OpenAI is working to improve the reliability and usefulness of its AI systems. One major challenge is “hallucinations,” where a model gives an answer that isn’t true, often due to training methods that reward guessing over acknowledging uncertainty.
This issue affects all large language models, including ChatGPT-5. Researchers found that GPT-5 has fewer hallucinations than other models but still experiences them. The problem persists because current evaluation methods encourage guessing rather than honesty about uncertainty.
The main problem lies in how these systems are trained and evaluated. In multiple-choice tests, taking a wild guess can be correct more often than leaving an answer blank. Similarly, when grading language models on accuracy, the focus is on getting right answers rather than admitting uncertainty.
To address this, OpenAI proposes changing evaluation methods to discourage guessing. This would encourage models to be more humble and admit when they don’t know. The company argues that simply adding new tests for uncertainty won’t solve the problem, but rather updating widely used accuracy-based evaluations.
Source: https://openai.com/index/why-language-models-hallucinate