OpenAI confirmed that its latest AI model, GPT-5, still produces errors known as “hallucinations,” where it generates answers that appear convincing but are factually incorrect. While the rate of these errors has declined compared to previous versions, the company considers this a persistent issue in how language models operate.
اضافة اعلان
These hallucinations stem from the way the systems are trained: by predicting the next word in a text. This method works very well for language tasks like grammar and style but struggles with precise or rare factual information. As a result, the model can produce confident-sounding responses that are inaccurate in reality.
Internal testing shows that improving overall accuracy does not necessarily reduce hallucinations. Some earlier models achieved higher rates of correct answers but also generated more confidently wrong responses. GPT-5, on the other hand, exhibits slightly lower overall accuracy but is better at avoiding guesses when knowledge is lacking, making it more cautious and less prone to hallucination.
OpenAI points out that part of the problem lies in current industry evaluation methods, which often focus solely on the percentage of correct answers while ignoring the danger of incorrect answers presented as facts. The company advocates for revising evaluation standards so that models are allowed to acknowledge when they do not know something, rather than forcing them to give confidently wrong answers.
For users, the main takeaway is that GPT-5 represents an advancement in reducing hallucinations, but it has not eliminated them entirely. The real challenge remains in improving training and evaluation methods to build more reliable models and narrow the gap between what seems correct and what is truly accurate.
