An experiment shows that ChatGPT’s hallucinations can be significantly reduced when the model is prompted to distrust its own answers. With appropriate prompts, the AI acts like a “critical reviewer,” treating unverified claims as false and clearly flagging uncertain information. As a result, responses become more cautious, analytical, and reliable.

Hallucinations cannot be eliminated completely, but they occur less frequently. The model is less likely to produce overly confident or fabricated answers and is more inclined to highlight potential errors. This approach demonstrates that well-designed prompting can significantly improve the reliability of large language models by encouraging critical self-checking and a structured way of dealing with uncertainty.