Join War on the Rocks and gain access to content trusted by policymakers, military leaders, and strategic thinkers worldwide.
We spare the reader another detailed story from Ukraine to establish the now-familiar point that battlefield technology changes faster than militaries can acquire, review, and field it. The villain in this story is easy to identify: bureaucracy. The fashionable remedies are equally familiar — cut the red tape slowing procurement, or revise the policy review requirements for autonomous weapon systems. But bureaucratic reform, however necessary, cannot eliminate the fundamental challenges machine-learning-based systems pose for rapid testing, evaluation, validation, and verification. Moreover, legal and operational reasons make it unwise to cut corners on compliance with domestic and international law.
Rather than relying on red-tape cuts or curtailed review to speed things up, we argue the United States and other NATO militaries can streamline weapons review and improve design programs by embedding legal and operational advisers within defense firms who can, at a company’s discretion, give input and advice. Such advisers would support companies and speed reviews, without adding oversight or overhead. The core idea is to change the review architecture: Rather than becoming a discrete hurdle at the end of development, review becomes a continuous process running alongside development. This is no silver bullet, but a concrete step militaries and industry may take together to meet today’s challenges.
A clarification first, since “review” covers several distinct processes. In the United States, a law of war weapons review, conducted as a matter of policy since 1974 and required today by U.S. Department of Defense Directive 5000.01, asks whether a weapon could be lawfully used. A separate senior-level review under Directive 3000.09 asks both before development and fielding whether (semi-)autonomous weapon systems satisfy policy requirements. Test and evaluation establishes actual capabilities and limits. Our proposal serves all three, which draw on the same underlying evidence, though time pressure is sharpest on the Directive 3000.09 review, our running example here.
Ukraine has shown that militaries can rapidly adapt platforms, software, electronic warfare, and doctrine on timescales measured in days. In the United States and NATO countries, this is generally not the case.
The mismatch between how quickly a system can change and how quickly it can be reviewed is sharpest for AI and machine learning-enabled systems. Fielded models benefit from retraining on current theater data, creating strong operational grounds for updating. But the test, evaluation, verification, and validation that review requires can take longer than a provider needs to develop the next iteration. That makes traditional procurement and review pipelines onerous at best and strategically disastrous at worst.
The challenge is especially stark for models with millions or billions of learned weights, as even small changes to a dataset or training process can shift model behavior in significant and often hard-to-parse ways. For such systems, a review beginning only at deployment-readiness fails twice over: it leaves procurement authorities far behind the technological curve, and any defect it uncovers forces costly rework of every iteration already built on top of it.
Two responses dominate the debate. The first is to cut red tape surrounding development and deployment. This must be done carefully, as many seemingly unnecessary rules are the residue of past, even deadly mishaps. Overly complicated processes, redundant or unclaimed rules, and requirements that mainly bar smaller entrants can and should be pruned. But the fundamental challenge remains: complex models must be examined, massive datasets assessed, and small changes evaluated afresh. Cutting tape alone cannot make review timely.
Given this, the second broad approach is to cut the review process itself. One could relax standards, accepting systems that are less predictable, discriminate, or controllable, or simply curb testing. Either approach would speed pipelines.
But such brute-force measures cut against standing U.S. policy and are neither ethical nor safe for the warfighters who use these systems. The United States is not a party to Additional Protocol I, and Article 36 of that treaty requires nothing of it. But the United States has conducted law of war weapons reviews as a matter of policy since 1974. Directive 5000.01 carries that requirement today, and Directive 3000.09 adds the senior-level policy review for autonomous and semi-autonomous systems. Both persist because the underlying judgment is sound, not because a treaty compels them. Law and ethics are also central pillars and force multipliers for the United States, NATO, and allied militaries, and trading them for faster acquisitions undermines the readiness such plans are meant to serve. Thorough test and evaluation can help establish critical system limits: the raw material for field manuals and training.
These reviews, and the procurement regimes around them, are too slow and risk leaving militaries technologically behind. But simply going faster can undermine safety and core values, harming readiness. The goal should be efficiency, not speed. We propose militaries embed legal and operational advisers within industry to observe development, document challenges and solutions, advise on request, and compile a portfolio of a system’s technical details and their ethical, legal, and operational implications.
This may sound like already existing mechanisms: In the United States, the Defense Contract Management Agency keeps resident personnel inside contractor facilities, and both the law of war weapons review and the Directive 3000.09 review are meant to occur at more than one point in a program’s life.
But these were built for a different problem. Contract-administration personnel oversee cost, schedule, and conformance to specification, not the legal and operational character of what is built. While acquisition has begun adapting to software timelines, milestone-based review remains too infrequent for today’s technology. We propose not new oversight but new cadence and competence. Military advisers would be present continuously, focusing on legal and operational questions, and documenting for the reviewing body rather than the contracting officer.
Why would this speed review, if it appears to just add more red tape? Documentation of earlier versions cannot stand in for evaluation of the final AI or machine learning system since changes between versions can be large. But embedded documentation changes the review architecture. When reviewers know what changed between versions, what testing accompanied each change, and which model behaviors remained stable, reviewing a deployment candidate becomes assessing deltas from already-examined baselines rather than starting from scratch. Reviewers inherit running records and can focus scarce test and evaluation resources on what is new. And because military advisers compile that record, it is more likely to include what reviewing bodies need than what industry wants to highlight. Running assessments are not reviews themselves, but they hand reviewers a ready-made first pass.
Our proposal calls for two kinds of embeds: legal and operational advisors. Legal advisers can forestall any potential development decisions that would theoretically render a system unprocurable, such as accepting data of unclear provenance or possibly compromised inputs, sparing firms further investment in unfieldable systems. They can also steer designs away from legal obstacles to fielding. Operational advisers keep development aligned with real military needs. They should sit high enough in the command hierarchy to know formation requirements, yet low enough to stay connected to operators and veterans who can fine-tune designs. Together, the advisers should ensure the record supports later field manuals and training.
Neither class of embed replaces technical evaluators and their burden of proof. For the largest programs, embeds should be backed by the military’s test and evaluation community, whether as a third class of embed or reach-back support. Legal advisers set and check the documentation standards reviewers require, industry engineers produce content that meets them, and the military side vouches for the record’s completeness.
To avoid compromising reviews, embeds should have no role in reviewing or procuring the systems they observe. Tying their career prospects to the thoroughness of their reports, rather than to whether systems are procured, mitigates this risk and helps ensure the report’s accuracy and fidelity.
Others have proposed embedding in the opposite direction, placing industry engineers inside military units to accelerate combat iteration. These are complements, not competitors. The faster engineer-in-unit cells spin the develop-adapt loop, the larger the documentation debt owed to the reviewing body. Our proposal is how that debt gets paid down.
Three industry objections deserve consideration: fear of micromanagement or burdensome oversight; fear that outsiders familiar with ongoing research and development could threaten intellectual property or competitive advantage; and reluctance to “show how the sausage is made” by letting the military see the faults and missteps in development.
Each objection has an answer. First, embedded advisers are advisers, nothing more. Industry decides whether to ask for advice, share information, or even confine advisers to their offices. The proposal accelerates review for firms that choose openness and creates competitive pressure on the rest, since guarded firms yield thinner documentation and, consequently, slower review, but it does not guarantee coverage. Firms may restrict access, leaving the military’s insight no worse than under the status quo. That limit is deliberate. Contracts could, in principle, require firms to grant embeds access to development records, test results, and facilities, but that would turn a voluntary benefit into another compliance obligation. Any such mandate must be weighed against the trust the voluntary model is designed to build.
Second, protecting intellectual property will require rules on how advisers are deployed. The specifics will depend on domestic intellectual property and secrecy laws but embeds should remain with a single company as long as possible, never rotating between competitors, and a non-compete rule should keep former embeds out of the defense industry for a suitable period after service. But these protections carry their own risk. Long, exclusive tenure is a recipe for capture, in which the embed comes to see matters as the company does. The promotion incentive described above mitigates this but cannot eliminate it. Periodic independent audits comparing embed reports with how systems later fared in review would provide that check, as would pairing embeds with staggered tenures on the largest programs.
The third worry is that letting military professionals see every trial, false start, and failure might undermine trust. This gets things backward. Openness and candor build trust, and in the defense domain, trust is a strong currency and among the hardest to come by. Embedded advisers give companies a running opportunity to be open with the military across a system’s full lifecycle. There may be some need to introduce military professionals to the realities of defense research to ensure they understand how both business and development are run. Showing how the sausage is made is only a problem if a company is putting garbage in the sausage.
The military would bear costs of its own. Embedding advisers requires billets from legal and operational communities that are already stretched thin, along with clearances and facility arrangements. This also raises worries of the military “picking winners” by virtue of which companies they provide advisors to. There is no fully satisfactory response to this, but a sensible approach is to assign advisors first to companies where research is most apt to outpace review. The proposal also raises hard questions of accountability. If a system observed throughout development nonetheless fails review, or fails in the field, who answers for it?
These costs are real. But so are the status quo’s: capabilities that sit unfielded for years, defects caught at final review that invalidate multiple generations of work at once, and reviewers left to reconstruct a system’s development history after the fact. The proposal can start small, with a pilot placing a handful of embeds in programs where development cycles most outpace review. To ease the strain, operational advisers can come from units with a stake in the capability, legal advisers from the judge advocate community, and technical support as reach-back rather than resident billets. Because embedded documentation reduces the load of final assessment, the net effect shifts effort rather than adding headcount.
A common view holds that law, ethics, and oversight hinder the rapid development and acquisition of emerging technologies. This assumption, however, misreads both the force-multiplying value of ethics and law and the critical importance of thorough testing, evaluation, verification, and validation for fielding systems that work. Without thorough documentation, field manuals and training rest on guesswork, and soldiers ignorant of systems’ limitations are likely to misuse them, to their own harm. “Move fast and break things” is a sensible mantra when breakage means a crashed app and subsequent patch. In the military, the “things” that break are our own warfighters. Curtailing evaluation risks filling arsenals with systems that are technically capable yet cannot be safely or effectively deployed. That is not readiness.
This calls for a review process that is efficient, not merely fast. Embedding military legal and operational advisers within industry keeps industrial partners informed of evolving needs, documents day-to-day trials and development solutions to accelerate test and evaluation, and builds cooperation to produce better systems and faster integration. It will not solve every acquisition problem, but it is a concrete step militaries can take to address the challenges posed by how fast new technologies are developed and iterated. While differences in legal regimes among allied states may hinder complete interoperability, better documentation tailored to faster review can improve each state’s internal processes. Most NATO members are parties to Additional Protocol I and carry an Article 36 obligation the United States does not, yet implementation across the alliance is uneven. A common documentation standard would give those states something their own obligations already require.
This architecture has a further advantage: it holds regardless of who runs government. Oversight is a pendulum. One administration loosens review to move faster; the next tightens it, re-examines inherited programs, or cancels systems with contentious records. A program begun today may span several administrations before fielding, so the greatest schedule risk may not be today’s review standards but tomorrow’s. Existing program documentation offers little protection, since snapshots go stale as systems iterate. A continuous, military-authored record fills that gap. It gives future skeptics nothing to hold a program up for and lets future standard-setters raise the bar without years of new testing. Continuous documentation gives each side what it wants without waiting for its turn in power.
Nathan Wood is an executive director of the International Society for Military Ethics in Europe and a junior research group leader at the Hamburg University of Technology, where he is leading the project “Military Defense Technologies and Ethics,” supported by the German Aerospace Center and the German Federal Ministry of Education and Research. He has published in numerous scientific and popular journals.
Jesse Hamilton is Departmental Lecturer in Philosophy and Public Policy at the Blavatnik School of Government, University of Oxford, where he is also assistant director of the Military Leadership and Judgment Programme. Before academia, Jesse served in the U.S. Army and subsequently worked in asset management.
Image: Midjourney
