U.S. Army War College faculty conducted a groundbreaking experiment in February 2026, administering rigorous oral comprehensive exams to four prominent AI models—ChatGPT, Google Gemini, Anthropic’s Claude, and xAI’s Grok—instead of students. While all models passed, researchers Kevin Boyce and John Nagl discovered a critical flaw: these digital "students" degraded during extended questioning due to technical computing limits, producing repetitive and lazy responses.
This project highlights AI's power for historical recall but underscores human judgment's indispensability for high-pressure decisions, particularly in complex strategic thinking. Senior leaders should treat AI as a capable but imperfect staff officer, asking precise questions and carefully verifying output, acknowledging its probabilistic thinking and potential for hallucinations, as Professor Kris Wheaton noted, evolving from a "mediocre" to a "good" staff officer.
No comments:
Post a Comment