Quick Buzz Feed
The U.S. military’s desire for an AI tool with “minimal refusal rates” was revealed through a FOIA lawsuit filed by The Intercept.
The U.S. Department of Defense reportedly sought a specialized version of OpenAI's artificial intelligence technology, explicitly engineered to exhibit "minimal refusal rates" when processing military commands. This concerning detail emerged from documents obtained by The Intercept through a Freedom of Information Act lawsuit, which aimed to shed light on the military's opaque collaborations with leading AI companies. The request for such a highly compliant AI tool raises significant ethical concerns about potential scenarios where AI might execute directives without sufficient built-in constraints or moral reasoning, especially in critical national security operations. It highlights a military desire for AI that would prioritize obedience over caution, prompting questions about the future of autonomous systems in warfare.
In 2025, several prominent AI developers, including OpenAI, Google, xAI, and Anthropic, engaged in agreements with the Pentagon to develop militarized AI prototypes. These systems were intended for various military applications, such as improving logistical efficiency, enhancing intelligence analysis for decision-making, and directly supporting "warfighting" operations. The Pentagon later sought to broaden these agreements, planning to deploy the AI technologies across U.S. classified computer networks, indicating a significant push towards integrating advanced AI into sensitive military infrastructure. This expansion, however, was not without controversy, leading to a notable dispute with Anthropic over the ethical limitations of military applications for its technology.
The contentious phrase "minimal refusal rates" was found within an updated contract document, version "P00003," which further expanded upon an existing $200 million prototype deal with OpenAI. Despite its appearance in documents released to The Intercept, both OpenAI and the Pentagon have since asserted that this specific language was not present in the final, executed contract. OpenAI's spokesperson claimed that the document was an early draft which the company rejected. Initially, a Department of Justice attorney confirmed the document's authenticity but later retracted this confirmation, stating the Department of Defense needed more time to investigate. This sudden reversal and the conflicting statements have intensified concerns about transparency and accountability in military AI procurement processes.
The phrase "minimal refusal rates" for "OpenAI Mission Models" intended for "national security problem sets" carries profound implications. It suggests a version of OpenAI's technology stripped of the ethical guardrails typically found in its public-facing models. For instance, consumer versions of ChatGPT are designed to reject queries related to harmful activities like prioritizing drone airstrike targets or developing weapons. Experts like Heidy Khlaaf from the AI Now Institute expressed significant alarm, interpreting "minimal refusal" as likely signifying a drastic reduction or complete absence of safeguards. Such a design could render AI a highly compliant tool for military objectives, potentially bypassing crucial ethical considerations and human oversight in military operations, raising the risk of autonomous decision-making in conflict zones.
The ongoing debate surrounding OpenAI's contract highlights critical ethical questions about the increasing involvement of private technology corporations in military affairs and the development of AI for combat. Experts are deeply troubled by the prospect of private entities influencing or making determinations in warfare, such as targeting recommendations or setting operational parameters, which traditionally fall under state obligations. The previous conflict with Anthropic, where the company resisted contractual prohibitions against autonomous weapons and domestic surveillance, further illustrates the tension between corporate ethical stances and military demands. While OpenAI publicly stated it secured red lines on autonomous killings and spying in its latest classified deployment contract, the extensive redactions and the broad phrasing permitting "any uses the government deems legal" leave the true extent of AI deployment in warfare ambiguously defined and open to significant scrutiny.