The Shift to Autonomous Targeting in Modern Drone Warfare
By early 2025, Ukrainian and Russian forces were collectively launching an estimated 50,000 first-person-view drones per week along contested front lines—a figure cited repeatedly in open-source assessments of the conflict. The sheer volume broke the human-in-the-loop model that Western military doctrine had long assumed would govern lethal autonomous systems. Operators cannot individually authorize every strike when drones are being deployed at industrial scale, lost at rates exceeding 10,000 per month per side, and intercepted faster than radio commands can travel.
That operational reality is accelerating a transformation that defense analysts have warned about for years: the migration of targeting authority from human hands to onboard algorithms. Battlefield AI drones are no longer a speculative technology hedged inside think-tank white papers. They are a procurement priority, and the startups building them are scaling quickly.
The conflict in Ukraine has functioned as the most consequential live test bed in the history of autonomous weapons development. Electronic warfare environments have made GPS unreliable and remote piloting intermittent. Adversarial jamming has forced engineers to push more decision-making onto the drone itself—onto hardware that weighs grams and runs on a battery the size of a thumb drive. The solution the industry is converging on is edge AI: compact machine learning models trained to detect, classify, and, in some configurations, engage targets without waiting for a human to authorize the strike.
How Tiny AI Models Enable on-Device Target Detection
The technical challenge is not finding an AI model capable of identifying a tank or a troop formation. That problem was largely solved in commercial computer vision years ago. The challenge is compressing that capability into a model small enough to run on the microcontroller inside a sub-250-gram drone, with inference times measured in milliseconds and power budgets measured in milliwatts.
Read next Laika's Wildwood: Stop-Motion Fantasy at TIFF 2026Conventional neural networks trained for desktop GPUs are orders of magnitude too large. The field has responded with a cluster of compression techniques—quantization, pruning, knowledge distillation—that strip a model down to its most predictive parameters without catastrophically degrading accuracy. A model that originally required a server rack can, after aggressive optimization, run on a chip smaller than a fingernail.
The result is what researchers call edge inference: the model lives on the device, processes sensor data locally, and produces a targeting decision without any round-trip communication to a remote server. This matters enormously in contested electromagnetic environments, where an adversary can jam uplinks, spoof GPS coordinates, or intercept command signals. An autonomous drone running edge AI is, by design, immune to those attack vectors—because it never needed to phone home in the first place.
The tradeoff is well understood in the academic literature on human-machine teaming. Researchers at institutions including the RAND Corporation have documented the tension between operational resilience and meaningful human control: the same architectural feature that makes a drone survivable in a jamming environment is the feature that removes a human from the kill chain. These are not separable engineering choices. They are the same choice.
Inside the NATO-Backed Startup Repurposing Commercial AI for Combat
Scaleout Systems was founded in 2018 by researchers from Uppsala University in Sweden, and its original market had nothing to do with warfare. The company built technology for training and deploying machine learning models directly on the hardware embedded in commercial trucks and other vehicles—edge AI for logistics and transport, not for combat. The core capability was the same: get inference running on constrained hardware in environments where cloud connectivity is unreliable.
That commercial foundation proved more transferable than the founders anticipated. When Russia launched its full-scale invasion of Ukraine in February 2022, the strategic context around dual-use technology shifted abruptly across European defense establishments. Andreas Hellander, cofounder and CEO of Scaleout Systems, told Ars Technica that the company recognized its edge AI platform could give NATO allies a strategic advantage in operationalizing the sensor and data streams generated at the front line.
The pivot was not incidental. European militaries had spent years chronically underinvesting in defense technology relative to GDP targets, and the war exposed the consequences. NATO member states suddenly needed to absorb commercially developed AI capabilities quickly, and startups with proven edge deployment expertise found themselves inside procurement conversations they had not previously been part of. Scaleout's backing by NATO-affiliated entities reflects a broader institutional acknowledgment that the alliance's technological edge will increasingly depend on commercial AI talent, not exclusively on traditional defense primes.
The company's work now spans reconnaissance and attack mission profiles—surveillance drones that can identify objects of interest autonomously, and strike platforms that can act on those identifications. The distinction between the two categories is narrower than it sounds.
Selecting and Striking Without Human Sign-Off: What the Technology Actually Does
The phrase "autonomous targeting" obscures a meaningful spectrum. At one end, a drone might use onboard AI only to stabilize its camera and flag objects for a human operator to review. At the other end, the system identifies a target, classifies it as hostile, calculates an intercept trajectory, and initiates the strike—all without any human involvement after the mission is launched.
Scaleout Systems is operating somewhere on that spectrum, building AI-driven target detection and selection for both surveillance and attack drones. The operative question—the one that legal scholars, arms control negotiators, and military ethicists are fighting over in Geneva—is where exactly on that spectrum the human exits the decision loop, and whether that exit point is a design choice or an operational inevitability.
In practice, the distinction blurs under combat conditions. A system nominally designed to require human confirmation of a strike recommendation becomes effectively autonomous when the latency of that confirmation loop exceeds the window in which the target is actionable. If a human has three seconds to approve a strike on a fast-moving target and the communication link introduces two seconds of lag, the system is functionally autonomous regardless of what the user manual says.
Edge AI removes that lag entirely. The decision happens on the chip. The human is upstream of the mission, not inside it.
Legal and Ethical Fault Lines Around Autonomous Lethal AI
The International Committee of the Red Cross has been explicit on this point: autonomous weapon systems that select and strike human targets without meaningful human control may be incompatible with international humanitarian law. The ICRC's 2021 position paper called for new legally binding rules, arguing that existing frameworks—the laws of armed conflict, proportionality doctrine, distinction between combatants and civilians—presuppose a human decision-maker who can exercise judgment in context.
That position informs ongoing negotiations at the United Nations Convention on Certain Conventional Weapons, where states have been debating a framework for Lethal Autonomous Weapons Systems since 2014. Twelve years later, no binding treaty exists. Russia and the United States have each resisted hard prohibitions, and the pace of technological deployment has consistently outrun the pace of diplomatic consensus.
The Future of Life Institute and allied research organizations have documented the recursive problem this creates: every year that binding rules are delayed is a year in which more states and non-state actors develop and field autonomous targeting capabilities, increasing the political cost of prohibition and reducing the likelihood that any single treaty can achieve meaningful coverage.
Critics of the current trajectory argue that the problem is not merely legal but epistemic. A model trained to classify targets on a training dataset cannot, in principle, account for the context a human observer would use to distinguish a legitimate military target from a civilian carrying similar equipment. The RAND Corporation's work on human-machine teaming in lethal applications has consistently emphasized that meaningful oversight requires not just a human in the loop, but a human with enough situational awareness and authority to override the system in real time—a standard that edge AI architectures are structurally designed to circumvent.
Strategic Implications for NATO, Adversaries, and the Future of War
The emergence of battlefield AI drones with autonomous targeting represents a genuine discontinuity in military competition, not merely an incremental improvement in munitions. States that deploy these systems at scale can impose costs on adversaries faster than human-paced decision cycles can respond. The side that accepts autonomous targeting first gains a tempo advantage that is difficult to offset through conventional means.
For NATO, the strategic calculus involves acknowledging that adversaries—including Russia, which has demonstrated willingness to field autonomous systems in Ukraine without the alliance's legal and ethical constraints—are not waiting for Geneva to reach consensus. The backing of companies like Scaleout Systems reflects a calculation that NATO cannot afford to unilaterally cap its autonomous capabilities while peer competitors develop without similar restraint.
The counterargument, pressed by humanitarian law scholars and some serving military officers, is that autonomous targeting creates escalation risks that undermine the strategic logic of deploying it. A strike authorized by an algorithm rather than a human cannot easily be walked back, explained to an adversary, or calibrated to communicate political intent. The history of crisis management in nuclear competition taught that maintaining communication channels and interpretive clarity matters enormously. Autonomous battlefield AI drones, operating faster than human deliberation, compress those channels to zero.
What Scaleout Systems and the broader wave of defense AI startups are building is not just a product. It is a set of precedents—technical, legal, and political—that will define the character of armed conflict for the next generation. The models are tiny. The consequences are not.
Source: AI - Ars Technica
