Large language models are often seen as "liberal arts students" excelling in transactional writing, but errors are not allowed in these highly regulated writings. Large language models frequently experience "hallucination" issues, such as inconsistencies with facts or fabricating content, which is currently the most troublesome problem for engineers and users.
However, some people believe that "hallucination" and "generation" go hand in hand, implying that machines are also creating. Last year, at the Jiangsu Province Youth Science Fiction Works Release Conference, a short story entirely generated by AI won second prize. In the same year, GenWorld also held two "Chinese AI Flash Fiction Competitions," requiring all entries to be generated by AI.
Generating an entire text using AI is not easy, and contestants adopted various innovative methods to control the chatbot's output. However, AI still has significant limitations in storytelling. A study showed that professional writers scored higher in creativity, literary quality, and stylization.
Science fiction author Mu Ming believes that AI writing tools are more like a "partner," with their greatest function being to encourage daily work. She points out that AI's ability is weaker when it comes to highly literary works, especially in the overall structure, plot, and character development of novels, where human authors still have the upper hand.
Mu Ming believes that there are issues with the Transformer architecture itself, such as insufficient attention token length. Additionally, the corpus used by the model can have an impact, but due to the use of large amounts of data, it's difficult to assess the degree of influence specific corpora have on its performance.
Ultimately, large models use a completely different "writing approach." Mu Ming started writing novels in 2016, initially improving herself by learning techniques from her favorite authors. She believes that AI currently struggles to fully imitate the creative process and thinking patterns of human writers.