ChatGPT doesn’t lie because it doesn’t know the truth

The entire debate surrounding ChatGPT and other models hallucinating or inventing responses is based on a significant misunderstanding.

The fundamental goal of GPT models is for machines to have linguistic abilities, allowing them to interact with users like a real person. These models have been trained on billions of web pages to generate content, not to precisely answer all possible questions. And there are technical and philosophical reasons for this.

To use this models a source of factual data, it is necessary to custom-train the model with generic (industry-level) or company-specific databases. The accuracy of the answers will be as good as your knowledge base. In other words, for the model to provide correct answers to objective questions, it needs to be trained on specific topics.

This may disappoint those who think of ChatGPT as a talking encyclopedia. However, it is good news for businesses that can invest in preparing the model to provide real and objective data for customer service, research, etc.

For those who enjoy coding, I go step by step into how and why GPT hallucinates and how to avoid it in this gist.