The American company Anthropic wants to impose limits on the American army for the use of its chatbot Claude, which the Trump administration refuses. France 24 explains to you why these red lines are both essential and insufficient.
Published on: 02/27/2026 - 6:28 p.m. Modified on: 02/27/2026 - 10:18 p.m.
The Pentagon wants to be able to put AI everywhere in the army, Anthropic a little less. The clash between the US War Department and OpenAI's competitor could decide the extent of the US military's use of chatbots and AI.
For the creator of Claude, the clock is ticking. War Secretary Pete Hegseth ordered Anthropic to relent by 5:01 p.m. ET on Friday. Faced with the company's stated determination to stand up, Donald Trump ordered his administration to no longer use Claude.
“I am ordering ALL federal agencies of the US government to IMMEDIATELY STOP all use of Anthropic’s technology. We don’t need it, we don’t want it, and we will no longer work with them,” thundered the US president on his Truth Social platform.
This first sanction might not be the last. The War Department further threatened Anthropic with being labeled a U.S. “supply chain [security] risk,” which would disqualify it from any contract with the U.S. government.
Two red lines
On paper, the Pentagon is only asking all AI vendors to allow the military to use chatbots “for any lawful purpose.” But Anthropic believes that this wording does not prevent the army from violating its two red lines in military matters.
The company does not want Claude to be used for large-scale surveillance of American citizens and for fully autonomous weapons, such as drones without a human pilot.
Anthropic's rebellion may seem surprising. After all, the army has already been using Claude since last July. The agreement makes it possible to integrate this chatbot into the confidential communications networks of the American government and "use it for the collection and processing of intelligence", specifies Sofia Romansky, coordinator of the study project on the responsible use of AI in the military domain at The Hague Center for Strategic Studies.
Read alsoNantianmen project: science fiction, a new ally of the Chinese military
Anthropic, however, seems to have been taken by surprise by the use of Claude in the context of the capture of Venezuelan President Nicolas Maduro in January. The company would have liked to know more about how its AI was used during this controversial operation by the American army and whether this use remained consistent with its ethical rules.
It was shortly after this first little hiccup in the relationship between the Pentagon and Anthropic that Pete Hegseth demanded to be able to use these AIs "for all lawful purposes" for the army.
The standoff aims to determine who, the state or private companies, “has control over the legal and ethical frameworks for the military use of advanced data processing systems in generative AI models,” summarizes Emil Archambault, researcher and specialist in emerging technologies in the military field at the German Council on Foreign Relations (DGAP).
An AI that presses the nuclear button?
The outcome of this confrontation is all the more important as there has been a “significant change in the position of the United States” regarding the military use of AI with the return of Donald Trump to the White House, assures Anna Nadibaidze, specialist in military applications of artificial intelligence at the University of Southern Denmark. The new administration “has redefined the concept of responsible use of AI by considering, to simplify, that it would be irresponsible not to integrate this technology everywhere and without limits,” explains this expert.
In other words, nothing would theoretically prevent Pete Hegseth and Donald Trump from allowing Claude or ChatGPT to decide when and against whom to use nuclear weapons.
Hence the importance of these “red lines” that companies like Anthropic would like to enforce. They could fill a void because “the legal frameworks are not yet defined and there are no clearly established standards that govern how these technologies can be used by the military,” notes Emil Archambault.
The two limits set by Anthropic "represent a bit of a bare minimum," said Ingvild Bode, director of the Center for War Studies at the University of Southern Denmark and an expert on the role of artificial intelligence in international security issues.
See also Killer robots and autonomous weapons: should they be banned?
Even if it's not much, "it's a good thing that a private company reminds us of these red lines," she adds.
They cover two crucial elements, according to the experts interviewed by France 24. On the one hand, the refusal to let Claude participate in a program of mass surveillance of American citizens "touches on a central ethical principle which is that of respect for private life", underlines Anna Nadibaidze.
On the other hand, Anthropic does not believe that “current AI systems are capable of reliably performing the tasks required for the control of these autonomous weapons,” explains Ingvild Bode.
It is therefore a technical red line. These chatbots remain large black boxes, even for their creator: they “are made up of so many different parameters that there is great uncertainty about the final decision taken. There are therefore reliability problems that are difficult to reconcile with autonomous weapons,” notes Sofia Romansky.
For example, in the case of autonomous combat drones, no one can guarantee that the AI “will know the difference between a combatant and a civilian, or even distinguish an enemy soldier from an ally,” she emphasizes.
“Insufficient” limits
In this context, these red lines are as much questions of principle as a “way of preventing a possible legal risk”, underlines Ingvild Bode. Anthropic would have no desire to be held responsible for dramatic consequences on the ground: “If it was demonstrated that an AI error had led to the death of civilians, this could have very costly consequences for the company,” adds Emil Archambault.
If all the experts interviewed recognize that the existence of these two red lines is important, “they remain insufficient”, assures Sofia Romansky. Thus, the refusal to participate in a massive surveillance program “only concerns American citizens. What about foreign residents?”, asks this specialist. This red line is also very vague: from how many citizens concerned can we consider this to be mass surveillance?
The same goes for autonomous weapons. “The boundary between a 'fully' autonomous or 'partially' autonomous weapon is poorly defined,” underlines Emil Archambault. Even when there is human intervention, "we know that there is a tendency to trust too much in the advice of AI. When everything is going very quickly and humans are under pressure [as in the case of a drone strike for example, Editor's note], they may be tempted to follow the recommendations of the algorithm without necessarily checking for themselves," explains Anna Nadibaidze.
So the problem is that by setting two red lines that are as necessary as they are imperfect, Anthropic is suggesting that all other cases are acceptable. Even though the example of semi-autonomous weapons suggests that the use of AI is problematic.
Anthropic's inflexibility on these principles has, according to the experts interviewed, the advantage of casting a harsh light on the other AI giants who do not seem to have the same reluctance.
The Pentagon will still be able to rely on ChatGPT (OpenAI), Grok (Elon Musk's xAI) and Gemini (Google), if Anthropic does not give in. Except that this standoff seems to have awakened certain consciences: more than 100 Google employees signed an open letter calling on their managers to imitate Anthropic.
This fight also risks leaving its mark if, faced with Pete Hegseth and Donald Trump, Anthropic decides to give in after all. In this case, "smaller companies, which also wanted to have an ethical posture, could say to themselves that it is useless because they will never have the means to resist government pressure", estimates Anna Nadibaidze.
The entire American AI ecosystem would then risk bending to the will of the Trump administration.
Keywords associated with the article
The content you are trying to access does not exist or is no longer available.
