OpenAI has identified the root cause of “hallucinations” in AI models, which leads to factually incorrect responses. This issue has become increasingly problematic as AI systems grow more advanced, undermining their reliability. Experts debate whether these hallucinations are an intrinsic flaw in large language models (LLMs). A recent OpenAI paper highlights that LLMs are incentivized to guess rather than admit uncertainty during training, which leads to more hallucinations. The traditional evaluation method rewards correct answers and penalizes wrong ones, encouraging guessing. OpenAI proposes a solution: modify scoring systems to penalize confident errors more than uncertainty, thus rewarding appropriate acknowledgment of not knowing. These changes could reduce hallucinations and enhance the development of more nuanced AI models. While OpenAI claims that their new GPT-5 has improved in this regard, user feedback remains skeptical. The AI industry must address these challenges as it faces rising costs and scrutiny over performance.
Source link
