Congress comes back to Washington on Monday with about five working weeks left in the year and a defense bill it never finished. The House passed its version, H.R. 8800, on July 22 by a vote of 216 to 212. The Senate never passed S. 4784; a July 14 vote to take it up failed 50 to 46, according to the independent tracker ndaa.ai. Leaders now expect a compromise text written by the four Armed Services leaders rather than a Senate floor fight.

Buried in that negotiation is the most detailed set of rules Congress has ever considered for autonomous weapons and military AI. The Senate text would write the Pentagon’s autonomy directive into law, add design requirements and a testing regime, and ban AI from any decision to launch a nuclear weapon. The House text does something different: it orders the Department of War to rewrite the directive itself within a year, to a list of nine required elements, and to report back.

Democrats won both chambers on Nov. 3, Breaking Defense reported, so whatever this conference produces becomes the floor for the next Congress. Here is what each bill actually requires, what the department already has to do under its own directive, and what has to happen before the Dec. 11 funding deadline eats the calendar.

Where the bill stands

The Senate is scheduled to be in session eight days in November and 14 in December, counting Fridays, with Thanksgiving recess Nov. 23 to 27 and a break starting Dec. 21, according to ndaa.ai. Government funding runs out Dec. 11. Sen. Mike Rounds said before the recess, “I think the NDAA will come together after the election.”

Both committees marked up in early June. The Senate’s autonomy sections came largely from a June 8 bill by Sens. Chris Coons and Jack Reed, with pieces of a Kelly-Subramanyam bill folded in, and the Republican majority accepted them into the base text, the Arms Control Association reported.

What the Senate text would require

Section 1647 of S. 4784 would add a new section 4577 to Title 10. Its opening line is not a restriction. It says the department’s policy is to “maximize uses of autonomy and artificial intelligence capabilities to the extent practicable,” while ensuring “appropriate human oversight.” The rules follow from there.

Every autonomous weapon or AI capability would be sorted into one of two levels of required human judgment, based on how much risk to life a failure would pose. To the maximum extent practicable, each system would have to let operators supervise it during a mission, include a way to intervene in or stop the use of force at any stage, fall back to manual control when jammed or spoofed, show operators enough status and decision logic to oversee it, and keep records of target selection data and operator actions for post-engagement review.

The section would codify the two-stage senior review in Directive 3000.09, before formal development and again before fielding, and assign it to the Under Secretary for Research and Engineering or a designee, in consultation with the Vice Chairman of the Joint Chiefs. A verification would be good for three years. Systems exempted under the current directive but lacking a valid verification would have to go back through review. The Director of Operational Test and Evaluation would have to confirm that covered systems work “in realistic operational environments against adaptive adversaries,” with a 12-item test list that applies in full to higher-risk systems.

Three uses would be prohibited outright: using autonomy or AI “for the decision to initiate the launch or detonation of a nuclear weapon,” using it to monitor or target people reasonably believed to be in the United States outside existing law, and employing lethal force by an autonomous weapon “without incorporation of appropriate levels of human judgment.” The AI Policy Network, which has urged conferees to keep the section, says it would also require contractors to produce a bill of materials covering AI software, data and hardware.

Section 1653 is shorter and aimed at the chain of command. Within 180 days of enactment, the Secretary would have to issue regulations ensuring that the use of force “remains subject to a clear, accountable, and identifiable human chain of command and control,” that decisions to authorize, start or stop force are “attributable to human commanders and operators,” and that no weapon system “shall eliminate human responsibility for the use of force.” AI-enabled weapons would need operational plans and training manuals that spell out operator roles, exercises that test for automation bias, and after-action processes that can name the responsible human. The Chairman of the Joint Chiefs would write a joint doctrine for autonomous warfare.

What the House text would require

The House bill has no match for either section. Its closest provision is Section 1524, which gives the Secretary one year to update department policy, “including by revising Department of Defense Directive 3000.09.” The scope is in one respect wider than the Senate’s: it covers not only weapons but AI systems that “support, recommend, or materially influence” decisions about force, including “operational planning, target development, weaponeering, or engagement recommendation.” That is the category targeting tools such as Maven fall into, as NPR has described them.

The House lists nine things the updated policy must contain, among them criteria for categorizing systems by mission context and consequence; “realistic and combat-effective requirements for operator intervention, override mechanisms, and operational resilience”; procedures to identify the commanders responsible for authorizing and terminating force; auditability; criteria for rapidly fielding systems after “material changes to software, models, data, or operational context”; and training for “calibrated reliance” on AI.

The House text states that nothing in it requires the department to pause any program. It asks for an interim report at 180 days and semiannual implementation reports through 2032 with public unclassified summaries. The Arms Control Association described the House approach as “less specificity and more deference to the Pentagon.”

What the directive already says, and the rewrite in the wings

Directive 3000.09, last updated in January 2023, requires that autonomous and semi-autonomous weapons be designed to “allow commanders and operators to exercise appropriate levels of human judgment over the use of force,” and it sets a senior review before formal development and before fielding, according to the Congressional Research Service. The standard is deliberately flexible. It does not require a human to pull every trigger, a point the Lieber Institute at West Point notes the U.S. has defended for years in Geneva against calls for binding “meaningful human control.”

The directive is also in motion. National Security Presidential Memorandum 11, signed June 5, ordered the department to update 3000.09 within 90 days and review it annually, DefenseScoop reported. Sen. Ruben Gallego wrote to Secretary Pete Hegseth on June 12 warning that a “compressed” rewrite could raise the risk of friendly fire and civilian harm. We could not find a public release of a revised directive as of Sunday. The Senate section would lock the review structure into statute; the House section would hand the rewrite back to the department with a checklist.

Our read: what it means for program offices, vendors and lawmakers

For program offices, the Senate text is the one to plan against, because its design requirements are specific and would apply regardless of acquisition pathway. Supervision, intervention, manual fallback, decision logging and three-year verifications are engineering and documentation work, not slogans. The two-level sort means low-risk systems get lighter treatment, but the department would have to defend each categorization.

For vendors, two items matter most. The Senate’s bill-of-materials requirement, as described by the AI Policy Network, would make AI supply chains a contract deliverable. And both bills push toward re-review after software or model changes, which favors companies that can show what changed and test it quickly.

For lawmakers, the choice is whether to legislate the standard or delegate it. The Senate codifies “appropriate levels of human judgment” and defines what it must include. The House keeps that judgment inside the department but demands a paper trail through 2032. A conference could plausibly take the House’s broader scope on decision-support AI, the Senate’s prohibitions, and a softened version of the Senate’s design list. That is speculation on our part; conference negotiations are not public.

What we don’t know

We do not know whether the four leaders have agreed on a text or when they will meet. We do not know whether the 3000.09 rewrite ordered in June has been completed inside the building, or what it says. We do not know how the department would handle the Senate’s requirement to re-review systems previously exempted under the directive, or how many systems that covers. We have not seen a reconciled AI title, and this piece relies on the filed House and Senate texts, not drafts in circulation.

What to watch

• This week: Any statement from Wicker, Reed, Rogers or Smith on a conference schedule, and whether the AI title is among the open items.

• Dec. 11: The funding deadline. If the NDAA slides into a year-end package, the autonomy sections could be traded away quietly.

• 3000.09: A revised directive from the department would change what the House section asks for and what the Senate section codifies.

• Section 1647’s prohibitions: The nuclear and domestic-targeting bans are the simplest provisions to keep, and the nuclear ban has bipartisan history going back to 2023.

• January: Whatever drops out now returns as the starting point for new Democratic majorities on both Armed Services committees.

Machine Front runs every Monday. Get it in your inbox by subscribing to the Autonomy Wire newsletter.