A significant National Security Memorandum signed by President Joe Biden mandates human oversight, safety testing, and various safeguards for military and intelligence uses of artificial intelligence (AI). This memo initiates a comprehensive review of how the Pentagon and intelligence agencies acquire AI, with regulatory recommendations and reforms expected next year.
However, the memorandum and the accompanying Risk Management Framework (RMF) do not introduce substantial new restrictions on AI-controlled drones, munitions, or other autonomous weapons—issues that are of particular concern to arms control advocates. Instead, the RMF mainly aligns with existing Pentagon policies outlined in DoD Directive 3000.09, which was revised last year to impose limitations rather than outright bans on autonomous weapons that already include computer-controlled systems, such as missile defenses. The focus of the new policy is primarily on AI applications involving data analysis and decision-making, especially regarding lethal force.
The memorandum refers to „a classified annex” that deals with sensitive national security topics, including countering adversarial AI that presents risks to the U.S. However, the annex’s contents remain unspecified, raising questions about the nature of the risks addressed. Another element of uncertainty is the upcoming election; while a hypothetical Kamala Harris administration may continue Biden’s policies, the GOP has signaled intentions to repeal current AI regulations, with former Trump advisors advocating for accelerated military AI development reminiscent of a „Manhattan Project.”
The Details Matter
Biden administration officials have expressed that the intent behind the new policy is to promote the adoption of AI by establishing clear guidelines, rather than stifling what they term “responsible” use. Consequently, the memo and RMF tend to avoid outright prohibitions, instead allowing AI development and deployment as long as an extensive set of best practices is adhered to.
One explicit prohibition in the RMF prohibits using AI to exclude a human from critical decisions regarding the launch or termination of nuclear weapon employment, reaffirming a long-standing U.S. policy that necessitates human oversight in this area.
The framework also restricts the use of AI in anticipating civilian casualties or collateral damage during military strike planning, mandating both technical safeguards—such as continuous rigorous testing—and human oversight by trained personnel. Moreover, AI-generated intelligence reports must be distinctly labeled to warn readers when the analysis is based solely on AI inputs.
The RMF includes a detailed list of “high impact AI use cases,” which are permissible if agencies implement specific safeguards. Many of these precautions are designed to protect human rights, including restrictions on using AI for tracking or categorizing individuals as potential threats. Human oversight is similarly mandated for AI that assesses eligibility for benefits such as political asylum or federal employment.
Before deploying any new AI or retroactively for existing systems, agencies must conduct thorough assessments, including a cost-benefit analysis of whether AI is the most appropriate solution compared to traditional methods. Guardrails include realistic context testing, pilot programs, failure mode analysis, assessments of training data accuracy, and continuous monitoring of AI performance.
Training for human operators of AI is a crucial aspect of the policy, ensuring they are equipped to identify errors rather than blindly trust technology—addressing concerns about „automation bias.” The policy does allow for some waivers of these safeguards, requiring documented approval from designated Chief AI Officers and annual reviews.
In summary, the National Security Memorandum and RMF emphasize regulation over prohibition regarding sensitive AI applications, aiming to facilitate rapid advancements in military technology while maintaining ethical standards and safety protocols. Experts like Michael Horowitz, a professor at the University of Pennsylvania and former deputy assistant secretary of defense for emerging capabilities, assert that the framework strikes an effective balance between delineating prohibitions and enabling swift adoption of AI, drawing on decades of experience in safe AI implementation.
