As artificial intelligence becomes entangled with questions of national security, OpenAI has set out, in a letter to Amnesty International, the limits it says it places on how its technology may be used by military and government clients.

A reply to mounting scrutiny

The letter, dated 12 May 2026, responds to questions Amnesty International had raised about OpenAI’s human rights policies and practices regarding generative AI in military contexts, following an earlier exchange dated 15 April 2026. OpenAI’s central argument is that democratic security and the protection of human rights are deeply interconnected, and that democratic governments need to understand both the capabilities and the risks of frontier AI systems. With appropriate safeguards and meaningful human oversight, the company argues, AI could help democratic societies protect people, deter aggression, and potentially reduce the likelihood of future conflict. Given that AI technologies are proliferating rapidly across a wide range of state and non-state actors, OpenAI frames the relevant question not as whether AI will be used in conflict, but how harm can be mitigated and legal and human rights standards upheld in an environment where such use is increasingly unavoidable.

Three red lines for the Pentagon

The company points to existing policy level restrictions and model level safeguards, including longstanding prohibitions on the use of its services to automate high stakes decisions in sensitive domains such as national security without appropriate human review, and prohibitions on using its services to facilitate illegal activity, including war crimes, crimes against humanity, genocide, torture or mass domestic surveillance. At the model level, the company says principles of human safety, human rights and human control are built into how its systems respond, shaping behaviour to avoid facilitating critical harms while reinforcing the importance of human oversight in consequential contexts.

For its work with the US Department of War, OpenAI describes three specific red lines written into its agreement: no use of its technology for mass domestic surveillance, no use to direct autonomous weapons systems, and no use for high-stakes automated decisions, alongside a blanket prohibition on illegal activity.

Layers of safeguards, or just words

OpenAI says these contractual commitments do not stand alone, but are reinforced by deployment limitations, safety controls, and oversight mechanisms, including embedded staff intended to identify and mitigate emerging risks, with a process to work with the Department to remediate any noncompliant activity identified. The company describes questions around AI and military applications as complex and evolving, not purely technical, and says it welcomes continued engagement with governments, civil society organisations, researchers, and industry.

Whether such assurances satisfy human rights groups is likely to depend on how they are tested in practice. Critics of AI’s growing role in defence have generally argued that self-imposed red lines are difficult for outside observers to verify and that embedded oversight arrangements can blur the line between independent scrutiny and cooperation with the client they are meant to watch. Supporters of closer cooperation between AI developers and democratic militaries counter that disengagement would simply cede the field to states and companies operating under fewer constraints. The exchange between OpenAI and Amnesty International is likely to be one of many as governments move to define how far, and under what conditions, AI can be allowed into the conduct of war.

Sources: OpenAI, letter to Amnesty International, 12 May 2026.