OpenAI said in its blog post on September 5 that hallucinations, which are plausible but false outputs generated by AI systems, remain a persistent challenge for large language models, including its latest GPT-5 system.
The company is calling for changes to evaluation methods that currently reward models for guessing rather than acknowledging uncertainty. OpenAI added that hallucinations can show up in surprising ways, even for seemingly straightforward questions.
According to the company, the problem stems in part from how models are tested. Evaluations typically measure accuracy alone, which encourages systems to take risks instead of being cautious. “If you do not know the answer but take a wild guess, you might get lucky and be right. Leaving it blank guarantees a zero,” OpenAI noted.
The company compared this to standardised tests that use negative marking. “Penalise confident errors more than you penalise uncertainty, and give partial credit for appropriate expressions of uncertainty,” it said, citing a paper led by OpenAI staff published on arXiv.
Data from OpenAI’s GPT-5 System Card illustrates the issue. On the SimpleQA benchmark, the older o4-mini model achieved a slightly higher accuracy rate of 24% compared with 22% for gpt-5-thinking-mini. But o4-mini produced wrong answers 75% of the time, versus 26% for the newer system, which abstained more often.
“Most scoreboards prioritise and rank models based on accuracy, but errors are worse than abstentions,” OpenAI stated. “Our Model Spec says it is better to indicate uncertainty or ask for clarification than provide confident information that may be incorrect.”
The company also outlined why hallucinations arise from the way language models are trained. Pretraining relies on predicting the next word from large amounts of text, without negative labels. While models quickly learn consistent patterns like spelling, they struggle with arbitrary factual details such as birthdays. “Arbitrary low-frequency facts cannot be predicted from patterns alone and hence lead to hallucinations,” the paper said.
OpenAI emphasised that hallucinations are not inevitable, but reducing them requires a shift in how performance is measured. “Accuracy will never reach 100% because, regardless of model size, search and reasoning capabilities, some real-world questions are inherently unanswerable,” the company said.
It concluded that the company’s latest models show lower hallucination rates but acknowledged that “confident errors” remain a fundamental challenge.
