It is a truth universally acknowledged that an AI company in possession of a powerful GPU cluster must be in search of a solution to hallucinations. After all, the prize is nothing short of the most transformative market of the modern era: a reliable, trustworthy AI assistant capable of handling critical tasks without fabricating facts. Yet, despite three years since OpenAI launched ChatGPT, progress in eliminating hallucinations has been modest—far from sufficient to enable widespread, high-stakes deployment across industries. The root cause may finally be coming into focus. In a new research paper titled “Why Language Models Hallucinate,” OpenAI’s researchers present a compelling explanation: language models hallucinate because their standard training and evaluation methods reward guessing over admitting uncertainty. This insight, while conceptually familiar, gains new weight through the paper’s rigorous analysis and proposed remedies. The core argument is that during training, models are incentivized to produce a response—any response—rather than say “I don’t know.” In traditional setups, models are scored based on how closely their output matches a reference answer, regardless of whether that answer is correct or fabricated. As a result, models learn to generate plausible-sounding text even when they lack real knowledge, much like a student who chooses an answer on a multiple-choice test just to avoid leaving it blank. The paper highlights that this design flaw persists in evaluation as well. Benchmarking systems often penalize models for not providing an answer, even when the correct response is “I don’t know.” This creates a feedback loop: models learn that guessing is safer than being honest, leading to frequent hallucinations. What’s new is the suggestion that this problem isn’t inherent to the architecture of language models but stems from flawed incentives. The researchers propose that by retraining models with evaluation metrics that explicitly reward uncertainty—such as using a “don’t know” option or penalizing overconfidence—hallucinations could be significantly reduced. They tested this idea in controlled experiments, showing that models trained under such revised protocols were far more likely to admit ignorance when uncertain, and when they did answer, their responses were more accurate. The results suggest that the path to trustworthy AI may not lie in bigger models or more data, but in smarter training objectives. If validated at scale, this shift could mark a turning point. It would allow AI systems to operate with greater reliability in sensitive domains like healthcare, law, and finance—where false confidence is dangerous. The implications extend beyond technical improvement; they challenge the very foundation of how we train and assess AI performance. For now, OpenAI’s findings offer a promising roadmap. Solving hallucinations may not require a breakthrough in neural architecture, but a fundamental rethink of what we reward in AI development. The crown of trustworthy AI may finally be within reach—if the industry is willing to stop rewarding guesses and start valuing honesty.