NEW
Why LLM Hallucinations Aren’t Bugs
Watch: Why Large Language Models Hallucinate by IBM Technology LLM hallucinations aren’t bugs-they’re a byproduct of how these models are trained, evaluated, and incentivized to perform. Understanding this requires examining the interplay between statistical prediction, evaluation metrics, and the limitations of training data. When models generate text, they’re not solving for factual accuracy but rather selecting the most statistically likely next word. This creates a system where confident, false statements emerge as a natural consequence of the design, as detailed in the The Nature of LLM Hallucinations section. Large language models (LLMs) are trained using next-word prediction, a task that rewards statistical fluency over factual correctness. For example, OpenAI’s GPT-5 “thinking-mini” model abstains from answering 52% of questions, while its counterpart o4-mini abstains just 1% of the time. The trade-off? O4-mini’s hallucination rate soars to 75%, compared to 26% for GPT-5. This stark contrast reveals how evaluation metrics -which prioritize accuracy over honesty-create a “guess-and-win” incentive. Models that abstain are penalized in leaderboards, even if their uncertainty is prudent in real-world scenarios.
