It's Hard to Learn from Machines (carlkolon.com)

A recent study titled “Human Learning from Artificial Intelligence: Evidence from Human Go Players’ Decisions after AlphaGo” highlights the challenges in bridging understanding between human players and AI systems in games like Go. While human players have improved since AlphaGo's introduction, the study reveals that this enhancement is primarily limited to the early game (before move 50). The research underscores that AI, such as AlphaGo, relies on complex heuristics—essentially learned rules for evaluating positions—which are difficult for humans to internalize and utilize. Instead of directly imparting knowledge, AI's decision-making must be interpreted by observing their gameplay, creating a gap in understanding. The implications for the AI/ML community are significant, particularly regarding explainability in AI. As AI systems, including large language models (LLMs), are trained via reinforcement learning (RL), the expectation that these models can articulate their reasoning is questioned. The study suggests that while LLMs like those used in Go may generate plausible explanations for their moves, these responses may not accurately reflect the internal heuristics guiding their decisions. This raises important concerns about the reliability of AI explanations and the broader applications of RL in developing interpretable AI systems, emphasizing the need for more integrated human-AI learning frameworks.