Some rare examples of AIs being underconfident (futuresearch.ai)

A recent analysis highlights a curious phenomenon in AI forecasting, where models such as Claude Opus 4.6 exhibit underconfidence, contradicting their own well-reasoned analysis. While previous AI models were criticized for presenting overconfident conclusions—often leading to inaccuracies in critical fields like legal analysis—this behavior raises new concerns. For instance, in evaluating the 2025 NYC mayoral election ballots, Claude made accurate calculations but only forecasted a 25% chance of surpassing the threshold, despite evidence suggesting a much higher likelihood. This tendency to downplay certain conclusions may be a byproduct of reinforcement learning from human feedback (RLHF), where AIs are trained to avoid extreme assertions. This underconfidence poses significant implications for the AI/ML community, especially concerning the reliability of AI when making predictions. The mismatch between nuanced analysis and conservative forecasting can lead to suboptimal decision-making. The findings underscore the importance of examining not just the final predictions of AI systems but also their analytical reasoning. As AI models continue to evolve, maintaining a balance between caution and accuracy will be crucial to enhance their utility in providing reliable forecasts and insights, particularly in dynamic or unpredictable environments.