In 2021, a UN report documented something that military analysts had long predicted but hoped to avoid: the first confirmed use of autonomous lethal drones against humans without direct human command. Turkish-made Kargu-2 drones, operating in Libya, reportedly hunted and attacked retreating forces using onboard AI - no operator in the loop.
That was four years ago. The technology has only advanced.
The Ukraine Laboratory
The ongoing conflict in Ukraine has become an unprecedented testing ground for AI-enabled warfare:
Drone swarms coordinate attacks on fortified positions, with individual units making real-time decisions about targeting and evasion.
Aquatic drones have struck naval vessels, demonstrating that autonomous weapons aren't limited to the air.
AI-powered targeting systems help identify military equipment in satellite imagery and drone footage, accelerating the kill chain from detection to strike.
Counter-drone AI attempts to detect, track, and neutralize incoming autonomous threats.
What was theoretical in 2020 is operational in 2025.
The Technology Stack
Modern military AI combines several capabilities:
Computer vision identifies targets, distinguishes military from civilian objects (with varying reliability), and enables navigation without GPS (which can be jammed).
Swarm coordination allows dozens or hundreds of cheap drones to operate as a unit, overwhelming defenses through numbers rather than individual capability.
Edge computing puts AI inference on the drone itself, eliminating reliance on communication links that can be severed or intercepted.
Reinforcement learning trains systems to adapt tactics based on what works, potentially evolving faster than human doctrine can respond.
The Ethical Chasm
The deployment of autonomous weapons has outpaced governance:
The fundamental question: Should machines make life-or-death decisions without human oversight?
The practical reality: In the chaos of combat, with milliseconds to react, the human operator may already be a fiction. "Human in the loop" often means a human somewhere watching a screen, not meaningfully controlling each decision.
The asymmetry: Nations that restrict autonomous weapons face adversaries who don't. The incentive to match capabilities is overwhelming.
The proliferation risk: Unlike nuclear weapons, which require rare materials and sophisticated infrastructure, autonomous drones can be built with commercial components. The barrier to entry is low and falling.
What Guidelines Exist?
Efforts to regulate autonomous weapons have produced more discussion than binding agreements:
The Convention on Certain Conventional Weapons (CCW) has debated lethal autonomous weapons systems (LAWS) since 2014. Progress has been minimal.
The US Department of Defense Directive 3000.09 requires "appropriate levels of human judgment" but leaves "appropriate" undefined.
Various nations have called for bans or moratoriums, but the countries most actively developing these systems have resisted binding restrictions.
The 2024 framework for military AI guidelines, while a step forward, remains voluntary and vague on enforcement.
The Scenarios That Keep Analysts Awake
- Accidental escalation: An autonomous system misidentifies a target, triggering a response that spirals beyond human control
- Proliferation to non-state actors: Terrorist groups or criminal organizations acquire autonomous weapons capability
- The vulnerability of critical infrastructure: Swarms designed for military use could just as easily target power grids, communications, or water systems
- The speed mismatch: When autonomous systems fight each other, events may unfold faster than human decision-makers can understand, let alone control
The Commercial Crossover
Many military AI capabilities derive from commercial technology:
- Object detection trained on public datasets
- Navigation systems from autonomous vehicle research
- Swarm algorithms from robotics competitions
- Edge AI chips designed for consumer devices
This dual-use nature means:
- Advances are rapid because commercial R&D dwarfs military budgets
- Export controls are difficult because the underlying technology is everywhere
- The line between civilian and military AI blurs
What Should We Do?
There are no easy answers, but several approaches deserve consideration:
Meaningful human control: Not just a human somewhere in the chain, but genuine human judgment on consequential decisions.
Accountability frameworks: When an autonomous system causes harm, who is responsible? The operator? The commander? The manufacturer? The algorithm designer?
Transparency requirements: Even if capabilities can't be restricted, understanding what systems are deployed and how they operate could reduce accident risk.
Technical safeguards: Kill switches, geographic limitations, rules of engagement encoded in software - imperfect but better than nothing.
International dialogue: Even adversaries have shared interests in preventing accidental escalation.
The Uncomfortable Reality
Autonomous weapons are not coming. They're here. The choices now are about how they're used, by whom, and with what constraints.
The AI community - which develops the underlying technology - has a stake in these outcomes. The techniques that enable drone swarms also enable beneficial applications. But we can't pretend the military implications don't exist.
This isn't about stopping progress. It's about ensuring progress doesn't outrun our ability to control it.
