Generative AI with Python

9.2    Limitations and Issues of LLMs

Still many possible issues exist in the context of large language models (LLMs). We’ll touch on the most relevant ones in this section: hallucinations, biases, misinformation, intellectual property, transparency, and jailbreaking.

9.2.1    Hallucinations

LLMs always produce an output if you don’t actively constrain them. Thus, they’ll even provide outputs if they don’t have the answer in their model weights. In the early days, this problem led to unsatisfactory results because the answers were factually incorrect. In some cases, the output was disconnected from the user query; it might be nonsensical or conflict with previous outputs in the same discussion.

Much effort was made to improve model performance, ...