The Israeli Defense Forces use AI systems to select strike targets in Gaza, with the final attack decision made in under 10 seconds. The IDF reports a 10 percent error rate. That number sounds tolerable until you work through the base-rate math.
A 10 percent error rate means the system correctly identifies 90 percent of targets. But if only 1 percent of the population in a given area are legitimate military targets, positive predictive value collapses. The vast majority of flagged individuals in such a scenario would be civilians. This is a textbook base-rate problem, well understood in medical screening and fraud detection, yet apparently treated as acceptable in automated kill chains where decisions happen faster than a human can meaningfully review them.
The speed compounds the problem. A 10-second decision window does not allow for contextual judgment. The human operator becomes a rubber stamp on a machine recommendation, not a genuine check on the system's output.
The Ukraine conflict has introduced a different category of risk: autonomous drone swarms that operate without a human-in-the-loop. Unlike individually piloted drones, swarms are immune to electronic jamming because they do not rely on a remote control signal. Once launched, there is no mechanism for human intervention.
This matters because it crosses a line that military ethicists have debated for decades: the delegation of lethal force to a system with no ability to recall or override. A jammed radio-controlled drone falls from the sky. A jammed autonomous swarm continues its mission.
Predictive policing applies similar statistical methods to domestic law enforcement, using historical crime data to forecast where future crimes will occur and directing patrols accordingly. The structural problem is that the training data encodes existing enforcement patterns, not actual crime distribution.
If police historically patrol certain neighborhoods more intensively, those neighborhoods generate more arrests. A model trained on this data recommends more patrols in the same neighborhoods, producing more arrests, which further reinforces the pattern. The result is a self-amplifying discrimination cycle that looks objective because it is driven by data, but the data itself is the product of biased enforcement.
A recurring structural feature across these applications is asymmetry. Defensive AI systems must cover every possible vulnerability. Offensive systems only need to find one that works. In cybersecurity, this means an AI-powered attacker can probe thousands of attack vectors while the defender must secure all of them. In information warfare, generating disinformation is orders of magnitude faster and cheaper than fact-checking it.
This asymmetry is not a temporary engineering problem. It is inherent to the relationship between attack and defense in systems where the cost of generating threats is lower than the cost of verifying them.
The mismatch between civilian and military AI deployment standards is striking. Companies hesitate to deploy customer-facing chatbots because a single embarrassing response could damage their brand. Meanwhile, military AI systems making lethal decisions are fielded with error rates that would be unacceptable in a spam filter.
A coherent framework would scale safety requirements with potential harm. Systems that recommend products need modest safeguards. Systems that select bombing targets need extraordinary ones. The gap between these two standards in current practice is difficult to justify on any technical or ethical basis.
In conflicts such as Gaza, the military uses AI systems for target selection with decision times under 10 seconds. The IDF reports a 10% error rate. In Ukraine, autonomous drone swarms operate without human control and are immune to electronic jamming, removing the possibility of human intervention once launched.
Predictive policing risks amplifying existing biases. When historically biased police data is used for training, self-reinforcing cycles of discrimination emerge. Increased police presence in certain areas leads to more arrests, which further distorts the statistics.
The AI threat is asymmetric because defensive systems must cover all potential vulnerabilities, while offensive systems only need to find one successful attack vector. This is particularly evident in cybersecurity and disinformation, where generating fake content is far easier than debunking it.
The reliability of military AI systems is problematic. Even a seemingly good detection rate of 90% can lead to a high number of false targets with imbalanced datasets. When only a small percentage of the population are legitimate military targets, the probability of misidentification rises dramatically.
Requirements for AI systems must scale with their potential for harm. The greater the possible damage, the higher the requirements for reliability and verifiability must be. Systems that make life-and-death decisions demand rigorous safety measures and comprehensive testing.
.
Copyright 2026 - Joel P. Barmettler