Tonight we put candles on pumpkins and tell each other stories about spectres that wander houses and halls. The military has its own set of tales for All Hallows Eve. They are less theatrical but no less unnerving. The ghosts that haunt twenty first century battlefields are not metaphysical. They are algorithms, emergent behaviors, legacy code, and design assumptions that behave in ways their makers did not fully foresee. These are the ghosts in the machine of modern warfare.
We must be precise about what I mean by ghosts. In the literature of robotics and multi agent systems the term emergent behavior is used to describe system level outcomes that are not explicitly encoded in any single component but arise from interactions among many simple agents. Swarm robotics research shows how useful, robust, and clever patterns can arise from minimal local rules. Those same patterns are a double edged sword when placed in kinetic or contested environments. In short, emergence explains how collective competence can also become collective surprise.
This is not merely theoretical. Military institutions and legislatures are wrestling with whether autonomy will shrink the circle of human risk or expand responsibility without any clear locus for it. The United States Department of Defense updated its governing policy on autonomy in weapon systems in January 2023, formalizing categories such as human in the loop, human on the loop, and human out of the loop, and prescribing rigorous review, verification, and testing before fielding. Those reforms matter, but policy papers and doctrine cannot eliminate emergent failure modes that occur when many systems interact or when adversaries probe weak edges.
International fora reflect the same tension. States and civil society debate whether to prohibit autonomous lethal systems that cannot be used with meaningful human control, or to regulate them more narrowly. Meetings under the Convention on Certain Conventional Weapons have continued these discussions through 2024 while human rights organizations and many states press for legally binding limits. The debate is no longer abstract. Leaders and diplomats warn that time is not infinite for crafting rules that preserve human judgment in the use of force.
Why call these phenomena ghosts? Because the path from specification to deployment is littered with residuals. Systems inherit training datasets and legacy interfaces. They learn from historical signals that embed bias and context that no longer apply. They are vulnerable to data poisoning and adversarial manipulations. They can be triggered by edge conditions that were not part of test suites. When these latent vectors activate in combination across networked platforms the result is a behaviour that appears to spring ex nihilo. Security analysts have started using metaphors like ghouls and phantoms to convey how subtle triggers can lead to outsized effects.
Consider some concrete failure modes that read like folktales for engineers. A swarm of small systems may develop a locally optimal coordination pattern that, in aggregate, blocks evacuation routes or concentrates kinetic effects on unintended infrastructure. An AI perception stack trained in benign environments may misclassify targets when confronted with a new sensor modality or with spoofed signals. A chain of automated decision aids can amplify a minor sensor error into a cascading command that requires human commanders to untangle consequences after the fact. These are not horror tropes. They are documented classes of technical risk studied in both the robotics and reliability literature.
Policy responses to these risks must be twofold and they must be honest about trade offs. First, governance should build on the precautionary architecture already present in some national doctrines. DoD Directive 3000.09, its associated responsible AI implementation pathways, and recent congressional oversight language require senior review, hardware and software verification and validation, and operational testing. Those measures are necessary but not sufficient. Systems engineering that presumes isolated testing will fail to capture multi system interactions that only materialize in the wild.
Second, the engineering stack needs design patterns that make ghosts visible. Those patterns include strict action boundaries and circuit breakers that forbid autonomous application of lethal force in ambiguous contexts, intentionally constrained communication channels between agents, aggressive adversarial testing and red teaming against plausible misuse scenarios, and layered human oversight that is both meaningful and realistically paced for the operational tempo. Sandbox deployments and staged fielding with rollback plans should be standard practice rather than hopeful exceptions. The technical literature on swarm resilience and partial failure modes offers concrete metrics and evaluation strategies that can inform those practices.
There is a political and moral argument that sits behind the engineering prescriptions. When machines make or materially assist life and death decisions we confront an accountability gap. If no individual can foresee and be held responsible for collective emergent behavior, then the social license for deploying such systems erodes. Human rights organizations and a broad array of states rightly insist that rules be found to preserve human control and legal accountability. At the same time, militaries argue that some autonomous functions can reduce risk to civilians and friendly forces when correctly constrained and supervised. Reconciling these views requires honest empirical testing, transparent governance, and international cooperation.
Finally, let us not romanticize the supernatural. The ghosts in our machines are not sentient. They do not conspire. They are the failures of our models to capture reality, the residues of old software, the unanticipated couplings of distributed agents, and the exploitation vectors an adversary will test. Halloween is a useful reminder that fear without understanding is paralyzing and that courage without caution is reckless. We should be afraid enough to act prudently and boldly enough to build systems whose failures are understandable, traceable, and repairable.
On this night of masks and misdirection, I offer a modest prescription. Treat autonomy as an experiment in social technology as much as a technical one. Demand robust, independent testing that exercises systems in realistic multi actor environments. Insist on design limits that enforce human judgment where it matters. And commit, in national and international fora, to frameworks that distribute responsibility rather than diffuse it into a phantom. The true way to lay these ghosts to rest is not by exorcism. It is by engineering, law, and ethics that render the invisible visible and the unpredictable manageable.