The argument about AI in warfare is usually conducted as an argument about robots. It is really an argument about tempo, and about how much of a decision a human can meaningfully own when the decision has to be made in seconds.

Mass, firepower and attrition have not stopped mattering. They have simply been joined by something harder to procure: the ability to observe, orient, decide and act faster than the other side can react. The uncomfortable end of that logic has a name, hyperwar, in which events outrun the oversight that is supposed to govern them.

What is already fielded

Project Maven is the clearest example. It ingests imagery, full-motion video and other ISR feeds, detects objects of interest, and passes them to a human analyst. It does not fire at anything. One targeting officer reported working through around 80 targets an hour with it, against roughly 30 without, and with fewer analysts. That is the entire proposition in one number: the same people, doing more of the same job, considerably faster.

In Ukraine, commercial satellite imagery, drone feeds and open data are fused through analytic pipelines that locate and classify equipment far faster than manual exploitation allowed (Georgetown Journal of International Affairs). In Gaza, systems reported as Gospel and Lavender have generated targeting options and maintained large databases of individuals assessed from behavioural and network signals. Those cases are where the debate actually lives: the output volume rises, the time available to challenge any single recommendation falls, and accountability becomes harder to locate.

Where it earns its place

Sensor fusion and ISR exploitation is the obvious one: cueing objects for a human to look at, re-identifying a target across passes, noticing that a logistics pattern has changed. Models small enough to run on the platform keep that working when the link does not.

Electronic and cyber warfare is less discussed and moving quickly. Jamming that adapts to the pattern it is attacking, spectrum allocated dynamically in contested space, intrusion detection that spots an adversary already inside the network. These are environments where the response window is shorter than human reaction time, which makes the case for machine speed far easier than it is for targeting.

Then there is the unglamorous work that wins campaigns: predicting component failure before it strands a vehicle, routing resupply around a threat, allocating scarce assets. No ethical committee is required, and the return is immediate.

Higher up, decision support proposes courses of action and battle management sequences effects. The limiting factor there is not accuracy but trust. A staff officer who cannot interrogate a recommendation will either ignore it or accept it uncritically, and both failure modes are expensive.

The failure modes that should worry people

Models can be deceived. Camouflage, spoofing, sensor jamming and poisoned training data are all cheaper than the systems they defeat, and an adversary does not need to break a classifier to win; degrading confidence in it is enough.

Accountability is the harder problem. When an assisted targeting decision is wrong, the question of who is answerable, the operator, the developer, or whoever prepared the data, has no settled answer (Queen Mary University of London). Distinction, proportionality and military necessity are human judgements, and they do not survive being delegated quietly.

Automation bias makes that worse. Under time pressure, people stop interrogating a system that has been right for a fortnight. A confirmation step that nobody has the time or information to refuse is not oversight, and designing out that comfortable illusion is mostly a human factors problem rather than a technical one.

And compressing everyone’s decision cycle simultaneously is not obviously stabilising. Two sides both acting faster than they can verify is a recipe for an escalation nobody chose.

Lessons from Ukraine

The pattern there is instructive: open satellite data processed into usable detections, classification happening onboard drones to cut the delay between finding and striking, and previously separate intelligence systems federated to shorten the loop (Tech Policy Press). The constraints are equally instructive. Misidentification remains a real risk, communications degrade exactly when they are most needed, and compute and bandwidth at the edge are finite in a way that laboratory demonstrations never are.

A sensible order of adoption

Begin with logistics and maintenance, where being wrong costs money rather than lives, and where the data is already there. Move to decision support rather than autonomy, and keep escalation with a human for anything lethal. Build the architecture to be modular and the models to be interrogable, because an operator who cannot see why will not act on what. Stress the system adversarially in wargaming, with jamming and spoofing and deliberate deception, not only in clean conditions. Mediate classification properly so coalition partners can be included without the data governance collapsing. And keep an independent team whose job is to find the failure modes, permanently, not as a pre-deployment exercise.

None of this is speculative any more. AI is already shaping how decisions get made and how strikes get authorised. It is an accelerant rather than a capability in itself, and the organisations that handle it well will be the ones that invested in validation, doctrine and human-machine teaming rather than in compute alone.