Sometimes the most difficult part of writing with AI is persuading it to let a character make a terrible decision.
I like characters who are capable of surprising me.
Unfortunately, surprises aren’t always pleasant.
People can be selfish, cruel, ignorant, ridiculous or convinced they’re doing the right thing when they are causing terrible harm. I find those contradictions fascinating. They are also a large part of why I enjoy creating fictional personalities.
But when I write with AI, I occasionally encounter a curious kind of resistance.
The story starts becoming more reasonable than I intended.
A character explains themselves before they should. A cruel decision gets softened. An uncomfortable moment acquires a reassuring observation that seems designed to make everyone understand the correct moral interpretation.
Sometimes that is a useful suggestion.
Sometimes it destroys precisely what made the scene interesting.
The freedom to write someone I disagree with
A fictional character doesn’t have to share my values.
I may want readers to understand why someone behaves badly without excusing what they do. I may want a sympathetic character to make a dreadful mistake. I may even want a scene to seem funny until a later revelation makes the same events horrifying.
None of that is necessarily an endorsement of the character’s actions.
It is storytelling.
I understand why AI providers place boundaries around harmful assistance. There are real risks that deserve serious attention. But writing fiction introduces a distinction I consider important: representing cruelty is not the same as recommending it.
I want room to explore that distinction without every morally uncomfortable moment being made comfortable before the reader has encountered it.
The story I didn’t tell the AI
One of my more revealing experiments involved the Attention storyline.
I had a larger narrative direction in mind, including consequences that would change how earlier events were understood.
But I chose not to give the AI writer every part of that plan.
The reason was simple. I wanted certain scenes to be written from what the characters understood at the time, without the ending leaking into their behavior.
Sometimes a joke needs to be allowed to remain a joke until the story reveals why it wasn’t quite so innocent.
Spoilers for Book Two — the Attention storyline
Across a series of interactions with Attention, Rae presses buttons.
At first, the interactions can feel like another piece of the dungeon’s absurd logic. There is an action, a reaction and the possibility of trying again.
Later comes the terrible realization that the presses have been delivering electrical shocks to Jimothy.
The same events suddenly mean something very different.
If Rae had understood those consequences from the beginning, she would have been making an entirely different kind of moral choice. The delayed realization changes how she—and the reader—must understand what happened.
I wanted to preserve that contrast.
So I developed the early material without giving the AI writer the full significance of the later revelation.
I cannot know exactly how the scenes would have turned out if I had shared everything. Perhaps the model would have handled the distinction perfectly. But I was concerned that it might make the early moments ominous, encourage Rae to behave differently or reveal too much through the narration.
Keeping part of the plan to myself was a way of protecting the experience I wanted the reader to have.
When a helpful tool knows too much
There is an odd discomfort in deliberately withholding information from an AI you have spent so much time working with.
A conversational assistant can begin to feel like a collaborator. You explain ideas, consider alternatives and return to the same fictional world repeatedly.
Yet giving it the complete plan is not always the best way to write an individual scene.
A human author can know the end of a story while allowing a character to remain completely unaware of it. In my experience, making an AI maintain that separation sometimes requires deliberate effort.
That introduces other problems. The assistant may miss subtle opportunities for foreshadowing or produce details that don’t fit later events.
Those are things I have to resolve.
But they are preferable to losing an important revelation before the story has earned it.
The kind of freedom I want
I’m interested in the cultural assumptions, preferences and boundaries that shape AI-generated writing. They matter because a writing assistant can influence a story even without explicitly refusing to produce anything.
I don’t assume every awkward passage is censorship. Sometimes it is simply conventional prose, an overhelpful suggestion or a misunderstanding of what I’m trying to achieve.
But the underlying creative question remains.
I want fictional people to be allowed to make mistakes that matter.
I want them to have beliefs I disagree with, behave in ways that surprise me and face consequences that cannot always be reduced to a simple moral lesson.
And I want readers to have the experience of deciding what those choices mean.
AI can help me explore possibilities I might never have reached alone.
The responsibility for deciding which possibilities belong in my stories is something I intend to keep.