The dispute between the Pentagon and Anthropic raises questions about the role of artificial intelligence in warfare. What is hidden behind the scenes of the conflict?
An unusual conflict has arisen between the Pentagon and the American artificial intelligence company Anthropic, the developer of the Claude language model (AI model). The US administration has included the company on a list of entities deemed a risk to defense supply chains – a rare decision usually reserved for firms from rival states such as Huawei.
The executive director of Anthropic, Dario Amodei/FOTO:X
The move could lead not only to the cancellation of a military contract worth about $200 million, but also to the phasing out of the Claude model from the US defense technology ecosystem. War Department contractors were to completely phase out use of the system within six months.
The origin of the dispute
Tensions rose after the Pentagon reportedly requested expanded access to the Claude model, requesting the ability to use it “for any lawful purpose” without the restrictions imposed by the company. Anthropic's chief executive, Dario Amodei, publicly denied the request.
He cited two major ethical limits: the use of artificial intelligence for mass population surveillance and the integration of AI models into weapon systems capable of selecting and hitting targets without human intervention.
Pentagon officials, including official Emil Michael, criticized the company's position, accusing Anthropic management of putting corporate principles before national security.
A larger debate about military artificial intelligence
At first glance, the dispute appears to be a typical one between the state and the tech industry: the military wants wider access to artificial intelligence tools, while developers try to limit how they are used.
But the stake is much higher. Advanced linguistic models – such as GPT‑5.2, Claude Sonnet 4 or Gemini 3 Flash – can analyze complex strategic scenarios, including simulations of nuclear crises, at a level of detail difficult to achieve by traditional human analysis, writes Oleksii Kostenko, head of the Scientific Laboratory of Immersive Technologies and Law of the Institute of Information, Security and Law of the National Academy of Legal Sciences of Ukraine, in Zerkalo Nedeli.
Meanwhile, other tech companies are taking a different approach. OpenAI, the developer of ChatGPT, announced that it has reached an agreement with the Department of Defense to use AI models in classified military networks, while maintaining bans on mass surveillance and autonomous weapons.
Also, the company xAI, founded by Elon Musk, signed an agreement to use the Grok model (AI model) in secure government systems, although some experts say that it cannot completely replace Claude.
What nuclear crisis simulations show
Military interest in these models is also linked to the results of recent academic research. In a study at King's College London, Professor Kenneth Payne analyzed how AI models make decisions in a nuclear crisis simulation.
In the experiment, AI systems were programmed to play the role of leaders of fictitious states and react to an international crisis. Models could choose between diplomacy, show of force, military escalation, or the use of nuclear weapons.
The results attracted the attention of researchers:
-in approximately 95% of the simulations, the models escalated the conflict to the level of tactical use of nuclear weapons;
-in 76% of cases, the escalation reached the level of strategic nuclear threat;
-de-escalation options were hardly used.
According to the researchers, the problem is not that algorithms “want” war. Rather, how these systems are optimized to gain a strategic advantage can sometimes push decisions toward higher risk options.
Time pressure and radical decisions
An important result of the study was the influence of time pressure on the behavior of the models. When the simulations included a deadline for achieving a strategic advantage, the level of escalation increased significantly.
For example, the GPT-5.2 model initially maintained a moderate level of tension in scenarios with no time limit. In deadline scenarios, however, his decisions became much more aggressive, approaching the level of a nuclear crisis.
Artificial intelligence as a strategic tool
For the military, the value of these systems lies not only in the logic of decisions, but also in their ability to quickly simulate the consequences of actions. AI models can simultaneously analyze the military, economic, environmental and political impact of a conflict, generating thousands of alternative scenarios.
This type of analysis could transform military planning: what used to require separate reports and lengthy analysis can now be done in real time.
Global competition
The debate in the United States also has a geopolitical dimension. Research indicates that China is accelerating the development of military technologies based on artificial intelligence, including military training systems and AI-assisted command platforms.
In this context, the dispute between the Pentagon and Anthropic reflects a larger question: how tightly the use of artificial intelligence in the military should be controlled.
Existing studies do not suggest that artificial intelligence should decide the use of nuclear weapons. However, they show that these systems can quickly become important tools in strategic analysis, military planning and risk assessment.
For many experts, the real challenge is not whether AI can think strategically, but who sets the limits of how these systems will be used.
In this sense, the dispute between the Pentagon and Anthropic could represent only the beginning of a global debate about the role of artificial intelligence in the wars of the future.
