's Ray-Ban glasses feature a planned facial recognition system called "Name Tag," which could identify individuals in public using their public Meta profile. Meta reportedly believes the current political environment is opportune for this launch, as civil society groups are focused on other concerns.
What Can Be Done?
Using AI for war and surveillance represents one of the most alarming abuses of the technology. While individual uses of AI can be productive, the aggregate potential for misuse is catastrophic.
There are practical steps individuals can take to protect themselves:
- Remove your data from data broker sites. The US military's surveillance plan relied on purchasing data from these networks. Reducing your presence on them limits your exposure.
- Advocate for new laws. Sign petitions, raise awareness, and communicate that this direction is unacceptable.
Ultimately, the rapid progression of AI from a tool of convenience to a weapon of war and surveillance demands a critical public conversation about who this technology truly serves and what safeguards are necessary to prevent its worst-case applications.