AI Writing Poetry Does Not Really Understand Its Meaning

When I opened my computer, I came across code I had written earlier for AI poetry. Online examples tend to be Tang poetry, while what this humble Taoist collected is mostly verse from the Daoist Canon. Though most of the output is unsatisfying, a few pieces are occasionally amusing.

Ultimately it is still deeply disappointing: when studying it I had hoped to understand how it grasps semantics, only to discover it is merely the result of probabilistic continuation.

Its principle is very simple — through an LSTM it predicts the most likely symbol at the next time step. In other words, it does not care whether the underlying content is poetry or the holistic meaning of English words.

Shifting perspective, it is still computing relationships between word vectors, then searching nearby for similar characters based on that relationship, and finally emitting poetry or words.

This clearly lacks any real theoretical foundation of intelligence and is only a simulation — more precisely, a kind of sophisticated parrot that has not truly understood the text, merely calculating by probability which symbol should appear next.

Of course it has its interesting side: if imagined more radically, after enough simulated learning on one person, an AI could develop response patterns fully matching that person's style, making even their thinking, decisions, and behavioral traits predictable. Applied to crime prediction, it would become like the future depicted in Minority Report.