In a major win for AI safety advocates, OpenAI has successfully forced the Pentagon to legally accept the “human-in-the-loop” principle as a condition of its new contract. This means that no OpenAI model can be used to make a “high-stakes” decision—specifically those involving the use of force—without direct human approval. This safeguard is a central pillar of OpenAI’s mission to ensure that artificial intelligence remains a tool for humans rather than a replacement for human judgment.
The battle for this clause was intense. Earlier in the month, Pentagon officials had pushed for “unfettered access” to AI models, arguing that the speed of modern warfare requires autonomous decision-making. Anthropic’s refusal to budge on this point led to their federal ban. OpenAI, however, was able to demonstrate that its models are actually more effective when paired with human expertise, leading to a compromise that is now “reflected in law and policy.”
OpenAI’s Sam Altman announced that the company will build “technical safeguards” into the deployment that literally prevent the models from outputting commands that bypass human oversight. These safeguards are managed by OpenAI’s own safety stack, which runs in a secure cloud. This ensures that even if a rogue commander tried to use OpenAI for autonomous strikes, the system would refuse to execute the command.
The agreement also includes a ban on using OpenAI for “social credit-style” systems, another major concern for civil rights groups. By being explicit about what the AI cannot do, OpenAI is setting a high bar for transparency in government contracting. The company has also urged the Department of War to apply these same “OpenAI standards” to every AI lab it works with, hoping to create a level playing field for ethical technology.
While the deal is a massive commercial success for OpenAI, the company insists that the primary goal is safety. By embedding its principles directly into the heart of the American military, OpenAI is attempting to prevent the “worst-case scenarios” of AI development from ever coming to pass. As OpenAI technology begins its work in the Pentagon, the world is witnessing the first major test of whether AI can truly be controlled in a theater of war.