Investment Rating - The report does not explicitly provide an investment rating for the industry or specific companies within it [20][21]. Core Insights - The report by OpenAI discusses the intrinsic mechanisms behind hallucinations in large language models, attributing these to representational shifts under probabilistic generation paradigms and biases in training data [8][9]. - It proposes a governance framework that includes both evaluation system construction and training process optimization to enhance AI trustworthiness [8]. - The study highlights that even with correct training data, the probabilistic nature of pre-training objectives leads to a certain rate of erroneous generation [9][10]. - The report emphasizes the need for improved evaluation mechanisms to mitigate hallucination risks, suggesting the incorporation of "confidence thresholds" in scoring systems [12]. Summary by Sections Event - OpenAI released a report titled "Why Language Models Hallucinate" on September 4, 2025, explaining the mechanisms behind hallucinations in language models and proposing a governance framework for AI trustworthiness [8]. Mechanisms of Hallucination - The report identifies that high sparsity of certain facts in training data, such as personal birthdays, contributes to hallucinations, with a "singleton rate" quantifying this sparsity [10]. - It establishes a theoretical basis for why high-frequency common knowledge is generally accurate while low-frequency long-tail knowledge is more prone to errors [10]. Evaluation Mechanisms - Current evaluation benchmarks use a binary scoring system that incentivizes guessing rather than abstaining from uncertain responses, which increases the tendency for models to generate fabricated answers [11]. - The report suggests that setting a "confidence threshold" (e.g., only answering when confidence exceeds 75%) could improve model reliability and align evaluation frameworks with practical safety requirements [12].