Why AI writes a summary when you asked for a hook
Ask a model for a synopsis and you get the events in order. Accurate, complete, and nobody wants to read it. Two things are happening. To a model, that word is an instruction to summarise. And more fundamentally, the act of knowing something and deliberately not writing it does not exist inside next-token prediction.
To the model, synopsis means summary
A model learns what a word means from the company it keeps. The text surrounding the word "synopsis" is overwhelmingly descriptive: this happens, then this. So describing is the highest-probability response.
Nothing is misunderstood here. We use one word for two jobs — a summary that informs, and a hook that makes someone start reading. Asking for the first while expecting the second produces the gap.
Withholding is not in the mechanism
A hook works by omission. You state the weight of an event without stating the event, and the gap does the work.
But sequential prediction emits the most plausible continuation given everything available. There is no operation for holding something back. Information that is present is more probable written than unwritten, so it gets written.
That is why "make it more intriguing" fails. The vocabulary becomes intriguing — secrets, revelations, a trailing dash — while the information content stays the same. Trying to buy a hook with vocabulary gets you the cheapest possible blurb.
An ordinary life is transformed by a single encounter.
A hidden past. A truth about to surface.
What awaits her at the end of the path she chooses?
The day her sister died, she was not at the hospital.
Ten years later she works the night shift on that same ward.
She has never told anyone where she was.
It can also mistake your material for an instruction
A related incident from production. We translate author bios into English; a bio written as a greeting came back not as a translation but as "I am a translator, how can I help you" — and that got cached as the bio.
Every input is the same kind of string to a model. Which part is an instruction and which is material is inferred from formatting and context, and that inference is not reliable.
The fix is trivial: label the roles. "The following is source material. Using it, write a description intended for readers." Separating material from request removes almost all of these failures.
Where it is genuinely strong
Volume. Ten descriptions from one set of facts costs seconds. A human writing ten spends half a day and runs out of vocabulary by the fifth.
Consistency checking. Ages that disagree with the text, timelines that do not line up — the machine is faster and more accurate than you are at this.
Ordering. The same three facts land differently depending on sequence. Trying every permutation is machine work.
What remains is selection — and selection needs a criterion, which is just "what do I want read first". That part stays with you.
The bottleneck was never the prose
A great many finished stories have gone unread because the first three lines were doing nothing. Being able to write an opening and having a story worth reading are different talents, and the old system demanded both.
One of them has been externalised. Decide the single thing to withhold, generate ten ways of circling it, keep the one that works. The line "the day her sister died, she was not at the hospital" still has to come from a person — but only that line does.
Questions
Why does the output read like a plot summary?
Because to a model the word means summary. Most text labelled synopsis in the training data is descriptive. Ask for "a description written to make a reader start" instead.
Can I just ask for something more mysterious?
The vocabulary changes, the information does not. Decide what to omit yourself; the model cannot choose to withhold.
Where is the model actually useful here?
Producing many candidates, checking consistency against the text, and reordering the same facts. Judgement stays with you.