Defense and international security

22.1AI, the new strategic weapon

In context

Palantir and defense contracts

The entry of generative AI into the armed forces happened largely through an intermediary: Palantir, an American data-analysis company whose Maven system serves as the digital backbone of many operations. It was via Palantir (and Amazon's infrastructure) that a first large model was integrated, in late 2024, into classified military networks, for intelligence analysis and planning support. In July 2025, the Pentagon awarded contracts (up to 200 million dollars each) to four frontier labs: Anthropic, OpenAI, Google and xAI. These deployments are already operational: systems coupling an AI model with Maven helped, in 2026, to identify and prioritize targets during real military operations, which starkly illustrates that the stakes are anything but theoretical.

It was in this context that a dispute broke out in early 2026. The Pentagon demanded access "for any lawful use," with no restriction imposed by the supplier, arguing that in urgent national-security scenarios it is for the law and the elected government, not a private company, to set the limits. But the labs did not all have the same policy: xAI accepted the broadest use, OpenAI and Google placed limits mainly on classified work, and Anthropic maintained two red lines (no fully autonomous weapons without human oversight, no mass surveillance of citizens). Faced with Anthropic's refusal to lift these two limits, the administration classified it as a "supply-chain risk" (a label usually reserved for foreign adversaries) and set about removing its model from defense systems; in the wake of this, OpenAI signed an agreement covering classified environments while claiming to retain similar prohibitions. A federal judge then granted an injunction, finding that the sanction amounted to retaliation. Beyond the twists and turns, the dispute crystallizes a question that will outlast this case: who should set the limits of military AI, the law and the elected government, or the usage charters of the private companies that design the models?

In context

The dual-use conundrum

One difficulty runs through all governance of military AI: most of the building blocks involved are dual-use, that is, inseparably civilian and military. The same vision model that sorts holiday photos can guide a drone; the same AI that accelerates drug discovery (Chapter 14) can, if diverted, help design a biological weapon (Chapter 24); the same chips that train an assistant equip a military compute center (Chapter 8). This ambivalence complicates everything: one cannot ban a "military" technology without affecting its legitimate civilian uses, which makes export controls (Chapter 8) and treaties delicate to calibrate. It also blurs the boundary of responsibilities: a "civilian" lab may see its models integrated into weapons systems (Section 22.1). Dual-use explains why the regulation of defense AI cannot be conceived separately from civilian AI: it is often the same technology, depending on how it is used.

22.2Lethal autonomous weapons

Diagram22.1. The place of the human in the decision. The further one slides to the right, the more the machine decides alone. For many, the ethical and legal red line lies at the move to full autonomy ("out of the loop"), where a machine chooses to kill without a human.

22.3Accountability, escalation and arms races

Debate

Three major risks stand out.

The dilution of accountability. As one researcher puts it, in a military context, "an unforeseen failure means a dead civilian." If an algorithm designates a target and a human approves it within seconds without sufficient information, who answers for the error? Autonomous weapons risk making the chain of accountability no longer merely hard to trace, but nonexistent.
Escalation through speed. Systems that decide at machine speed could trigger an escalation faster than humans can stop it, by accident or through automatic chaining, in the manner of "flash crashes" in finance (Chapter 18) but with lethal consequences.
The arms race. The fear that the adversary is advancing pushes each power to accelerate, creating a race dynamic that is hard to halt, echoing the rivalry over chips (Chapter 8) and the scenario in Chapter 24.
Opacity and the nuclear question. Two specific concerns add to this: the inexplicability of a learned system (why did it designate that target?), hardly compatible with the requirements of the law of war; and the integration of AI into nuclear command systems, where a broad consensus calls for keeping the decision strictly human.

22.4Toward international governance?

Key takeaways (Chapter 22)

All the major powers make AI a matter of military superiority (intelligence, cyber, logistics, decision support, drones), with contracts between states and labs.
Lethal autonomous weapons raise the question of the place of the human in the decision to kill ("in," "on" or "out of" the loop); a legal vacuum and a debate over who should set the limits (the law or the suppliers).
Three risks: the dilution of accountability, escalation through speed, and the arms race.
International governance is being attempted at the UN (Geneva), but runs up against the definition of autonomous weapons and the logic of the race. An urgent and unresolved challenge.

So ends Part V, which has surveyed the major real-world domains transformed by AI. Part VI, the last of the course, rises toward the ultimate questions: philosophical, existential and political.

22.1AI, the new strategic weapon#

22.2Lethal autonomous weapons#

22.3Accountability, escalation and arms races#

22.4Toward international governance?#

Key takeaways (Chapter 22)

22.1AI, the new strategic weapon

22.2Lethal autonomous weapons

22.3Accountability, escalation and arms races

22.4Toward international governance?