TALK: Reasoning Models // Non-Reasoning Humans
Mauro Viganti, ICRC
ABSTRACT
As artificial intelligence becomes increasingly integrated into military decision-support systems (AI-DSS), human operators may progressively rely on large language models to support complex and time-sensitive decisions. While AI can alleviate cognitive overload, excessive reliance may encourage cognitive offloading or even “cognitive surrender”; conversely, opaque AI recommendations can generate algorithmic aversion. Recent research suggests that providing human-readable explanations of AI recommendations can mitigate this aversion and foster more calibrated trust. In military decision-making, access to a model’s chain-of-thought (CoT) could therefore provide a targeteer or commander with an additional means to scrutinise, challenge, or reject an AI-generated course of action.
The emergence of GPT-6 Astra and Claude Mythos, however, may signal a changing paradigm. Previous research found that reasoning models struggled to control their CoT and that controllability decreased through reinforcement learning, providing cautious optimism that CoT could remain a useful monitoring signal. Astra and Mythos substantially reverses this trend: their CoT controllability is markedly higher than previous reasoning models and, unlike them, increases during reinforcement learning. At the same time, Astra displays stronger “no-CoT” capabilities—the ability to solve tasks or act without requiring verbalised reasoning—and produces shorter and less informative reasoning traces.
As hybrid models become less dependent on explicit CoT to reach an answer, task-solving and observable reasoning may increasingly decouple. This creates a potential trust– faithfulness gap for military AI DSS: human operators may increasingly depend on visible reasoning to calibrate trust in AI-generated courses of action precisely as that reasoning becomes less reliable as a window into the processes generating them. The talk examines this emerging tension and asks what should replace or supplement CoT as a basis for meaningful human scrutiny and calibrated trust.
Mauro Viganti
Mauro Vignati is an advisor on digital technologies of warfare at the International Committee of the Red Cross (ICRC), focusing on aspects such as the militarisation of digital space and AI, the exploitation of the cognitive domain as a means of confrontation, and the impact of digital technologies in armed conflicts on civilians, civilian critical infrastructure, and society as a whole. Mauro gained 20 years of experience in various cybersecurity positions.
