Machine finds hidden questions in stories
There is a particular kind of silence that falls over a reading group when someone realizes the text in front of them contains a puzzle nobody noticed. Not a hidden message in the cryptographic sense — no acrostics, no shifted alphabets. Something simpler and stranger: a question buried so deep inside a narrative that the narrative itself does not appear to be about a question at all. That silence is the territory a September 2026 paper from Daniel Eisner maps out. The work, titled EnigmaForge: The Question Is Hidden in the Story, sits in the cs.AI category of arXiv and describes a system that does something deceptively modest: it finds the interrogative at the heart of a text and makes it explicit. [1]
What the System Actually Does
Strip away the framing and the mechanism is straightforward. Given a story, EnigmaForge identifies the central question the story is implicitly answering — or failing to answer. [1] This is not summarization. Summarization tells you what happened. This tells you what the text is about at the level of inquiry. A fable about a fox and a crow is, at one layer, about flattery. At another, it is about whether trust can survive desire. The system surfaces that second layer. It does not generate the question from nothing; it extracts a question that was already load-bearing, already holding the narrative together, but never stated.
The distinction matters because it separates this from the long tradition of question-answering systems. Those take a question as input and return an answer. EnigmaForge reverses the polarity. It takes a narrative as input and returns the question. That inversion is the whole trick, and it is not a trivial one.
Why Inversion Is Hard
Language models are good at compression. They can read a thousand words and produce a hundred that capture the gist. But compression preserves what is frequent and discards what is singular. A buried question is often singular — it appears once, obliquely, in a detail that seems decorative. The crow drops the cheese. The fox walks away. The moral is obvious. The question underneath — what did the crow want more than the cheese? — is not in the text. It is in the gap the text leaves.
Finding that gap requires something closer to literary reading than to information retrieval. It requires holding the whole story in mind and asking what it would mean for the story to be wrong. That is a different cognitive operation than pattern matching, and it is the operation Eisner’s system attempts to automate.
The Concrete Gain
Here is where the argument sharpens. The gain is not that machines can now read stories. The gain is that a specific human bottleneck — the one where a teacher, an editor, or a researcher has to read a corpus and identify the unasked questions — becomes tractable at scale. A single careful reader can do this for a hundred texts. A system can do it for a hundred thousand, and then a human can verify the interesting ones. The human judgment does not disappear. It moves upstream. Instead of finding the questions, the human evaluates the questions the machine found. That is a real shift in where attention goes, and it is the kind of shift that historically precedes institutional change.
The Historical Parallel Nobody Draws
When indexes were first printed — alphabetical lists of what was inside a book — scholars worried that readers would stop reading deeply. The opposite happened. Indexes made deep reading possible at scale because they freed the reader from the search. The same pattern recurs with search engines, with citation databases, with the reference managers that now populate every academic workflow. Each tool did not replace the intellectual act. It relocated the intellectual act to a higher level of abstraction. EnigmaForge belongs to that lineage.
The Limit Case
There is a failure mode worth naming, and Eisner’s framing invites it. A system that extracts questions can also extract the wrong ones — questions that are plausible but not central, questions that reflect the training distribution rather than the text. The paper’s contribution is not that it solves this. It is that it makes the problem visible. Once you can see the extracted question, you can disagree with it. Disagreement is the beginning of understanding. A system that produces a wrong question is more useful than one that produces a confident answer, because the wrong question can be corrected. The confident answer just sits there.
What This Changes
The institutional consequence is quiet but real. Literature departments, legal teams, intelligence analysts, and anyone whose job involves reading a lot and asking what it means now have a tool that surfaces the question before the human arrives. The human still decides whether the question is worth asking. But the first pass — the exhausting, error-prone, unglamorous first pass — is no longer the bottleneck. That is not a replacement. It is a lift. The machine hands you the question the story was hiding, and then it steps back. What you do with that question is still, entirely, yours.