OpenAI罕见发论文：我们找到了AI幻觉的罪魁祸首

Core Insights - The primary challenge in AI is the phenomenon known as "hallucination," where models confidently generate false information, making it difficult to discern truth from fiction [1][2][3] - OpenAI has acknowledged that while GPT-5 exhibits fewer hallucinations, the issue remains a fundamental challenge for all large language models [1][2] Definition and Examples - Hallucination is defined as the situation where a model confidently produces incorrect answers [4][5] - OpenAI provided examples where different chatbots confidently gave incorrect titles for a doctoral thesis and incorrect birth dates for the same individual [4][5] Evaluation Methods and Incentives - Current evaluation methods incentivize guessing rather than admitting uncertainty, leading to persistent hallucinations [6][11] - Models are often scored based on accuracy, which encourages them to guess rather than abstain from answering when uncertain [6][11] Proposed Solutions - OpenAI suggests that evaluation metrics should penalize confident errors more than uncertain responses and reward appropriate expressions of uncertainty [12][13] - The company emphasizes that merely adding uncertainty-aware tests is insufficient; widely used accuracy-based evaluations need to be updated to discourage guessing [12][13] Nature of Hallucinations - Hallucinations arise from the nature of language models predicting the next word without clear "true/false" labels, making it difficult to distinguish valid from invalid statements [15][16] - The randomness of certain factual information, like birthdays, contributes to the occurrence of hallucinations, as these cannot be reliably predicted [15][16] Misconceptions Addressed - OpenAI refutes the notion that hallucinations can be eliminated by achieving 100% accuracy, stating that some real-world questions are inherently unanswerable [17][20] - The company also clarifies that hallucinations are not inevitable and that smaller models can better recognize their limitations compared to larger models [19][20] Future Directions - OpenAI is reorganizing its Model Behavior team to focus on improving how AI models interact with users, indicating a commitment to reducing hallucination rates further [21][22]