Relocating Initiative and Interpretation with Agentic AI

Militaries around the world are deploying large language model-based agents for data fusion, intelligence analysis, and battle management. LLM-based agents in those capabilities, this paper argues, present a unique and potentially unmitigable risk to ability to retain human judgment in the kill chain.

What Makes LLM Agents Different

We’ve had AI tools on the battlefield for decades. Target recognition software. Automated sensor filtering. Rule-based alert systems. These tools are powerful, but they’re fundamentally deterministic: a human decides in advance what the system should flag, trust, or discard. And in most cases up ‘til now, human remains in or on the loop when such tools are deployed on the battlefield.

Large language model-based agents are categorically different. An LLM agent combines a language model with memory, external tools, and an orchestration layer that can reason, plan, and act and it can do so dynamically, without requiring pre-programmed rules for every contingency.

To understand the stakes, consider what a data fusion agent might actually do. A battle station receives feeds from hundreds of sensors: drones, satellites, ground patrols, intercepted communications. The data is incomplete, conflicting, and constantly changing. An LLM agent can read a fragmented drone observation, cross-reference it with a partially redacted signals intelligence report and a commander’s voice-transcribed intent, and surface a coherent operational picture, autonomously, and in real time.

This isn’t a hypothetical. Anduril’s Lattice C2 platform already describes itself as an “AI-powered battle-management platform” that turns “data into decisions at scales and speeds beyond human capacity,” and while there is no reason to think Lattice is an LLM agent, LLM agents are on their way to fill out these types of capabilities. The U.S. Department of Defense’s January 2026 AI Strategy explicitly calls for “unleashing AI agent development and experimentation for AI-enabled battle management and decision support, from campaign planning to kill chain execution.”

Four Features. One Problem.

The paper identifies four features that make LLM agents so operationally attractive for these roles (data fusion, and similar). The argument is that these same four features are what make human judgment substantively ineffectual where these agents operate.

Initiative: The agent can task sensors, switch data sources, and decide when to seek human input without being told to. It doesn’t go back to the operator each time it needs more evidence. It acts.

Interpretation: The agent can reason over unstructured, ambiguous, conflicting data: raw feeds, voice transcripts, imagery metadata, and produce a coherent picture. It doesn’t need data pre-formatted into rigid schemas.

Goal-Orientedness: Given a high-level directive, the agent decomposes it into sub-goals and reprioritizes them dynamically as context shifts, without waiting for a human to rewrite the task queue.

Dynamic Memory: The agent remembers past observations, past decisions, and past confidence levels and actively draws on that history when reasoning about new inputs. It can connect a current sensor reading to a chain of prior outcomes.

These features are genuinely impressive. They’re also why LLM agents outperform legacy AI systems for data fusion. The problem is that they systematically relocate what the paper calls normative and epistemic authority from humans to machines.

When an AI agent decides which data to surface, which sensors to trust, and how to translate a commander’s intent into action the human who ultimately “decides” is reasoning over a reality the machine already constructed.

By the time a human operator sees a recommended course of action, or data about a potential target, the key decisions that actually drive the final outcomes have already been made, and they have been made by the system, not by that human operator.

Why This Breaks Existing Governance Frameworks

The international community has spent years negotiating what “human control” means for AI weapons. The GGE-CCW rolling text calls for “context-appropriate human judgment and control” to ensure compliance with international humanitarian law. The REAIM process identifies the “erosion of human moral agency” as a central concern. NATO, the EU AI Act, and domestic U.S. policy all echo similar requirements.

LLM agents used for data fusion make these requirements unachievable, not because humans are absent, but because the human presence that remains is disconnected from the decisions that actually determine outcomes.

To illustrate: imagine a police officer who selectively presents evidence to a prosecutor -including the facts that point toward guilt, but omitting the video-supported alibi. The prosecutor charges the defendant in good faith, based on the evidence provided. Months later, the defendant is found innocent. Who drove that outcome? Not the prosecutor who filed the charges but the officer who curated what the prosecutor saw.

On the battlefield, the LLM agent is that officer. It curates what the human sees. The human who signs off on the final strike is reasoning over a picture the system already built. That is not context-appropriate human judgment. It is window dressing.

Answering the Objections

“Human judgment happens in development and testing, not just deployment.”

It’s true that human engineers design these systems and human evaluators test them. But the goal-orientedness and dynamic memory of LLM agents mean that neither developer intent nor commander intent can sufficiently explain any particular outcome. The gap between pre-deployment human judgment and real-world battlefield outcomes is too wide, and too shaped by the system’s own in-context reasoning, to meet the purposes for which human judgment is required in the first place.

“Human oversight and ‘on the loop’ monitoring is sufficient.”

Real-time human oversight of a stochastic LLM agent processing thousands of data streams simultaneously is not feasible. Any meaningful oversight would itself have to be mediated by another AI system which raises the same questions one level up. “Humans overseeing agents overseeing agents” is a governance structure, but it’s a different governance structure than the one existing policy frameworks were designed for.

“A human still makes the final targeting decision.”

This is the most intuitive objection, and the paper takes it most seriously. Yes, a human presses the button. But if the primary determinants of which target is identified, assessed, and presented for engagement were all decisions made by the agent what is that final human authorization actually authorizing? Reasoning over a curated subset of evidence chosen by an LLM agent is not the same as exercising independent judgment. The button-press is not where the decision was made.

Two Paths Forward

The paper does not argue that this problem is unsolvable. It argues that existing responses which all insist on nominal human involvement are the wrong solution.

A ban on LLM agents for data fusion and battlefield management is the logical implication of the argument, and the paper takes it seriously as a normative ideal. But a blanket ban is unrealistic: these systems are already being built, sold, tested, and in some cases deployed. Calling for a ban without a credible enforcement mechanism risks becoming a governance fiction of its own.

The more actionable path involves four concrete reforms:

First, stop treating nominal human involvement as a de-risking strategy. In parts of the kill chain where effectual human judgment is not structurally possible, insisting on it as a checkbox doesn’t produce safety or accountability- it just gives the appearance of both.

Second, invest seriously in new TEVV (testing, evaluation, validation, and verification) methods designed for the specific opacity and stochasticity of LLM agents, including red teaming, real-world evaluation, and human-machine team testing.

Third, develop meaningful technical oversight mechanisms: counterfactual real-time processing, guardrails, ensemble models, and elements of the orchestration layer that are not themselves interpreted by an LLM.

Fourth, if we retain human judgment elsewhere in the kill chain, be explicit about what that judgment is actually about — and design training and responsibility attribution frameworks accordingly. Human oversight of a governance agent overseeing a data fusion agent is oversight of something, but it is not oversight of the decisions that drive battlefield outcomes.

Why This Matters Now

This is not a speculative paper about future risks. The Defense Innovation Unit announced a $100 million prize for autonomous vehicle orchestrators in January 2026, explicitly seeking systems that allow humans to command “through plain language” which is a description of an LLM agent. The U.S. AI Strategy for the Department of War calls for “agentic networks for warfighting.”

The governance conversation is still operating on assumptions built for an earlier generation of AI. The GGE-CCW rolling text was drafted with more deterministic systems in mind; systems where human intent could plausibly be traced through the decision chain to an outcome. LLM agents break that traceability at the architectural level.

This is not an argument against AI in military contexts. It is an argument for honesty: about what these systems actually do, about where human judgment actually operates, and about what it would genuinely mean to govern them responsibly. The alternative, namely insisting that humans remain “in the loop” while the loop has moved somewhere they cannot reach is a governance fiction that protects no one.

Davidovic, J. (2026). “The End of Human Judgment in the Kill Chain? Relocating Initiative and Interpretation with Agentic AI.” arXiv:2604.06300 [cs.CY]. This research was sponsored by the Research Council of Norway, grant number 352870. The author presented this work at the 2026 REAIM Summit.