Study: Soldiers Trust AI Targeting Less Than Humans — Until You Explain It

Most policy discussion about AI in high-stakes decision-making assumes the risk runs one direction: give operators a confident-sounding algorithm and they will defer to it, oversight becomes theatre, and "human in the loop" turns into rubber-stamping. A new peer-reviewed study, flagged by Bruce Schneier, tests that assumption directly — and finds the opposite, at least under one set of conditions.

What the study actually did

Researchers Ryan Shandler, Michael L. Gross and Yahli Shereshevsky, publishing in the *Journal of Conflict Resolution*, reconstructed the interface and functionality of a real Israeli Defense Forces AI targeting decision-support system, drawing on open-source materials, FOIA requests, and military interviews. They ran two experiments totaling 2,015 active-duty and veteran Israeli soldiers, presenting 12 targeting scenarios — target images, attack methods, success ratings, collateral-damage estimates, confidence scores — with each scenario randomly attributed to either an AI system or a human analyst.

The dependent measures were strike-approval rate, decision confidence, and response time — a clean design for isolating how the *source label* alone shifts human judgment, independent of the underlying recommendation.

Aversion, not bias — but conditional

Participants approved 63% of AI-attributed recommendations versus 68% of identical recommendations attributed to a human analyst (p < 0.001), with lower stated confidence (3.91 vs 4.00 on a 5-point scale). This is algorithmic aversion: people trusted the machine less than an equivalent human source, not more. The effect wasn't uniform — in high-collateral-damage scenarios the approval gap widened to 7.3 percentage points, while low-risk scenarios showed no significant difference. In other words, skepticism of AI scaled with the stakes, which is roughly the behavior a governance framework would want.

The part that should worry oversight designers

The second study added a simple explainability layer — three data points showing what the AI's recommendation was based on. That single change was enough to eliminate the aversion gap entirely (67% vs 65%, no longer statistically significant). Explanation didn't just build appropriate trust; it fully closed the gap between how people treated machine and human recommendations, including in the high-stakes scenarios where healthy skepticism had been strongest.

That is a genuinely double-edged result. Explainable AI is a stated goal of frameworks like ISO/IEC 42001 and the EU AI Act's human-oversight provisions, and this data shows explanation features do change behavior — operators engaged more deliberately with the AI's reasoning rather than reflexively discounting it. But the same data shows explanation can also flatten the very skepticism that acted as a brake in the riskiest scenarios. A three-data-point summary is a thin basis on which to fully re-equalize trust in a collateral-damage decision, and the study doesn't tell us whether that trust was better *calibrated* afterward or simply less resistant.

Why this matters beyond the battlefield

The same dynamic applies to any organization deploying AI decision-support in consequential settings — SOC alert triage, fraud holds, medical or safety-critical recommendations. "Human in the loop" is frequently treated as a checkbox control, but this study is rare empirical evidence that the loop can function as intended — and that interface design, specifically how much explanation is surfaced and how, is what determines whether it does. Auditing an AI system's human-oversight control shouldn't stop at confirming a person can override the model; it should test whether the interface pushes that person toward genuine judgment or toward automatic concurrence.