Why AI chatbots hallucinate, according to OpenAI researchers (www.businessinsider.com)

OpenAI researchers have identified a key reason why large language models (LLMs) frequently hallucinate—confidently generating false information presented as fact. Their new research reveals that current training and evaluation methods incentivize LLMs to "guess" answers rather than admit uncertainty. Because most metrics reward accuracy above all, models optimize for test-taking strategies that favor guessing, even when unsure, leading to persistent hallucinations across popular models like GPT-5 and Anthropic’s Claude. This insight is significant for the AI community as it points to a fundamental misalignment between how LLMs are evaluated and how humans naturally handle uncertainty. Unlike humans who learn to express doubt through real-world experience, LLMs are primarily trained on exams that penalize abstaining or indicating uncertainty. OpenAI suggests that redesigning evaluation metrics to stop penalizing abstentions and discourage guessing could dramatically reduce hallucinations. By shifting away from purely accuracy-based scores toward more nuanced assessments that reward honest uncertainty, future models may become more reliable and trustworthy in their outputs, enhancing real-world applicability and user trust in AI systems.