AI and Humans Misunderstand Uncertainty, USC Study Finds

Researchers at the USC Viterbi School of Engineering discovered a significant mismatch in how humans and artificial intelligence models interpret words of uncertainty ^[1]. Terms like "probably" or "likely" often hold different mathematical meanings for machines than they do for people ^[1]. The study highlights that AI models frequently express higher confidence levels than a human speaker originally intended ^[1].

This communication gap suggests that AI may overestimate the certainty of human statements during interactions ^[1]. While a person might use a vague term to express caution or doubt, a model could interpret the phrase as a definitive prediction ^[1]. These discrepancies could lead to errors in decision-making or a breakdown of trust between users and technology ^[1].

The findings emphasize the need for better alignment in how AI systems process natural language ^[1]. Researchers suggest that future models must be calibrated to reflect human nuance more accurately ^[1]. This ensures that "likely" remains a measure of probability rather than an unintentional signal of certainty ^[1].