The Weapon Passed Legal Review, Then the Software Changed
Every legal opinion rests on a statement of facts. A weapon’s legal review is no exception. Under U.S. Army Regulation (AR) 27-53, the program office that wants a weapon reviewed supplies a description of the system, its intended use, its anticipated effects, and the technical information behind them. The judge advocate’s opinion answers the question those facts pose. If the facts later change, the opinion does not become wrong. It becomes an opinion about a different weapon.
For most of the Army’s history, changing the facts meant changing the hardware, and changed hardware is hard to miss. Here, “software” includes executable code, trained artificial intelligence (AI) models and their weights, and the rules and parameter settings that govern the weapon’s behavior; it does not include physical hardware or live sensor observations. For an AI-enabled weapon, the facts that matter most to a lawyer now live in software: what the system treats as a target; how confident it must be before it acts; where it will and will not operate; and how much room a soldier has to stop it. Software of that kind is updated often, and an update leaves the launcher, the airframe, the serial number, and the name exactly as they were. This post is about the gap that opens when the legal review stays attached to the hardware while the facts move with the code, and about a short document that would close it.
One Weapon, Three Updates
Consider a hypothetical counter-drone interceptor that passed Army legal review before fielding. The review request described a system that detects small quadcopters, engages only inside a geofence (a boundary drawn in software beyond which the system will not act), fires only after a soldier confirms the target, and aborts within two seconds of a stop command. The opinion found the weapon lawful on those facts.
Over the next year the program issues three updates.
The first repairs a display fault. Engineers re-run the tests the weapon passed before fielding and confirm that target recognition, soldier confirmation, and the abort command behave as before. The facts in the review request are unchanged.
The second retrains the recognition model, the part of the software that decides what the sensors are looking at. After retraining, the weapon can also classify small fixed-wing aircraft as targets. Nothing about the launcher changed. The description of what the weapon identifies, which the opinion took as a given, has changed.
The third shortens a timing delay to make the system more responsive. It also has a side effect: the abort command now takes effect in four seconds, not two. The engineer who made the change recorded it as a performance fix.
A judge advocate would see at once that the second and third updates alter facts the opinion relied on, and that the third is the dangerous one, because nobody thought of it as a change to the weapon. The law can handle all three. The open question is whether anyone in the program will put the facts in front of a lawyer before the new version reaches the field.
The Law Already Says When to Look Again
AR 27-53 anticipates changed weapons. Paragraph 6f provides that if a fielded weapon is changed so that it is no longer the system or capability described in the review request, including a substantial change to its intended use or anticipated effects, the earlier review is no longer valid and a new review is required. Paragraph 6g addresses substantial changes in a weapon’s effects or accuracy. Paragraph 9 allows a program that is unsure whether review is required to put the question, through its servicing judge advocate, to the office The Judge Advocate General (TJAG) has designated for weapons reviews, ordinarily the National Security Law Division. The same paragraph bars fielding or employment of a weapon that needs a review and has not had one.
These are legal standards, applied by lawyers. They do not turn on the label an engineer gives to an update. The regulation’s weakness is not in its rules but in its intake. It assumes that someone will recognize a changed weapon and bring it in.
Earlier contributions to Articles of War have mapped the legal side of that problem. Andrew McKee and Christina Colclough examined review of autonomous weapons modified at the tactical edge, Tobias Vestner and Nicolò Borgesano addressed iterative review of evolving algorithms, and Ronald Alcala applied paragraph 6f to field-modified weapons in Ukraine. What remains is the handoff inside the program office, where the decision to ship a new version is made.
Where Software Decisions Are Actually Made
The Army’s software-intensive programs increasingly follow the Department of Defense (DoD) Software Acquisition Pathway, which expects frequent releases, testing integrated into development, and control over which version is approved for use. Three pieces of that process matter here, and each has a plain meaning.
A configuration is the exact combination of executable code, AI model versions, reference data, and configuration settings installed on a weapon, in ordinary language, the version actually in use. The approved version is the baseline. A configuration control board (CCB), a standard term in DoD configuration management, is the program’s release committee including the engineers, testers, and managers who decide whether a proposed change becomes part of the approved version. Before that decision, the program runs regression tests, which are the earlier tests repeated to confirm that the change did not break anything else. The program records every proposed change in a change package, the file that holds the change, its test results, and the board’s decision.
This is the machinery that decided to ship all three updates in the hypothetical. It is well suited to one question, whether the software works, and blind to another, whether the weapon described in the legal opinion is still the weapon on the launcher. A CCB has no authority to answer the second question, and should not. What it can do is notice it.
A Statement of Changed Facts
I propose that the program manager add to every change package for an AI-enabled weapon a short authority-impact statement. The name is less important than the content. Think of it as the statement of changed facts that a client owes a lawyer whose opinion the client is still relying on.
The statement would answer four questions in a page or less. What will the weapon do differently? Which facts in the last review request does the change touch: the targets it identifies; where it operates; how a soldier is involved; what effects it produces? What test evidence shows the difference, or shows that there is none? And who must decide whether the earlier review still covers the result?
Two features keep the statement honest. First, it compares the proposed version not only with the previous release but with the version the last legal review actually described. Three small updates can carry a weapon a long way from its reviewed description while each one looks minor on its own. Second, it treats a label as no evidence. “Maintenance patch” or “performance fix” describes the engineer’s intent, not the weapon’s behavior. The third update in the hypothetical shows why. If the program cannot say what changed, the statement records that it cannot, and the question goes to a lawyer.
Two limits should be stated plainly. The statement is no legal opinion, and the board that receives it gains no authority to decide the weapon’s legality. Nor is it an addendum that clears future software changes in advance. No document can do that, because the facts of a change are unknown until the change is made. Each release gets its own comparison between the weapon about to be fielded and the weapon the lawyers reviewed.
Three Decisions That Must Stay Separate
The statement would feed three different decision-makers, and it must not blur them.
The first is the Army weapons legal review under AR 27-53, by the Judge Advocate General‘s (TJAG) designee. For a change like the first update, the statement records why the earlier review still applies and preserves the test evidence. For a change like the second or third, it goes to the servicing judge advocate early, while engineers can still produce evidence or alter the design, and if paragraph 6f is engaged, a new review follows. The regulation itself addresses urgency. Paragraph 7c allows expedited review in appropriate cases, and paragraph 6i provides that operational urgency governs the speed of a review, not its conclusion.
The second is the senior review that DoD Directive 3000.09 requires for certain autonomous weapon systems, before development and again before fielding. The directive calls for a fresh senior review when a system’s algorithms, intended missions, operating environments, target sets, or expected adversary countermeasures have changed enough to fall outside the earlier approval, and it exempts some categories of systems. A software change can move a system outside its approval, and it can also end a system’s eligibility for an exemption. The statement should say which.
The third is testing. The directive requires verification and realistic testing, including additional testing after certain changes, to confirm that critical safety features have not been degraded. Passing a test provides evidence of behavior. It does not supply a legal review, and a legal review does not substitute for a test.
The statement gives those three processes one shared account of what changed. It does not let any of them stand in for another, and it never lets “no one checked” be recorded as “nothing changed.”
The Weapon in the Launcher
The division of responsibility this proposal relies on already exists. The program manager delivers software. The release board controls the approved version. Test organizations characterize behavior. Commanders employ the weapon under rules of engagement. TJAG’s designee conducts the legal review. The authority-impact statement moves no authority among them. It moves information, at the one moment when the information exists and the decision has not yet been made.
A weapon’s hardware may stay the same through dozens of updates. The facts a lawyer reviewed need not. The Army should be able to show, before a new version reaches soldiers, that the weapon in the launcher is still the weapon in the legal opinion, or that it has had the further review its changes require. It should not have to find out after an incident.
***
Burak Oktenli is an independent researcher based in Washington, DC, focusing on assurance, verification, and the governance of machine authority in autonomous and AI-enabled systems.
The views expressed are those of the author, and do not necessarily reflect the official position of the United States Military Academy, Department of the Army, or Department of Defense.
Articles of War is a forum for professionals to share opinions and cultivate ideas. Articles of War does not screen articles to fit a particular editorial agenda, nor endorse or advocate material that is published. Authorship does not indicate affiliation with Articles of War, the Lieber Institute, or the United States Military Academy West Point.
Photo credit: U.S. Marine Corps, Sgt. Joseph Helms
