How a perfectly consistent character can become someone I no longer recognize.
One of the more frustrating experiences I’ve had with AI writing is watching a character slowly become more ordinary.
Not in one dramatic transformation. Nothing so convenient.
It happens one reasonable sentence at a time.
Someone who previously had a wonderfully unusual way of seeing the world gives a perfectly sensible answer. A character who might have made an inappropriate joke responds with polite concern. A peculiar reaction becomes a familiar one.
I read the scene and nothing seems particularly wrong.
Then I read another.
And another.
Eventually I realize that the characters are beginning to sound as though they have all attended the same course in appropriate conversational behavior.
The story still makes sense. Everyone remembers their relationships and objectives. The personalities have simply become smoother, safer and less distinctive.
Knowing facts isn’t knowing a person
I can describe a character in remarkable detail.
I can give the AI their history, their habits, their relationships, their favorite expressions and examples of conversations from earlier chapters.
That helps. Considerably, in some cases.
But people are more complicated than descriptions.
Dude’s unusual priorities aren’t merely a collection of quirks to insert whenever he appears. Chef’s practical thinking shouldn’t turn every conversation into a demonstration of practicality. Hex doesn’t become convincing simply because the narration remembers that it is a hand.
What matters is how those qualities affect the choices characters make when something unexpected happens.
And that is considerably harder to specify.
The danger of teaching a character to perform itself
As I developed better character references, I began catching more of the obvious mistakes.
But I also encountered an unexpected consequence.
The more thoroughly I described someone, the easier it sometimes became for the AI to produce a recognizable imitation of their personality.
An imitation can be accurate in all the obvious ways and still feel artificial.
The character makes the expected joke. They display the appropriate concern. They react according to their established traits.
And somehow I don’t believe them.
Real people aren’t constantly demonstrating who they are. They become tired of their own jokes. They surprise their friends. They borrow one another’s expressions. Sometimes they react badly to something they would normally handle well.
Fictional people need that freedom too.
When every reaction seems designed to prove consistency, the character begins to resemble a very conscientious actor playing themselves.
When more rules made things worse
I have experimented with using separate AI writing and review sessions, supported by increasingly detailed character references.
There are genuine benefits. A second perspective can notice a weak conversation, an unlikely reaction or a personality that has drifted away from its earlier portrayal.
But I’ve also seen how accumulated corrections can make the prose oddly defensive.
Characters start explaining their motivations. Scenes seem determined to demonstrate that every decision makes sense. The writing becomes very good at avoiding obvious mistakes and noticeably worse at surprising me.
It is a peculiar problem to create for yourself.
You spend a great deal of effort protecting the characters, only to discover that the protection has begun restricting them.
Sometimes the right response is another example or a better description.
Sometimes it’s removing a rule.
Sometimes I simply have to read the scene and ask whether I recognize the person I imagined.
Keeping people alive on the page
I don’t expect a character to behave unpredictably for the sake of unpredictability.
I want the surprises to come from somewhere: fear, experience, affection, frustration, pride or a change in how that person understands the world.
If someone who is normally funny becomes silent, I want the silence to matter. If someone who usually avoids danger takes a risk, I want the choice to reveal something.
And if someone changes over the course of a book, I want that change to feel like part of a life rather than a failure to follow a character description.
AI can help me remember an enormous amount about the people I’ve invented.
But remembering someone isn’t quite the same as understanding them.
That difference remains one of the most interesting—and difficult—parts of writing with it.