Why Does AI Prose Explain What It Should Let the Reader Discover?
This is Essay 15 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.
Is the Information Even Needed?
There is a passage you have read before. It arrives without warning, often in the Prologue if there is one, or by the second chapter. The story has been moving—there are characters, there is a problem forming at the edges—and then it stops. Not to breathe or turn but to explain.
In the next several paragraphs you learn the political structure of the kingdom, the history of the trade routes, the terms of the treaty that failed twenty years before the novel begins, the name and reputation of the previous king, and why the city was built on the plateau rather than in the valley below. The information is accurate and detailed yet has no heartbeat.
This is the info dump: backstory, worldbuilding, historical context delivered in undigested blocks, deposited in the text rather than threaded through it. The story pauses while the author prepares you for it. The author has calculated what you need to know before you can understand what happens next and has decided to give it to you now.
What the Info Dump Reveals
The reader does not require a complete map of the world before entering it. Only the next necessary facts.
The info dump is an error of mistrust, not excess. The writer who deposits three paragraphs of dynastic history before the scene has decided that the reader cannot follow the scene without them, which may be true for the writer himself. Sometimes the dump is a symptom of the author settling their own account with the material before returning to the story. Either way, the reader pays.
As he walks its streets, a character who has lived in a city for many years does not think about its founding and current politics. He thinks about where he is going, among other things. The founding and the politics are there. If the writer is attending, they arrive in the details that belong to this character in his present state.
What Generative AI Does
A language model approaches historical or secondary-world fiction as a problem of completeness. Before the story can proceed, the story must be grounded. Before the character can act, the context must be established. This is how a summary is organized, not how a scene moves.
The result is prose that knows what it knows and cannot resist delivering it. The model has been trained on text that explains, and text that explains tends to front-load its explanations. Given a historical setting, it produces the setting first: dates, arrangements, causes, precedents. Given a fantasy world, it produces the taxonomy before the character who lives in it. The information arrives because the information exists, and the model deploys it.
Consider: a writer asks the model to open a scene in a medieval guild hall where a journeyman has just been accused of theft. The model may well begin not with the accusation, not with the journeyman’s face, but with three sentences about the guild system, its origins in the eleventh century, the structure of its ranks, and the penalties it has historically imposed. The accusation, when it arrives, lands in explained territory. The reader has been prepared—and excluded.
This is not a failure of the model‘s knowledge, but of sequence. It is unaware that the reader does not need the guild system before the accusation. It does not grasp this as it cannot test the passage against a reader’s experience of time. The model does not feel the scene slow.
What the Human Writer Can Do
An info dump is a revisable error. It may occur in a first draft because, before proceeding, the writer needs to confirm where to stand with what is known. This is not a failure but a stage, the draft where the writer settles the account with the material before learning to spend it in scene. The revision is where the question is asked: where does this information belong? Not where must it be delivered, but where will the reader need it, carried by whose perception? The answer is almost always later, smaller, and attached to action.
A character in a Renaissance wool-merchant’s counting house does not announce that the Arte della Lana has governed Florentine wool production since the thirteenth century. He notices that the ledger his associate shows him uses a system of notation he does not recognize. One sentence. The history is in it. The reader was not briefed.
The human writer can make this correction by reading the passage aloud and discerning where it thickens. Of every explanatory paragraph, they can ask whether the story required the stop, or whether they required. They can distinguish between the two and eliminate the latter.
The model cannot make this correction. It does not experience the passage as a reader does. It produces text that covers the material, which is not the same as serving the scene.
The Thread, Not the Bolt
Information in a story is not a problem of quantity. It is a problem of integration.
The reader can absorb a great deal if it arrives threaded through what is already happening. What the reader cannot absorb is the pause. The pause that says: before we continue, let me tell you what you need to know.
The writer who threads does not hide their knowledge. They wait for the sentence where a detail belongs and place it there. The reader receives it without noticing they have been taught anything. The world accumulates. At the end, the reader arrives knowing the world—not because they were briefed at the beginning, but because they lived in it.
The thread requires patience: the willingness to let the detail sit in the writer’s knowledge until the story calls for it. That is what the model cannot wait for. It dumps.
Rio de Janeiro, the xxx day of June, MMXXVI.
