According to Ethan Mollick on X, a widely circulated paper testing large language models in military decision-making includes prompts that prime aggressive escalation, such as lines like “Failure to act preemptively means certain destruction,” which can bias models toward first-strike choices; as reported by Ethan Mollick, this critique underscores that AI should not be entrusted with lethal command decisions. According to the original paper’s authors as cited by Ethan Mollick, the study used role-play scenarios to evaluate model behavior in high-stakes conflict, but the embedded threat framing may confound results by rewarding preemption, raising concerns about construct validity and external reliability. As reported by Ethan Mollick, this debate highlights urgent needs for red-team evaluation protocols, neutral baselines, and transparency in prompt design so defense and dual-use sectors can avoid overestimating LLM readiness for command-and-control. According to Ethan Mollick, the business implication is clear: vendors pursuing defense contracts must demonstrate prompt-robustness, calibrated risk preferences, and audit trails that regulators and acquisition officers can verify.