What DoDD 3000.09 Actually Requires
The directive, last updated in January 2023, establishes a two-stage senior-level review covering autonomous weapon systems the department develops or fields — with specific exemptions for semi-autonomous systems without autonomous modes, certain local defensive applications against time-critical or saturation attacks, defense of deployed unmanned vehicles, and non-lethal non-kinetic force against materiel. Before formal development begins, a system requires sign-off from the Under Secretary of Defense for Policy, the Vice Chairman of the Joint Chiefs of Staff, and the Under Secretary for Research and Engineering. Before fielding, the same approval process runs again, substituting the Under Secretary for Acquisition and Sustainment for the R&E position.
The substantive standard — the phrase that matters for system design — is that autonomous and semi-autonomous weapons must be designed to "allow commanders and operators to exercise appropriate levels of human judgment over the use of force." The CRS Insight by Kelley Sayler, first published March 13, 2026, and updated October 5, is precise on what this means operationally: it does not require manual human control of every engagement. It requires broader human involvement in decisions about how, when, where, and why a weapon will be employed. A human must assess the operational environment and decide to deploy the weapon; once that deployment decision is made, the system may operate autonomously within those parameters. The requirement also encompasses a human determination that the weapon will be employed "with appropriate care and in accordance with the law of war, applicable treaties, weapon system safety rules, and applicable rules of engagement."
Three categories matter for design purposes. Fully autonomous systems — human out of the loop — select and engage targets without further human intervention after activation. Human-supervised systems — human on the loop — allow an operator to monitor and halt engagement. Semi-autonomous systems — human in the loop — only engage individual targets or target groups that a human operator has selected. The directive does not prohibit any of these categories, but it requires that all three enable the human-judgment standard at the deployment-decision level.
The directive also includes a waiver authority: the Deputy Secretary of Defense may waive the senior-level review process in conditions of urgent military need. FY2026 NDAA §1061 (10 U.S.C. §130g) added a requirement for written congressional notification within 30 days of any such waiver.
What NSPM-11 Set Out to Change
NSPM-11's §3(a) issued a single directive relevant to DoDD 3000.09: update the directive within 90 days to ensure "the deliberate adoption of AI systems that respect the chain of command and operational authorities," review it annually to account for "the rapidly evolving capabilities of AI systems," and keep it consistent with NSPM-11's §2 policy. That §2 policy is what gives the directive its substance: it calls for eliminating "unnecessary barriers to rapid deployment" (§2(a)) and for the national security enterprise to "adapt commercial or open-source AI technologies" (§2(b)).
Sen. Ruben Gallego (D-Ariz.) wrote to Secretary Hegseth on June 12 to raise "serious questions about whether sufficient analysis of how prospective changes may aggravate risks to U.S. servicemembers and U.S. power projection has been conducted." The letter flagged the risk that a revised directive could accelerate fielding, compress test, evaluation, verification, and validation requirements, or curtail the ability of commanders and operators to exercise human judgment over the use of force — the design-side safeguards with the most direct acquisition implications.
The commercial AI angle is the technically significant one. A CRS Insight by Kelley Sayler, first published March 13, 2026, and updated October 5, analyzes the Pentagon-Anthropic dispute that the revision may be trying to address. According to the CRS analysis, Anthropic was reportedly willing to modify its usage policies for DoD but refused to authorize uses involving fully autonomous weapons or mass domestic surveillance. Anthropic CEO Dario Amodei stated publicly that "today, frontier AI systems are simply not reliable enough to power fully autonomous weapons." On August 27, a federal district court vacated DoD's supply-chain risk designation of Anthropic; on September 25, a D.C. Circuit panel upheld DoD's authority under 41 U.S.C. §4713, meaning DoD may continue to enforce its directive to remove Anthropic products from its systems, barring a contravening ruling by the U.S. Supreme Court or the full D.C. Circuit. That reliability claim is not primarily a doctrinal question — it is a systems engineering question about whether the failure modes and adversarial manipulation surface of current commercial foundation models are compatible with compliant operation under DoDD 3000.09's requirements. NSPM-11's push toward commercial AI adaptation lands directly on this gap.
The Deadline That Passed and What the Silence Means
The 90-day window closed in early September 2026. No revised directive has been published. The Department of War provided a direct update to the Senate Armed Services Committee on September 28: the Department "is ... updating the 3000.09 policy for lethal autonomous weapons systems as directed by NSPM-11." The revision is in progress; the delay may reflect the complexity of aligning the update with the statutory landscape before Congress finalizes the FY2027 NDAA.
The pressure to move commercial AI into defense applications is genuine and heavily resourced in the budget request. The Defense Autonomous Warfare Group carries a $54.6 billion FY2027 request — $53.6 billion of it mandatory funding dependent on a reconciliation bill; authorizers have so far backed only the $1.0 billion discretionary portion, per CRS R49023. The Autonomous Warfare Command's acquisition ambitions depend on a procurement environment that can absorb commercial AI quickly. When the revision does appear, FY2024 NDAA §251 (P.L. 118-31) requires the Secretary to brief the defense committees within 30 days after making any changes to DoDD 3000.09 — providing Congress visibility into the revision, not a pre-change veto.
What the NDAA Would Codify — and Why That Changes the Stakes
The Senate Armed Services Committee's FY2027 NDAA, S. 4784, divides the autonomous systems question between two provisions. Section 917 creates a permissive authority for a Robotic and Autonomous Systems Command (RASCOM; new 10 U.S.C. §167a) — a separate construct from AutoWarCom, which Hegseth announced on September 30. The autonomous weapons governance framework is in Section 1647 (new 10 U.S.C. §4577). The committee summarized it as covering "human judgment, validation and testing requirements, prohibited uses, and a centralized incident reporting repository." Section 1647 borrows DoDD 3000.09's two-stage logic and human-judgment language, but changes who reviews — USD(R&E) or a Secretary's designee in consultation with the VCJCS, rather than the directive's three-official approval — and adds obligations the 2023 directive does not have: risk-level categorization, a three-year verification validity period, re-verification of previously approved systems, DOT&E verification and validation testing, and statutory prohibitions on certain use cases including autonomous nuclear launch decisions. The SASC reported S. 4784 on June 15, after an 18-9 committee vote on June 10. On July 14, a cloture motion to proceed to the bill failed 50-46; as of CRS's September 1, 2026 update, no further Senate floor action had been taken.
Retired Air Force Lieutenant General Jack Shanahan, the first director of the Joint Artificial Intelligence Center, made the case for statutory codification at a Center for a New American Security conference on June 11: a human must remain the "final decider," and there is a "legislative piece of that if you want to really take care of this at a high level, don't just rely on an individual policy in the Pentagon." The operational logic is straightforward: a departmental directive can be updated by the Secretary in 90 days. A statutory requirement takes an act of Congress to modify.
H.R. 8800, which the House passed 216-212 on July 22, takes a narrower approach in §1524. It requires DoD to update its autonomous weapons policy within one year of enactment, and prescribes substantive content: the update must include "requirements to preserve existing human command responsibility for the use of force involving autonomous systems or artificial intelligence-enabled systems, including procedures to identify the human commanders or operators responsible for authorizing, supervising, and terminating such use of force." Section 1524 also requires "appropriate and operationally responsive requirements for approval, validation, oversight, and authorized operational use," and "realistic and combat-effective" intervention requirements. That structure would constrain an NSPM-11-directed revision that weakened accountability, while also endorsing faster fielding.
No conference has been scheduled. If the Senate subsequently takes up S. 4784, passes it, and the chambers reconcile, programs face a range of outcomes: from a Senate-style framework adding statutory prohibitions, risk categorization and DOT&E testing, to a House-style directive update that endorses faster fielding while preserving accountability structures. Either way, the statutory architecture around DoDD 3000.09 is becoming more elaborate, not less.
What Programs Need to Decide Now
The compliance uncertainty creates a real design decision for any program developing autonomous systems. The CRS identified four potential congressional responses: codify the directive, impose additional reporting and notification, restrict appropriations for specific systems, or limit funding for AI-model applications deemed too risky. All four are structurally possible in the FY2027 NDAA. The question for program managers and system architects is which design posture minimizes rework across the range of plausible outcomes.
Three considerations follow from the analysis.
Design to the existing human-judgment standard. Congressional action in both chambers, while also endorsing faster AI adoption, moves toward preserving rather than eliminating human command responsibility requirements. A system designed to meet the January 2023 directive's requirements — clear human deployment-decision authority, explainable boundaries on autonomous engagement, termination capability — provides the best available compliance baseline. Note that the Senate's §1647 framework would add obligations the 2023 directive does not have: risk-level categorization, three-year re-verification, DOT&E testing, and an incident reporting obligation. Design to the 2023 standard now, but build documentation practices that can support those additional requirements. A system designed to a hoped-for relaxed standard faces architectural rework risk if statute constrains that relaxation.
Treat the two-stage review process as a planning input, not a bureaucratic overhead. The waiver path exists for urgent military need, but the FY2026 NDAA added congressional notification for waivers. Programs betting on waiver availability to accelerate development are taking on oversight risk that has become more visible. Design the review touchpoints into the acquisition baseline.
Separate the commercial-AI-integration question from the autonomy standard. NSPM-11's push toward commercial and open-source AI is about integration pathways, not about removing the human-judgment requirement. A program that integrates a commercial foundation model for perception, planning, or effects while maintaining compliant human deployment-decision architecture is well-positioned under any plausible outcome. One that relies on commercial AI to collapse the human decision loop in lethal engagement is structurally at risk regardless of what the directive says.
The directive may change. What will not change is congressional interest in knowing who is accountable when an autonomous system makes a lethal decision.
Documenting that accountability — the human decisions, the autonomy boundaries, the TEV&V record, the chain of human command authority — is what both the current directive and any plausible statutory successor will require a program to show. Spartan X's AI assurance and advisory work focuses on building exactly that kind of behavioral evidence architecture: the technical and operational artifacts that let a defense program demonstrate, rather than just assert, compliance with human-judgment and accountability requirements. Arbiter, Spartan X's AI verification capability under development, addresses one part of that problem — surfacing AI behavior outside expected parameters across varied operational inputs, which is the class of evidence a statutory review process would need to trust.
Sources and further reading
- DoDD 3000.09, Autonomy in Weapon Systems (January 25, 2023)
- NSPM-11, Artificial Intelligence in the National Security Enterprise (June 5, 2026)
- CRS Insight IN12669: Pentagon-Anthropic Dispute over Autonomous Weapon Systems, Kelley Sayler (first published March 13, 2026; updated October 5, 2026)
- Senator Gallego letter to Secretary Hegseth on DoDD 3000.09 (June 12, 2026)
- P.L. 118-31, FY2024 NDAA, §251 (December 22, 2023)
- P.L. 119-60, FY2026 NDAA, §1061 — 10 U.S.C. §130g (December 18, 2025)
- S. 4784 as reported, FY2027 NDAA — Senate Armed Services Committee (June 15, 2026)
- H.R. 8800 as passed, FY2027 NDAA — House of Representatives (July 22, 2026)
- FY2027 NDAA Executive Summary, Senate Armed Services Committee (June 2026)
- DoW Update on Implementation of the 2026 National Defense Strategy, submitted to SASC (September 28, 2026)
- CRS R49023: FY2027 Defense Budget — Funding for Selected Weapon Systems (updated September 3, 2026)
- CRS IN12704: FY2027 NDAA — Status of Legislative Activity (updated September 1, 2026)
- Air & Space Forces Magazine: Lawmakers Eye Four-Star Command for Unmanned, Autonomous Systems (June 11, 2026)
- Institute for Law and AI: The NDAA as a Vehicle for AI Governance (July 2026)
