BL and AI — what the genre built, and what cannot be handed over
What this genre accumulated over decades is not prose quality. It is resolution of relationship: the asymmetry between two people, how distance changes, what is deliberately never said. That happens to be the exact region where a language model is weakest, and the reason is structural rather than a matter of taste.
A genre with a everyday word for "not my reading"
This culture has a term other genres largely lack: a common, everyday way to say that a work does not match your interpretation of the same characters.
That the term is in daily use tells you the situation. The baseline assumption is that finding work matching your own reading is unlikely.
Which follows from the combinatorics. Who, in which position, at which stage of the relationship, weighting which canonical detail. The branches are effectively unlimited, and not every branch has someone writing it.
So readers have read the seventy-percent match and swallowed the rest. Being able to build your own reading removes that thirty percent. Competent prose is not the scarce resource here, and never was.
Why a model flattens the dynamic
Left alone, a model closes distance. Advancing is more probable than withholding, because most written romance advances. Two episodes in, everything has been said out loud.
It also resolves obstacles internally. Given two people with a problem, the likely continuation is a conversation in which they understand each other. Obstacles that live inside the characters get talked away.
Both behaviours are the same phenomenon: the distribution favours resolution, and the distribution wins unless something outside it intervenes.
Hand it the relationship, not the people
In practice, specify three things before any character detail. The asymmetry — who holds what. The distance — what is currently permitted. The unsaid — what one of them knows and will not say.
Put the obstacle outside the two of them: position, time, a third person. External obstacles do not dissolve when the characters have a good conversation.
And state the limit per episode. "This time their hands touch; they do not look at each other." Without a ceiling, one episode consumes the entire arc.
The part that stays yours
What a model cannot supply is why this particular pair, read this particular way, matters to you. That comes out of having read closely, argued about it, and noticed something nobody wrote down.
One line of that is enough. "He apologises by fixing something, never by speaking." The hundred scenes that follow from it are machine work.
The change worth noticing is not that AI writes prose. It is that a reading held by one person can now become a finished work without that person also having to be a writer.
Questions
Why does the relationship resolve so quickly?
Advancing is more probable than withholding. State a ceiling for each episode and put the obstacle outside the two characters.
Can a model handle subtext?
Poorly, by default — unstated information is more probable stated. It can hold an explicit instruction not to say something, so specify the unsaid rather than hoping for it.
What should I supply myself?
Your reading: how these specific people behave in scenes nobody wrote. That is the only material with no average behind it.