OpenAI has acknowledged that artificial intelligence models will never reach 100% accuracy, saying some real-world questions are inherently unanswerable despite advances in search and reasoning.
In a blog post accompanying a new research paper published Friday, the maker of ChatGPT said hallucinations — where AI systems confidently generate false information — persist due to flaws in training and evaluation.
“Most evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty,” the company said. It noted that smaller models often generate fewer hallucinations by recognising their own limits.
The research comes weeks after the launch of GPT-5, OpenAI’s flagship model, which the company says reduces hallucinations but does not eliminate them. Researchers proposed overhauling accuracy benchmarks to penalise “confident errors” and reward models that admit uncertainty.
“Fixing scoreboards can broaden adoption of hallucination-reduction techniques,” the paper concluded, urging developers and regulators to rethink evaluation standards.
