Why Do AI Similes and Metaphors So Often Miss the Mark?
This is Essay 3 of 22 in the series Why Choose 100% Human Authors, addressed to readers who want to understand what a human author does that a machine cannot and why that difference matters to the experience of literary fiction.
The Comparison That Contradicts Itself
A simile or metaphor asks the reader to understand something by seeing it in terms of something else. The mechanism depends on the comparison illuminating a quality the two things share. When that relationship is absent, the figure of speech fails to illuminate. It confuses.
Consider this kind of sentence, which anyone who has read AI fiction will recognize:
The explosion tore through the valley like a held breath.
Or this:
Her grief was a bright, cheerful color that no one could look at directly.
Or yet another:
The ship’s horn split the silence like a whisper finding its way home.
Each of these has the grammatical shape of a comparison. Each is assembled from words that, individually, belong to the language of literary fiction. And on examination, each is nonsense—not because the words are wrong, but because the relationship they claim does not exist. An explosion is not like a held breath: the two share no quality that the comparison could reveal. Grief is not a bright, cheerful color: the contradiction is not paradox, it is confusion presented as insight. A ship’s horn does not split silence the way a whisper does: the whole point of a whisper is that it does not split anything.
These are not rare failures. They are among the most consistent signatures of machine-generated literary prose.
Why the Machine Produces Them
To understand why AI systems generate these broken comparisons, it helps to understand what a simile looks like from the outside.
From the position of a system that has read millions of texts but experienced none of them, a simile is a structure: X is like Y, or X was Y. The system has learned that literary prose uses these structures, that they are valued, and that the Y term tends to be drawn from a different semantic domain than the X term. A physical event is compared to an emotional state. An emotion is compared to a natural phenomenon. A sound is compared to a texture, a color to a temperature.
What the system has not learned is the underlying logic that makes the crossing of domains meaningful. That logic is always the same: the two things compared must share a genuine quality that the reader can feel. Her voice was silk works because silk and a certain kind of voice share smoothness, a physical quality the reader can sense. His anger was a weather system moving in from the coast works because anger and weather share the quality of arriving before one is ready, of changing everything in its path. The reader does not need to be told what the shared quality is. They feel it, and the feeling is the point.
A language model assembles the structure—the like, the crossing of domains—without verifying whether the shared quality exists. It produces comparisons that look like metaphors and read like riddles with no answer.
The Difference Between Paradox and Confusion
A skilled author may create comparisons that appear contradictory on first reading. This is intentional, and it works because the contradiction resolves into meaning upon examination.
The silence after the guns was louder than the guns had been. This appears to contradict itself. Silence cannot be loud. But the reader understands at once what is meant: the absence of the noise made the noise’s presence more felt. The apparent paradox is a fitting description of an experience anyone who has lived through a sudden quiet after sustained noise will recognize.
This is not what AI prose produces when it creates contradictory comparisons. The AI contradiction does not resolve. Her grief was a bright, cheerful color does not deepen upon reflection into emotional truth. It remains two things that have been placed next to each other without a shared quality or a felt experience to justify their proximity.
The difference between paradox and confusion is the same as between a lock and a wall with a lock painted on it. One opens. The other only looks as though it might.
What a Human Author Does Instead
A human author arrives at a comparison through a process that is not assembly but discovery. The writer knows the thing they are trying to describe and searches for the second term not by crossing domains at random but by asking: what does this remind me of? What shares its qualities?
The search may be long. A comparison that appears effortless on the page may have cost the author an hour, or a rewrite, or a long walk during which the right image surfaced. What makes it seem so seamless is that the shared quality is so well-matched that the reader does not notice the mechanism. It shows the reader something they had not seen before, in a light that makes them think: yes, that is it.
That recognition is what a broken AI comparison can never produce. In its place it yields something closer to: I suppose I see what was intended—or, more often, what on earth does that mean?
At the moment of confusion, the reader does not always know that the writer was a machine. But they feel something slip. The world the fiction was building becomes less solid, and the reader becomes aware of the page in a way that good fiction never permits.
What a human author offers is the accuracy of a mind that has tested the comparison against experience and found that it holds. The metaphor that earns its place in a sentence is not the most surprising comparison, or the most elaborate, or the most literary. It is the one that is true. And truth that is felt, verified, and discovered rather than generated cannot be reproduced by a system that has never felt anything at all.
Rio de Janeiro, the xxv day of June, MMXXVI.
